Here’s Your Number, Now Please Wait in Line: The Asylum Backlog, Federal Court Litigation, and Artificial Intelligence in Agency Adjudication

Introduction

The U.N. High Commissioner for Refugees has estimated that, by the end of June 2021, there were nearly 4.4 million pending asylum applications worldwide.1 Many asylum seekers suffer heinous abuses in both the countries from which they flee and the countries through which they travel to reach sanctuary.2 To make matters worse, most flee with only the “barest necessities.”3 It was reported that in 2020 the United States granted asylum, both affirmatively and defensively,4 to 31,429 persons5 —16,864 of whom were granted asylum affirmatively.6 As of April 18, 2022, there were 194,840 affirmative asylum applications pending in the United States.7

Although statistics on wait times for affirmative asylum applications have not been made public since 2018, some advocacy and legal-services organizations have reported that their clients wait an average of more than four years.8 The gravity of these delays cannot be understood as an abstract number; rather, it must be understood by the human toll those delays exact. For example, Aaron is an Ethiopian man who was engaged and had hoped to bring his fiancée over to the United States once he was granted asylum.9 However, while he languished in the backlog for over five years, his fiancée married someone else because she couldn’t wait for him any longer.10 Similarly, Ibrahim, a Pakistani human rights activist, has been waiting for his asylum interview for nearly six years while separated from his wife and children, who remain in danger in Pakistan due to his activist work.11 His youngest daughter was three when he fled;12 if ever reunited with her, Ibrahim will have missed more than two-thirds of her life.

Like anyone else, Aaron, Ibrahim, and the other asylum seekers with claims pending in the backlog live complex lives that can be irreparably damaged by indefinite waits. While the fact that many asylum seekers must wait over four years for claim adjudication may seem inconsequential, it takes years away from the lives of tens of thousands of individuals—years that can never be given back.

While a foreign national13 waits for a decision on her asylum application, she is entitled to remain in the United States and pursue employment authorization.14 Some asylum seekers might be content with the opportunity to stay in the country and work, even without stable legal status.15 Others face unique and precarious circumstances related to their flight. These individuals and their families might be especially harmed by a multiyear wait because they cannot access particular benefits available only to those with definite legal status, particularly asylum. This Comment specifically identifies two subsets of asylum seekers who are in this position. The first subset is comprised of asylum seekers who require medical care but are unable to afford, or gain access to, medical services without the federal public benefits available to those granted asylum. In the second subset are asylum seekers who wish to be reunited in the United States with immediate relatives who face threats of persecution in their home countries but cannot undertake the journey to seek asylum themselves.16

This Comment focuses on these two groups of asylum seekers for two reasons. The first reason is my normative belief that individuals facing such threats should be prioritized for humanitarian relief. The second reason is that there is a legal basis for focusing on these groups because of the factors the U.S. government already considers in need of particular attention.17

In this Comment, I do not attempt to propose a solution to the backlog itself. Rather I ask, and seek to answer, whether there is a way that individuals in the specified subsets might access reliable and consistent relief under the current system despite the backlog. Because the threats of harm to these subsets of asylum seekers require particular attention, this intervention attempts two separate but related tasks. First, it assesses whether the current “solution,” or avenue to relief, on which advocates and asylum seekers have relied—federal court litigation—is actually viable for its purported ends. Second, drawing on existing government practices and recent technological developments, it proposes a novel path to relief and argues that the proposed path is a superior avenue to relief than litigation.

In Part I, this Comment provides background on the history and legal framework of asylum in the United States. Part II turns its attention to a method some asylum seekers have used to attempt to escape the seemingly indefinite queue—bringing claims of unreasonable delay under the Administrative Procedure Act18 (APA) in federal court. Part II.A outlines the legal basis and basic framework of unreasonable delay, and Part II.B reviews courts’ dispositions of these claims to show that courts have overwhelmingly denied plaintiffs relief.

Part III then argues that the reliance on federal court litigation to provide relief to the identified subsets of asylum seekers is misplaced because the judiciary is neither a well-positioned institutional actor nor better positioned than Congress or the executive branch to deal with this type of problem. Specifically, this Comment argues that (1) these claims of unreasonably delayed asylum adjudication ask courts to engage in a task for which they are institutionally ill-suited; and (2) these claims ask for a judicial remedy that raises serious separation of powers concerns. Although judicial review of unreasonably delayed agency adjudication has led to relief in some cases, 19 and has found support in scholarly literature,20 this Comment contends that the asylum backlog and its attendant problems are ill-suited for judicial
intervention.

Although the judiciary may be an imperfect institution for a solution, all human institutions are imperfect, so institutional deficiency is not, by itself, a reason for abandoning litigation as a tool. Instead, because another institutional actor is more capable of providing relief, I advance that advocates and asylum seekers should turn to this actor instead of the courts. This actor is the executive branch, and more specifically, the agency administering affirmative asylum: U.S. Citizenship and Immigration Services (USCIS). Drawing on the wealth of recent literature on the use of artificial intelligence in agency administration, I propose in Part IV a new agency-side adjudicative mechanism to provide relief to the identified subsets. Part IV.A discusses the use of artificial intelligence in federal agencies. Part IV.B outlines the structure of the proposed mechanism, wrestles with some of the details of implementation, and discusses how it can serve as the impetus for further innovation in the agency. Part IV.C assesses how administrative law doctrines can address some of the serious legal concerns that arise when administrative agencies use artificial intelligence in core agency functions. Finally, Part IV.D concludes by reflecting on the comparative advantages for asylum seekers of an agency-side solution to litigation.

I.  Asylum in the United States

A.    History and Legal Framework

The Refugee Act of 198021 (“Refugee Act”), amending the Immigration and Nationality Act22 (INA), “establish[ed] an explicit asylum provision in [U.S.] immigration law for the first time.”23 Under the Refugee Act, a foreign national who is determined to be a “refugee” as defined by the INA24 and is physically present in, or arrives to, the United States may be granted asylum.25 A refugee is an individual outside of her country of nationality26 who is “unable or unwilling to return to, and is unable or unwilling to avail himself or herself of the protection of, that country” because of past persecution or a well-founded fear of
persecution.27

A foreign national may apply for asylum through two legal avenues—“affirmative” and “defensive” asylum.28 Defensive asylum is claimed by foreign nationals only in removal proceedings, and it is raised to seek relief from deportation.29 The agency charged with adjudicating defensive asylum claims is the Executive Office for Immigration Review (EOIR) in the Department of Justice; the proceeding is adversarial in nature, and the adjudicator is an immigration judge.30 Affirmative asylum is a process that allows a foreign national who is not in removal proceedings to apply to the U.S. government for asylum.31 Foreign nationals must apply for asylum within one year of the date of their last arrival in the United States unless changed circumstances materially affected their eligibility or extraordinary circumstances caused the delay in filing.32 Notably, the proceedings for affirmative asylum are nonadversarial.33 This Comment concerns the affirmative asylum process and its attendant problems.

Affirmative asylum is governed by 8 U.S.C. § 1158 and administered by USCIS. Authority to establish procedures and requirements by which a foreign national should apply for asylum is vested in the Secretary of the Department of Homeland Security (DHS).34 As such, DHS regulations and USCIS policies are the application process’s most important rules.35 To affirmatively apply for asylum, one must file a Form I-58936 with USCIS, which will then schedule an interview with an asylum officer. After further review, the asylum officer will make her decision on the application.37

DHS regulations do not provide a timetable for adjudication, but the INA itself does put forth a “soft” timetable. Section 1158 provides that, absent exceptional circumstances, the initial interview or hearing on the application “shall commence not later than 45 days after the date an application is filed” and the final decision “shall be completed within 180 days after an application is filed.”38 However, § 1158 explicitly does not create a private, judicially enforceable substantive or procedural right based on these timelines.39 Thus, in practice, it often takes years for applicants to get their interviews scheduled.40

While waiting for a decision, an asylum seeker may remain in the United States and pursue employment authorization.41 If granted asylum, that person may remain in the United States and may also travel outside the country with permission from the proper authorities. The asylee is legally authorized to continue working and is eligible for certain federal humanitarian and immigration-related benefits and services.42 Further, individuals granted asylum eventually become eligible to become legal permanent residents (“green-card holders”)43 and may begin fulfilling the requirements to become U.S. citizens.44

B.    Cause of Delay? The Asylum Backlog

From early on, it was apparent to some that the asylum procedures created by the Refugee Act were incapable of dealing with the strains of mass adjudication. Because of the Cold War, Congress was paying “scant attention” during the process of drafting asylum adjudicatory procedures, and the administering agency, the Immigration and Naturalization Service (INS),45 did not advocate for specific procedures.46 Soon after the Act’s passage, to justify aggressive exclusion, deportation, and interdiction policies against asylum seekers (particularly Haitian asylum seekers), the Reagan administration and INS officials argued that the Refugee Act had not contemplated, and therefore was ill-equipped to deal with, sudden massive influxes of asylum seekers.47 These concerns about the asylum system came to fruition in the 1990s when the country saw a drastic increase in migration.48 New asylum filings rose to nearly 150,000 per year,49 and in 1995, in large part because of the pronounced lack of resources and effective processing,50 the total number of applications in the backlog peaked at about 464,000.51

In response to the backlog crisis, the INS and Congress adopted a variety of reforms in 1995. First, applicants could no longer apply for employment authorization immediately after filing; rather, they had to wait until either asylum status had been granted or 150 days had passed since the date of filing, whichever was earlier.52 Second, the INS implemented the “Last-In First-Out” (LIFO) scheduling system, under which the newest applications were prioritized for processing while the agency slowly worked backwards through pending cases.53 The policy rationale behind both changes was deterrence: There was growing concern that frivolous applications were being filed simply to delay removal and gain work authorization.54 Officials reasoned that if the INS could swiftly weed out frivolous claims before the applicants gained work authorization, others might be deterred from filing frivolous applications. Lastly, Congress approved appropriations that increased INS’s staffing of asylum officers from 150 to 600 by 1997.55

The reforms’ effects were considerable. The backlog dipped to fewer than 350,000 cases in 1997,56 and by 1999, new filings had decreased by 80%.57 By 2006, the number of pending cases was down to about fifty-five thousand.58 The decline in the backlog was coupled with an increase in the percentage of claims granted.59 However, due to the LIFO system, prereform applications suffered the brunt of the backlog because they became the last to be adjudicated. USCIS noted that by the end of 2003, most of the 262,118 pending cases had been filed before the 1995 reforms.60 That meant that over two hundred thousand individuals, despite being nonremovable and eligible to work, had been without legitimate legal status for at least eight years.

Between 2005 and 2014, the backlog remained under one hundred thousand cases,61 and in 2014, USCIS switched from LIFO to the “First-In First-Out” (FIFO) prioritization schedule.62 Although there had been a slowly increasing upward trend, the next few years saw a sharp climb in the number of applications filed and the total number pending.63 By April of 2022, there were 194,840 pending applications.64

Recent years have seen an expansion of the reforms from the 1990s. In 2018, USCIS reinstituted the LIFO schedule.65 In 2019, asylum-officer staffing increased to 771.66 Despite these changes, the COVID-19 pandemic—which caused office closures and decreased processing efficiency67 —and various Trump-era policies68 have further contributed to lengthening wait times. Although the 1995 reforms demonstrate that agency-side responses to the backlog can be effective, they also show that asylum seekers who happened to arrive to the country earliest may bear the greatest burden of slow agency responses. Although the number of new filings decreased from 2020 to 2021,69 it remains unclear whether this trend will persist in 2022. Even so, it is unlikely to save pre-2018 applicants from the same fate—indefinite uncertainty in legal status—that pre-1995 applicants faced.

II.  A Potential Lifeline: Bringing Claims of Unreasonable Delay in Federal Court

To find relief from their legal limbo, some asylum seekers have brought claims of unreasonable delay in federal courts under the APA. These litigants hope to show the judge how long they’ve waited and the harms that they have suffered as a result so that the judge might compel prompt adjudication—particularly because USCIS has implemented a priority schedule that deliberately causes further delay. To give away the punchline: in nearly all such cases, plaintiffs’ claims have been dismissed for failing to state a claim for which relief could be granted.

Part II.A briefly outlines the legal theory of unreasonable delay under the APA and discusses the factors that courts use to decide whether agency inaction is sufficiently ripe for judicial review. Part II.B then provides a descriptive overview of courts’ analyses that have led to the dismissal of plaintiffs’ claims.

A.    Unreasonable Delay Under the Administrative Procedure Act

The APA functions as the backbone of administrative law. While USCIS is subject to the requirements imposed by the INA, it is also—like any other federal agency—subject to the APA’s mandates. The APA requires that agencies, with “due regard for the convenience and necessity of the parties or their representatives and within a reasonable time,” conclude matters presented before them.70 The APA also provides that a person who is “suffering legal wrong” or “adversely affected or aggrieved” by agency action is entitled to judicial review of that action.71 This has been interpreted to mean that an individual must establish that the injury of which she complains “falls within the ‘zone of interests’ sought to be protected by the statutory provision whose violation forms the legal basis for [her] complaint.”72

The APA defines the forms of judicial review. In relevant part, it provides that a court may “compel agency action unlawfully withheld or unreasonably delayed”73 unless judicial review is precluded by statute or the agency action (and the timeline of that action) has been legally committed to agency discretion.74 Although this review authority seems quite broad, a court may compel agency action only once it determines that the agency action is (1) discrete,75 (2) final,76 and (3) “legally required”77 and that (4) the challenger’s injury falls within the statute’s zone of interest.

The question that has challenged courts for decades in unreasonable delay cases is when a failure to take a particular action becomes sufficiently “final” to be ripe for judicial review.78 In Telecommunications Research & Action Center (TRAC) v. FCC,79 the D.C. Circuit announced six factors to guide a court when deciding whether delayed agency action is ripe for review and may be compelled.80 The factors, which other courts have adopted,81 are as
follows:

1. The time agencies take to make decisions must be governed by a “rule of reason.”
2. Where Congress has provided a timetable or other indication of the speed at which agencies are expected to proceed, this may supplement the rule of reason.
3. Delays that may be reasonable in the sphere of economic regulation are less tolerable when human health and welfare are at stake.
4. The court should consider the effect of expediting delayed action on agency activities of higher or competing priority
5. The court should also consider the nature and extent of the interests prejudiced by the delay.
6. There need not be a finding of impropriety or bad faith by the agency to find unreasonable delay.82

The D.C. Circuit has since emphasized that it is especially important to consider competing priorities (the fourth TRAC factor) and that the “ultimate issue, as in all such cases,” is whether the delay satisfies the first TRAC factor, the rule of reason.83 Finally, the Supreme Court has held that, when considering the relevant factors, the “evidence of the passage of time cannot, standing alone, support” a claim of unreasonably delayed administrative action.84

In sum, to bring a claim that the adjudication of her asylum application has been unreasonably delayed, a plaintiff will need to demonstrate that (1) her interest falls within the “zone of interests” contemplated by the statute, (2) adjudication of an asylum application is a sufficiently discrete task, (3) adjudication is a task to which USCIS must legally tend, and (4) the agency’s inaction is sufficiently final to be ripe for judicial review. This final point is the most difficult to get a handle on, both theoretically and in practice, because the question of unreasonable delay is not decided with reference to some externally determined length of time that is “unreasonable.”85 Rather, courts focus on the specific facts of each case to weigh the plaintiff’s interests against the government’s;86 however, as will be discussed in Part III, broad institutional considerations may also slip into the analysis, further complicating the inquiry.

B.    Courts’ Disposition of Claims of Unreasonably Delayed Asylum Adjudication

Despite the deference generally given to the executive branch in immigration-related matters and the various provisions barring judicial review scattered throughout the INA,87 most courts have concluded that they have subject matter jurisdiction over these claims.88 There is consensus that plaintiffs cannot seek mandamus89 to compel action pursuant to the timetables set forth in § 1158 because § 1158(d)(7) precludes a private right of action90 based on these timetables.91 There is, therefore, no legal duty of prompt adjudication owed to the plaintiff that may be compelled. Courts also agree that § 1158(d)(7) is not so broad as to preclude judicial enforcement of the APA’s more general requirements discussed in Part II.A—namely that agency action not be unreasonably delayed.92

Courts generally agree that the adjudication of an asylum application is a sufficiently discrete action upon which USCIS has a lawful duty to act, even if the INA’s specific timetable for application processing is not enforceable against the agency.93 The question of whether USCIS’s failure to act is sufficiently final—in other words, whether the adjudication is unreasonably delayed—is where nearly all claims fail.94

To assess whether plaintiffs have a legitimate claim of unreasonable delay, courts have applied the TRAC factors at the pleading stage.95 Under the first TRAC factor, the LIFO schedule is generally considered a “reasoned response to a systemic crisis” and thus constitutes a sufficient rule of reason.96 Plaintiffs sometimes argue that LIFO cannot be a rule of reason because it deliberately causes longer wait times for older applicants. But these contentions are dismissed on the ground that the delay is “a byproduct of a reasoned attempt to address mounting issues with the asylum application process.”97 The second TRAC factor is generally unimportant in these cases. Some courts have not considered the statutory timetable at all because it is unenforceable by § 1158’s own terms,98 while others have found that this factor tips in the government’s favor because its nonmandatory nature supports the conclusion that LIFO is a rule of reason.99

The third and fifth TRAC factors—the human welfare and “extent of the interests prejudiced” factors100 —are often considered in tandem. Their resolution usually depends on each case’s facts, but several courts have rejected arguments that the harms visited upon waiting asylum seekers—for example, the psychological effects of uncertainty, the inability to petition to bring over one’s family who may be suffering persecution in their home country, and the inability to access federal public benefits or medical services—are sufficient to tilt these factors in the plaintiffs’ favor because such harms are “inherent in the process of seeking asylum.”101 Other courts have “declin[ed] to hold that ‘relief in this Court [will] never be available to an [asylum seeker] facing a delayed adjudication’” but the issue was that the plaintiff’s complaint was “completely devoid of facts” relating to the third and fifth TRAC factors.102 However, unless these courts denounce the analyses of Zhu v. Cissna103 and Fangfang Xu v. Cissna,104 it is not clear what facts a plaintiff would need to allege that are not “inherent in the process of seeking asylum.”105

The largest axe to plaintiffs’ claims has been the fourth TRAC factor, which states that a court should consider the effect of compelled action on agency activities of a higher or competing priority. Courts have recognized the immense backlog with which USCIS must deal and have noted that granting relief to a particular plaintiff simply allows her to “leapfrog[ ] other asylum applicants, and undermin[e] USCIS’s reasons for implementing the LIFO system.”106 Relying on the D.C. Circuit’s reasoning that judicially mandated queue-jumping “produce[s] no net gain,”107 courts have concluded that this factor heavily favors finding that agency action has not been unreasonably delayed.

III.  Searching for Relief in the Wrong Places: Misplaced Reliance on Federal Court Litigation

Despite the lack of success thus far, the comments of some courts108 might give hope to applicants and advocates alike that continued litigation efforts may prove fruitful. This Part advances the argument that continued reliance on federal court litigation as the primary means of gaining relief is misplaced for two reasons: First, claims of unreasonably delayed asylum adjudication ask courts to engage in a task that not only exceeds their institutional competencies but also falls squarely within the competencies of the administering agency. Second, these same claims ask courts to intervene in agency business in a way that raises serious separation of powers concerns. From a court’s perspective, these considerations taken together strongly counsel in favor of judicial restraint. And from an asylum seeker’s perspective, these factors both illuminate the deficiency of the judiciary as an institution that can deal with this problem and counsel in favor of looking elsewhere for a more viable avenue to relief.

A.    Limits of Institutional Competence

In a variety of contexts, like statutory and constitutional interpretation, scholars have argued for a comparative approach to analyzing which institution can make a legal decision most accurately and efficiently.109 The following sections concern themselves with similar arguments and demonstrate that the tasks required for resolution of asylum-context unreasonable delay cases—judgment of whether an agency’s prioritization of tasks and allocation of resources are “correct” or “reasonable”—exceed the judiciary’s competencies and are better left to another institutional actor. Part III.A.1 first briefly highlights some of the judiciary’s competencies before delving into case law exploring how the problems posed by unreasonably delayed asylum adjudication claims (1) are ill-suited for judicial resolution, and (2) fall more within the administering agency’s competencies. The most important component of this subsection is that it contains both a negative and positive claim: courts are not good at the task at hand, but the administering agency is adept at it. Part III.A.2 jumps into the intricacies surrounding USCIS’s administration of the asylum program to elucidate how the concerns animating the negative and positive claims sketched out in Part III.A.1 are acutely implicated when courts are called to decide these unreasonable delay claims.

1.   Courts are institutionally ill-equipped to review an agency’s prioritization of tasks and allocation of resources.

When choosing to allocate decision-making authority between different institutional actors, scholars have argued that assessing each actor’s relative strengths and limitations is important.110 Given that plaintiffs ask courts to decide unreasonable delay claims, I first sketch some of the competencies of the federal judiciary. On a granular level, tasks in which courts are particularly well versed include interpreting statutes, applying clear law to facts,111 ruling on the admissibility of evidence, and providing reasoned justifications for results reached.112 At a higher level, by virtue of judges’ experience and background, courts have a “degree of special competence” when deciding and reviewing questions of procedure and process.113 Another characteristic of the judiciary, seen as an advantage in many instances, is the insulation, and resultant impartiality, of its decision makers.114 The foregoing does not purport to be an exhaustive portrait of the judiciary’s strengths; nor does this subsection go so far as to argue that when a particular controversy before a court strays outside these competencies, the court should abdicate its responsibility to decide under the pretense of allowing the more competent institution to decide. However, cognizance of courts’ competencies helps contrast the specific task presented in unreasonably delayed asylum adjudication cases—reviewing agency decisions about task prioritization and finite resource allocation—from those tasks briefly outlined here.

Review of agency inaction must start with the recognition that agency actions are informed by and impact countless other parties. Agencies often are not singularly concerned with the administration of just one task. Most are in constant contact, coordination, and sometimes conflict with other agencies, regulated parties, and stakeholders. Agencies also often juggle a variety of tasks that, even if not in conflict, compete for priority. Accordingly, agency policies and priorities often reflect a reasoned compromise, a balancing the agency revisits and recalibrates. If one were to draw out these associations, it would appear as a “large and complicated web of interdependent relationships”115 that is “‘many-centered’ [so] that a pull at any one point changes an entire web of relationships.”116 This is particularly true of USCIS, whose work not only implicates hundreds of thousands of applicants and beneficiaries of many immigration benefits but is often also tied up with the prerogatives of other agencies like the U.S. Immigration and Customs Enforcement (ICE), the U.S. Customs and Border Protection (CBP), the Federal Bureau of Investigation (FBI), and the EOIR.

It is well recognized by the federal judiciary that adjudication of agencies’ internal task priorities and resource allocation goes beyond courts’ institutional competence. Courts commonly defer to, or even sometimes refuse to review, such agency determinations, as seen below. This deference, grounded in judgments concerning each institution’s comparative competencies, extends across many areas of law and to different types of agency action. For instance, in Heckler v. Chaney,117 the Supreme Court held that the Food and Drug Administration’s (FDA) decision not to take enforcement actions against drugs used for lethal injections in capital punishment was not subject to judicial review under the APA.118 The Court stated that when decisions like these are left to agency discretion by Congress, they are “general[ly] unsuitab[le]” for judicial review.119 It explained that enforcement decisions “involve[ ] a complicated balancing of a number of factors,” including how to budget agency resources according to competing concerns and overall policy goals.120 The Court further emphasized that this balancing is “peculiarly within” not the courts’ but the agency’s expertise.121 This deference is not unique to agency enforcement actions. In Lincoln v. Vigil,122 the Court was asked to review the Indian Health Service’s decision to reallocate portions of a lump-sum appropriation from one program to another. In holding that the decision was committed to agency discretion and thus unreviewable, the Court observed that agencies are “far better equipped than the courts to deal with the many variables involved in the proper ordering of its priorities.”123

Heckler and Vigil are admittedly distinct from the claims discussed in Part II.B, in that the agency’s decision-making in the asylum-delay context is not committed to agency discretion as a matter of law because the APA directs courts to “compel agency action unlawfully withheld or unreasonably delayed.”124 Nonetheless, their lessons are generally applicable.125 As such, two concerns animate judicial decision-making in more closely analogous circumstances: (1) that courts are not competent to review complex, and often political, agency decisions that balance various policy- and fact-based factors; and (2) that agencies are comparatively—if not uniquely—better positioned to make these decisions, like resource allocation, when it comes to the administration of their tasks.

To see how courts have applied these lessons to claims specifically concerned with mass-adjudication delay and requests for compelled adjudication, consider these cases from the D.C. and Fourth Circuits. In In re Barr Laboratories,126 the D.C. Circuit was asked to issue a writ of mandamus compelling the FDA to either approve or disapprove Barr’s drug and antibiotic applications.127 The court denied the request on the basis that it had “no basis for reordering agency priorities” and observed that “[t]he agency is in a unique—and authoritative—position to view its projects as a whole . . . and allocate its resources in the optimal way.”128

In Blanco de Belbruno v. Ashcroft,129 a case related to the problems of the defensive asylum backlog, the Fourth Circuit upheld regulations streamlining the review of certain defensive asylum claims as permissible constructions of the INA.130 Central to its reasoning was that, given a situation in which there is a backlog of tens of thousands of asylum claims and in which the agency “operates in an environment of limited resources,” how the agency “allocates those resources to address the burden of increasing claims is a calculation that courts should be loathe to second guess.”131

In both cases, the courts recognized two points bearing on the fundamental question whether the delay was reasonable: First, the courts acknowledged that they, compared to the agency, are at an informational disadvantage for assessing how one allocation of finite resources over another will affect the whole of the agency’s prerogatives. Second, the courts understood that these decisions are at a basic level ones of policy and that such decisions are better left to a politically accountable actor.

2.   How claims of unreasonably delayed asylum adjudication specifically implicate the concerns of judicial competence that animate deference to agency task prioritization and resource allocation.

It is important to understand that neither the backlog nor USCIS’s Asylum Division exists in a vacuum. Addressing the affirmative asylum backlog is not the only task to which Asylum Division officers must devote their time and resources, and within USCIS, the Division is one of many that could benefit from more funding and manpower.132

In addition to adjudicating affirmative asylum applications, Asylum Division officers are also tasked with making credible-fear determinations for foreign nationals subject to detention and expedited removal by CBP and ICE.133 The individuals about whom the Asylum Division must make a credible-fear determination (1) are being detained at the government’s expense, (2) necessarily require attention and resources to be diverted from CBP and ICE’s other enforcement priorities, and (3) are at an immediate risk of being returned to a country in which they have expressed a legitimate fear of harm or persecution. Given that, it is especially important for the Asylum Division to budget adequate resources for credible-fear determinations. The reasons are twofold: First, USCIS must ensure it has sufficient resources devoted to these determinations so as not to impede on other agencies’ functioning. A failure to efficiently make a credible-fear determination requires the agencies to continue expending resources to keep the individual detained and leaves the agencies unable to move the individual to the next proceeding, whether that be expedited removal or affirmative asylum with USCIS. Second, because these determinations involve serious humanitarian concerns—the risk that individuals may be deported to a country where they may be harmed—USCIS must make sure that these determinations are carefully handled with adequate human resources.

Additionally, in May 2022, DHS (the larger agency in which USCIS is housed) and the Department of Justice jointly published an interim final rule that shifts initial responsibility for adjudicating an asylum claim from EOIR’s immigration judges to USCIS’s asylum officers when a foreign national subject to expedited removal shows a credible fear.134 The rule provides that during the credible-fear screening process, the asylum officer will create a documented record. If the officer determines the foreign national has a credible fear, USCIS will retain jurisdiction to conduct an “Asylum Merits interview” based on the credible-fear record instead of referring the individual to removal proceedings to have her asylum claim adjudicated by an immigration judge.135 Further, even if the asylum claim is denied, the asylum officer may consider the foreign national’s eligibility for other forms of relief from removal.136 As with their original credible-fear duties, these changes make it necessary for the Asylum Division to allocate attention and resources to these claims to avoid exacerbating the delays and dysfunction already endemic to immigration courts.

Moreover, the Asylum Division is just one section in a larger agency that administers many other immigration benefits, including refugee admissions, naturalization and legal permanent residency determinations, deferred action, and family reunification programs.137 The various sections or divisions that administer these programs have their own delays, backlogs, and resource constraints, and none can demand the entirety of USCIS’s appropriations and manpower. Even here, within the limits prescribed by Congress, USCIS must balance a multitude of competing factors to determine which division should get how much. The cases discussed in Part III.A.1138 counsel that the agency—here, USCIS—is the entity most capable of making informed decisions regarding the appropriate way to optimize the use of its limited resources. Given the agency’s competing priorities and budget constraints, tradeoffs are inevitable. But by delegating authority to the agency to administer these programs, and thereby vesting significant discretion in it, Congress ensured that USCIS would have the flexibility to address the most pressing challenges first.

Two objections can be made to this argument. The first is that the remedy for unreasonable asylum-adjudication delay is not a complete reordering of agency priorities or an in-depth interrogation of how the agency pursues its goals; it is simply an order compelling expedited adjudication. I address this in Part III.B.139 The second objection relates to a line of unreasonable delay cases concerning adjustment of status (AOS),140 another immigration benefit administered by USCIS.141 In AOS cases, courts have been far more willing to find that delays are unreasonable, thereby compelling adjudication. Therefore, one can argue that if courts can mandate adjudication of AOS cases, doing so for asylum cases is not so far beyond their competencies.

In the AOS cases that courts deciding asylum cases have looked to, the reason for delay can be summarized as follows: an individual applying for legal permanent residency in the
United States was previously associated in some way with an organization or collective meeting the definition of a “Tier III terrorist organization”142 in their home country. The affiliation generally bars a foreign national from an adjustment of status, but Congress has given the agency discretionary authority to “exempt certain terrorist-related inadmissibility grounds as they relate to individual aliens.”143 In 2008, USCIS adopted a policy of withholding the adjudication of cases that might benefit from the expanded discretionary authority.144 Plaintiffs sued over the resulting wait. In many of these cases, the individual’s application had gone unadjudicated for years not because of any backlog or reasoned decision about priorities but because the application had been held indefinitely “in case [the government] might, at some unspecified point in the future, consider an exemption” to the Tier III inadmissibility bar.145

Given these facts, when courts weighed the TRAC factors in AOS cases, they generally concluded that (1) seemingly indefinite deliberation was not a rule of reason and that (2) because the delays were the product of discretionary deliberation, not reasoned decision-making about resource allocation, compelled adjudication wouldn’t unduly infringe on competing agency priorities.146 The facts which led to these conclusions about the TRAC factors in AOS cases stand in stark opposition to the reality of the asylum backlog. In the asylum context, plaintiffs are left in limbo not as a matter of pure discretion but because USCIS necessarily must prioritize one application over another for each application in the queue. Although individuals may disagree with the LIFO schedule as a policy matter, USCIS here concluded that this specific method of prioritization would be best despite the delay that necessarily accompanies any given scheme.

Herein lies the dispositive difference between the two lines of cases: The AOS cases asked the courts to decide whether USCIS’s exercise of pure discretion exceeded the bounds reasonably set by Congress in its statutory directive. This is a task for which courts are much better suited than weighing whether a particular allocation of resources is “reasonable,” “optimal,” or “correct.”147 Nevertheless, because courts recognized that USCIS was still entitled to deference and leeway to deliberate, case law developed in such a way that many courts have concluded that deliberating for five or more years is unreasonable.

B.    Separation of Powers Concerns Counsel Against Judicial Interference

Beyond the issues of institutional competence discussed in Part III.A is a closely related separation of powers problem: too much judicial intervention by way of compelled adjudication will unduly interfere with the executive branch’s prerogatives. The affirmative asylum system is just one of the various immigration programs that USCIS administers. An asylum seeker’s application is one of nearly two hundred thousand, and inherent to any adjudicative backlog is both the reality of finite resources and the decision(s) to allocate portions of those resources to other tasks. Deciding how to distribute limited human and monetary resources in an agency that suffers from backlogs in many of its programs148 —finding the right balance to strike among competing and overlapping interests—starts to look like a series of complex policy questions. As discussed in Part III.A, courts recognize that agencies are the comparatively more competent institution to make these decisions, but it is also necessary to recognize that those decisions were delegated by Congress to the executive branch, not the judiciary.

A court order compelling agency adjudication requires one of two actions: (1) forcing the agency to recalibrate the political decision of what its priorities should be by requiring this application be prioritized over another; or (2) forcing the agency to reallocate resources from other functions to maintain current priorities and fulfill the court’s order. Either outcome invariably requires the court to substitute its judgment for that of the executive branch.149 Moreover, in other contexts of unreasonable delay, the courts have noted that if the fundamental problem is a lack of funding, the “problem [is] for Congress, not the courts, to address.”150

One might assert that USCIS cannot be so strapped for resources that a single order compelling adjudication is such a vulgar display of judicial impropriety. This is true. Any single order is unlikely to have a noticeable effect, yet courts must consider how their decisions serve as precedent for similar orders in the future. Judges are acutely aware that “[i]f a judicial remedy were available to private parties every time an agency failed to act, the establishment of agency priorities might turn on who wins the race to the courthouse door.”151 Although any one court order might be insignificant, the aggregate effect of compelled adjudications would—in a sense—put the judiciary in a managerial role over how USCIS prioritizes asylum applications.152 The case
Hui Dong v. Cuccinelli153 is illustrative.

The plaintiff in Hui Dong filed suit on October 31, 2020, seeking an order compelling adjudication of her asylum claim that had been pending for just one year and two months.154 On January 15, 2021, the government moved to dismiss.155 But unlike the courts in the cases discussed in Part II.B, the district court declined to apply the TRAC test at the pleading stage, reasoning that it was “premature to rule” on a fact-intensive inquiry at the pleading stage, particularly when other courts had also refused to do so.156 Accordingly, on March 2, 2021, the court denied the motion to dismiss and allowed the case to move forward.157 Threatened with further litigation, the government and the plaintiff entered a stipulation eight days later to stay the case while USCIS moved
Hui Dong’s asylum application forward.158 Then on June 11, 2021, the plaintiff voluntarily dismissed the suit, presumably because she had received what she wanted: a decision.159

Although not a formal court order, Hui Dong’s outcome raises the same concern of incentivizing certain behavior by current and future applicants. If surviving a motion to dismiss is seen as a method of informally “compelling” adjudication, applicants will file claims and fight to survive dismissal just to pressure the agency to modify its priorities. This would result in indirect judicial interference in USCIS operations, benefitting only applicants with “the wherewithal to retain private counsel and to bring suit in District Court.”160

As stated earlier, this Comment is principally concerned with evaluating potential avenues to relief from backlog-related multiyear waits for two subsets of asylum seekers stuck in the queue: (1) asylum seekers requiring medical care who are unable to access medical services without the federal public benefits available to those granted asylum, and (2) asylum seekers wanting to be reunited in the United States with immediate relatives who are facing threats of persecution but are unable to otherwise come to the United States. There is not a method of which I am aware to ascertain what percentage of total applicants belong to these subsets. Accordingly, it is irresponsible to ignore the risk of too much court interference in agency business when faced with uncertainty over how many individuals might fall into one or both of the groups deserving of prioritized adjudication. Courts might then be unlikely to grant relief out of fear that many other similarly situated applicants would file suit, bringing the worry of judicial establishment of agency priorities to fruition.

Against this Section’s contentions, a final argument can be levied that separation of powers considerations actually cut the other way, in that the judiciary should be playing an important role as a check on the executive branch’s failures to act. Absent malfeasance or gross negligence,161 I contend that several considerations support the position staked out in this Section. First is the evaluation of comparative institutional competencies, discussed earlier, which favors the agency as the proper decision maker. Second, Congress has explicitly delegated authority to execute and manage this program to a body of the executive branch, and the congressional delegations for administering the asylum system are quite broad.162 Third, implementing duly enacted legislation—particularly in a complex scheme like the asylum program—requires discretionary decision-making and filling in policy gaps. Here, the duty to “take Care that the Laws be faithfully executed” is entrusted to the executive branch, not to the judiciary.163 And finally, because decisions about how to prioritize set tasks given limited resources are, at the end of the day, policy decisions, the more flexible and accountable coordinate branch should probably hold the reins.

The concerns of institutional competency and separation of powers, taken together, illustrate that the judiciary is simply a deficient institution for dealing with the increased likelihood of harm that the identified subsets suffer on account of the backlog and long wait times. That said, all human-made institutions are in some respect deficient, so the ultimate question is not one of institutional competency in absolute terms but one of competency as compared to what, to whom.164 This Part has endeavored to show that the requests lawyers and asylum seekers make of courts in these asylum-context unreasonable delay claims are not ones the judiciary can effectively answer. Yet, if the courtroom is the best available way forward despite all its shortcomings, then continued litigation efforts are undoubtedly warranted. The path of the law does not end where the problems become quarrelsome and thorny. Nevertheless, the cases discussed in this Part suggest that there is another institutional actor better suited to directly confront the issue. To this actor I now turn.

IV.  A Different Path Through the Red Sea: Looking to Agency-Side Solutions

Underlying the proposition that courts should defer to agencies’ decisions concerning resource allocation is the premise that agencies are simply better at balancing the relevant considerations and coming to reasoned decisions.165 In Part III, I argued that federal court litigation is a far from optimal way of addressing the harms that individuals in the specified subsets face, but the purpose of this intervention is not simply to trash existing solutions or institutions. The purpose is instead to explore what potential avenues to relief our current law and system of governance might have to offer.

Open-ended congressional delegations of authority allow agencies to adjust existing internal policies and create new procedural and substantive rules to respond to new problems and directives. The problems of mass-adjudication delay are not unique to the immigration system. So just as other agencies—drawing on institutional expertise and new technological developments—have responded to mass-adjudication backlogs through novel substantive and procedural policymaking, USCIS should do the same to provide a reliable and consistent avenue for relief.

In this Part, I propose a new agency-side adjudicative mechanism using artificial intelligence (AI) that USCIS should implement to create a better—albeit imperfect—remedy for asylum seekers especially aggrieved by long wait times. Part IV.A briefly reviews relevant uses of AI in federal administrative agencies. Part IV.B sketches out how the novel adjudicative mechanism may be implemented with AI in USCIS, responds to some counterarguments, and argues why this mechanism should be the agency’s first foray into AI-driven adjudication. Part IV.C assesses the efficacy of existing legal mechanisms to ensure transparency and accountability when agencies use AI in “core agency functions.”166 Finally, Part IV.D closes by reflecting on the comparative advantages of an agency-side solution to federal court litigation.

A.    Artificial Intelligence in Administrative Adjudications

With rapid advancements in artificial intelligence technology, AI “has begun to permeate many aspects of U.S. society”167 and “promises to transform how government agencies do their work . . . [by] reduc[ing] the cost of core governance functions, improv[ing] the quality of decisions, and unleash[ing] the power of administrative data.”168 This Comment uses “AI” to denote, specifically, recent forms of machine learning (ML) that train models to learn from data through a range of methods capable of “recognizing patterns in a range of types of data.”169 Although there is a large private-public gap in the use and development of AI,170 scholarship shows that federal and state governments’ AI toolkits are diverse and that a number of agencies throughout the federal administrative state now use AI in “core” agency functions.171 Among those “core” functions is the administrative adjudication of government benefits and privileges.

The Social Security Administration (SSA) provides two relevant examples of mass agency adjudication augmented by AI, and a brief discussion of them will lead to an enhanced understanding of Part IV.B’s proposal. The SSA is likely “the largest adjudication agency in the western world” and among its tasks is the administration of two disability insurance programs.172 The SSA offices that are charged with scheduling disability hearings, handling appeals, and reviewing administrative law judge (ALJ) decisions often “experience a significant backlog of claims.”173 In order to expedite claims, the SSA over the past decade has developed two AI-driven mechanisms that have dual goals of (1) reducing case processing times and (2) identifying claimants in greater need of faster processing because they are most likely to have a disability and thereby qualify for benefits.

The first system I discuss is the SSA’s “Quick Disability Determination” (QDD) system—an AI model that “identifies claims where benefits are likely to be awarded and where the information needed to make the disability determination can be obtained quickly.”174 The QDD model identifies claims by using scores based on factors like treatment protocols, medical history, and symptoms.175 The model makes predictions based on features selected from prior human-made determinations.176 Predictions made by the QDD are then reviewed by a human examiner.177 The second AI tool developed and employed by the SSA is used to expedite claims at the ALJ hearing stage by predicting which claims that were denied on reconsideration have a high likelihood of ultimately qualifying for benefits. This model evaluates both procedural features, like outcomes at the hearing level, and personal features, like age and impairment, to make its predictions. Cases with higher probabilities are moved ahead in the queue “under the notion that disabled claimants should receive their decisions as soon as possible.”178

The SSA deploys a few other AI tools to facilitate its daunting adjudicative tasks, but these two specifically demonstrate that the use of AI to address the possibilities of harm exacerbated by adjudicative backlog not only is a feasible endeavor but also can be a successful one.179

B.    A Second Track: Petitions for Reprioritization

This Section proposes a new, public-facing adjudicative mechanism within USCIS that would be better than federal court litigation at securing relief for the identified subsets of asylum seekers. Part IV.B.1 sketches the basic contours of the new mechanism through which asylum seekers should be able to petition for reprioritization within the existing asylum queue. Part IV.B.2 then details how AI technology may be incorporated in the implementation of the proposal contained in Part IV.B.1. Specifically, it discusses some technical details of the AI model as applied to the novel mechanism and argues why this proposal should be USCIS’s first experimentation with sophisticated AI technology in mass adjudication.

1.   Basic structure of the reprioritization mechanism.

I propose that USCIS institute a new public-facing mechanism, or “track,” exclusively for asylum seekers falling within the specified subsets to petition for expedited adjudication of their asylum applications. To be clear, this is not an adjudication on the merits of the underlying claim. The adjudication is to qualify the asylum seeker for reprioritization within USCIS’s existing priority schedule. Under the current LIFO schedule, there are four priority categories, with rescheduled interviews and applications filed in the past twenty-one days being the highest priorities.180 Individuals likely to fall within the specified subsets and to take advantage of my proposed mechanism likely belong to categories three and four, and a favorable adjudication under the reprioritization mechanism would move them up to the priority two category.181

Although the reprioritization mechanism currently does not exist, its proposed contours—as will be discussed shortly—should be familiar to USCIS and thus would benefit from the agency’s institutional expertise. Under this proposed system, an individual would file a reprioritization petition with USCIS that specifies which of the two harms she suffers from, provides details about her claim, and includes evidence corroborating her claim. The type of the petition, the forms of evidence considered, and the nature of the decision-making would mimic those of the informal adjudications in which USCIS already engages for other immigration-related benefits. The most analogous is the currently existing “expedite process” system.182 Under the expedite request system, applicants to most of the programs that USCIS administers may request expedited processing of their application by communicating to the agency the basis of their request, explaining why their circumstances warrant expedition, and including evidence corroborating the claim for the need of faster processing.183 USCIS then considers all the information before it to decide whether a favorable exercise of discretion is in order without going through a full hearing or making specific findings of fact. These are the features of USCIS’s existing informal adjudications on which elements of this mechanism would be modeled. But to be clear, the proposed reprioritization mechanism is different from the expedite request system, and Part IV.B.2 further elaborates on this distinction and the advantages of the reprioritization mechanism specifically.

The actual petition for the reprioritization mechanism would be comprised of two substantive components. The first would require the applicant to provide basic biographical information and information pertaining to the underlying asylum application, such as the date of filing, receipt number, brief restatement of the claim, etc. The second component would be bifurcated into two sections relating to the two harms that provide grounds for reprioritization: the inability to access medical services needed to treat a current ailment and the inability to petition to have immediate relatives facing persecution in their countries of origin be reunited with the individual here in the United States. Applicants may petition under one or both bases. Each section would contain questions designed to gather as many details as possible about the claimed harm to be analyzed by the AI model.184 Further, applicants would be required to include evidence corroborating their claims and answers to the greatest extent they could. Such evidence might include, but would not be limited to, country-conditions reports, affidavits, declarations, witnessed or notarized signed statements, medical records, or physician evaluations.

The decision on the reprioritization petition would not be based on the merits or likelihood of success of the petitioner’s underlying asylum claim. Instead, the decision would turn on whether the evidence demonstrates a high probability that the individual suffers from one of the two identified harms and thus warrants reprioritization. If the applicant satisfies her burden, her original I-589 would be moved up to what is now Priority Two under LIFO. If she fails, her application would remain where it is in the queue, but she could submit another reprioritization petition with additional information and/or evidence for reconsideration at any time after the original denial. To prevent individuals from flooding USCIS with frivolous reapplications, asylum seekers would be limited to one resubmission. Further, any misrepresentations or suspect claims made by the applicant here could be used by the asylum officer to reach an adverse credibility finding when deciding on the merits of the underlying claim.

2.   Using artificial intelligence to implement the reprioritization mechanism.

Having described the broad contours of reprioritization in the previous Section, the following sections dive into the possibility of implementing the system with AI and the host of complexities that it would entail. Specifically, Part IV.B.2.a sketches how such a mechanism could be implemented with AI technology. Part IV.B.2.b then argues (1) that USCIS’s accumulated institutional expertise could be brought to bear on making the AI model successful and (2) that the reprioritization mechanism—as opposed to other USCIS adjudicative processes—is the best subject for the agency’s first use of sophisticated AI technology in a core agency function.

a. Logistics behind using artificial intelligence.  A comprehensive agency response like the reprioritization mechanism would entail substantial costs upfront and over time. A major concern stems from the very problem of the backlog: Would the mechanism produce benefits greater than the costs incurred by potential exacerbation of the asylum backlog, or would it simply create another mountain of applications through which no one has the time to sift? Here is where introducing AI would be most helpful and innovative. This Section attempts to outline how the proposed reprioritization system could be implemented with AI and address the hurdles attendant to implementation.185

First, for the AI system to make decisions on petitions, all—or nearly all—of each petition would need to be submitted electronically. For the model to function, the raw data must be structured in such a way that makes it analyzable by the model.186 While some agencies use ML to support automatic handwriting detection,187 this would be unwieldy when applied to hundreds of applications and supporting documents—either handwritten or typed—submitted on paper. Thus, use and modification of USCIS’s online portal for electronic filings would make this AI-infused mechanism more effective.188 Further, because supplementary evidence, such as medical records or affidavits, would likely be submitted in word processing or PDF formats, significant efforts and costs would be required to “extract, standardize, and validate such information from [the] raw records.”189 USCIS would not be starting from scratch in this endeavor, though; it could likely make use of the programs already developed and used by other agencies, like the SSA.190

Second, unlike the SSA models described in Part IV.A, the reprioritization model would not predict the likelihood of success on the merits. To reprioritize on that basis would defeat the purpose of LIFO, and deterrence of frivolous claims is, for purposes of this Comment, assumed to be a legitimate government interest. Instead, the algorithm would have to predict something akin to a “high likelihood of harm to applicant” or the “existence of a pressing humanitarian concern,” as tailored to the two identified subsets. Because these are abstract concepts that cannot be measured directly, proxies are necessary.191 This is where USCIS’s institutional expertise and the existence of similar adjudicatory procedures would be of great importance.

b. Drawing on institutional expertise and why the reprioritization mechanism should be USCIS’s first dance with sophisticated artificial intelligence.  As mentioned earlier, USCIS already has an “expedite request” process192 that allows applicants to petition for expedited processing of an application for one of the many immigration benefits falling within the agency’s purview.193 Among the criteria considered by the agency is the broad category of “emergencies and urgent humanitarian reasons,”194 which is also used as a criterion for granting relief in other circumstances.195 Because of these expedite requests, USCIS already has substantial experience considering and making decisions on similar claims with the same sorts of evidence that should be present in the reprioritization claims. Analyzing prior expedite request decisions would allow USCIS to identify the factors that the model would use in its predictions, isolating the data from which the model can “learn.”

This would both lower the costs of implementing the mechanism and make it more likely that petitions evidencing a high risk of harm would be identified and reprioritized. That said, relying on decisions and factors from a different—yet analogous—adjudicatory process could create the problem of a mismatch between the patterns from prior analyses and current, slightly different goals. I discuss how this should be addressed in greater detail in Part IV.C.3.

A potential challenge to my proposal is that because the proposed AI reprioritization mechanism would draw on data from the existing expedite system, it makes more sense to simply integrate AI into the expedite system itself. For the following reasons, this Comment rejects that contention and asserts that USCIS’s first foray into AI technology should be through the reprioritization mechanism. Ideally, AI would eventually be used in the expedite system and wherever else possible in USCIS’s processes to increase case management efficiency and decrease wait times without sacrificing procedural safeguards or the care with which each claim should be decided. But first testing and using AI in the reprioritization mechanism—which is more narrowly tailored—would be the wisest first step toward agencywide use of AI.

The expedite system is available for nearly all immigration benefits administered by USCIS. As a result, the agency receives many petitions in number and kind. Therefore, the costs of creating, training, and implementing an AI algorithm would be substantially greater for expedite requests than for petitions for reprioritization within the asylum queue. These greater costs are good reason for pause when potential hiccups and efficacy are still unknown. If first applied in the asylum context, the model would consider a smaller range of factors and would have to make a narrower range of predictions. The model would also adjudicate fewer petitions than the expedite system in absolute terms. USCIS would then be better situated to learn from and refine the system because (1) it is in one sense “simpler” than a model would need to be in the expedite request context, and (2) the costs of failure or malfunction would be significantly smaller. Moreover, because USCIS would have already built up its internal technical capacity through the creation of the reprioritization model, the agency would be able to draw on that capacity when developing new models and expanding its AI use.196

Finally, the pilot program for AI in USCIS should be for pending asylum claims rather than for applicants in another queue. Because of the usually tumultuous circumstances surrounding individuals’ flights from their home countries and the reality that asylum seekers remain in the country for years without legitimate legal status, asylum seekers generally face acute humanitarian concerns. If USCIS is to start experimenting with AI in adjudication, that capacity should be used for those who are most likely to need expedient adjudication of their claims.

Building technical infrastructure and data capacity for this mechanism, and any future AI mechanisms, could strain agency resources. Building technical personnel capacity, too, would push the agency to “grapple with budgetary and other human resource constraints.”197 USCIS is, for the most part, a fee-funded agency,198 but requesting congressional appropriations would likely be necessary for the implementation of this mechanism, as well as for any future expansion of the agency’s use of AI. These steep upfront costs, however, could very well be worth the long-term gains of increased processing efficiency and consistency in adjudication.199 Finally, there is a tension between the proposed reprioritization mechanism and the overarching goal of eliminating the asylum backlog. While instituting the reprioritization system might exacerbate the backlog to an unknown extent, it would serve the important function of providing relief to those who would otherwise suffer the most from the backlog. The use of AI would minimize the severity of this trade-off and would plant the seed for more expansive use of AI in USCIS.

C.    Transparency and Accountability with Automaton Administration

The increasing salience of AI in government has raised serious legal questions for which—unsurprisingly—there are no clear answers. Much of current scholarship has focused on broad constitutional questions raised by AI-driven governance,200 but I am more interested in what current administrative law has to say—as it is this body of law that will govern day-to-day implementation, expansion, and refinement.

Administrative law is “premised on transparency, accountability, and reason-giving.”201 Current doctrines support these premises in two principal ways: (1) ex post judicial review, and (2) ex ante notice-and-comment rulemaking. Because of the nature of judicial review of agency decision-making, reason-giving—offering a justification for the relevant conduct—by administrative agencies is a principle permeating U.S. administrative law.202 Reason-giving can also occur even if the agency doesn’t think it’ll be dragged into court. For instance, although most of the benefits administered by USCIS are discretionary,203 the agency must still provide its reasons for denial, including an explanation of factors that tilted in the applicant’s favor and led to an unfavorable
decision.204

With the reprioritization mechanism, however, it is unclear to what extent proper transparency and accountability could be obtained under those two methods. Parts IV.C.1 and C.2 examine how well-suited judicial review and notice-and-comment procedures are for ensuring transparency and accountability in the reprioritization mechanism. Then, drawing on recent literature, Part IV.C.3 argues that the novel proposal of “prospective benchmarking” is particularly effective for the reprioritization system.

1.  Judicial review.

Under current law, it is unlikely that transparency and accountability in the AI reprioritization mechanism could be ensured through judicial review. The first reason is that it is unclear if the model would itself ever be challenged and reviewed. For successful petitioners, the reprioritization would likely “consummate[ ] the agency’s decision process,” and those individuals would have little reason to challenge the mechanism.205

Rejected petitioners might be more likely to challenge the model, but the harmless-error rule would probably shield it from judicial scrutiny.206 Although they don’t involve AI, several cases challenging an SSA ALJ’s failure to classify certain disability claims as “critical” for expedited processing for disability determination demonstrate this point. In Webb ex rel. Z.D. v. Colvin,207 the plaintiff’s disability claim had been denied by the ALJ after the ALJ had previously declined to classify it for expedited processing. Before the district court, the plaintiff argued that the ALJ’s failure to select her claim for expedited processing was “plain legal error” and warranted reversal of the ultimate disability determination.208 In Bowers v. Saul,209 the ALJ did not classify the case as eligible for expedited processing even though the category of qualifying conditions that the ALJ should have consulted did include the plaintiff’s condition. Here, too, the plaintiff argued that the ALJ “committed reversible error by failing to hear Plaintiff’s case in accordance with” the expedite program.210 In both cases, the courts refused to reverse the ALJ’s disability determination because plaintiffs had failed to show that the procedural misstep was prejudicial to their disability claims.211 The courts explained that simply not receiving faster processing was insufficient to establish prejudice; prejudice would have had to be demonstrated by showing that classification for expedited processing “might have led to a different decision of disability.”212

If judicial review of a failure to categorize a claim for expedited processing simply turns into review of the merits of the claim itself, judicial review of the reprioritization mechanism and the underlying asylum application are unlikely. First, decisions under USCIS’s analogous expedite request system are “within the sole discretion of USCIS.”213 I have failed to find a case in which USCIS expedite decision-making was itself challenged,214 and one court has noted that USCIS guidance indicates there is no opportunity to receive judicial review of a denial.215 It is likely a similar phenomenon would occur with the reprioritization mechanism where plaintiffs challenge the delay, not the decision. Second, a denial by USCIS of the underlying asylum claim of a foreign national otherwise without legal status results in an “automatic appeal”216 to an immigration court217 —where decisions are made by an immigration judge and reviewed on further appeal by the Board of Immigration Appeals (BIA). After all of that, if a foreign national appealed the BIA’s decision, it is unlikely that a federal court of appeals performing a merits review would reevaluate the reprioritization denial.

The second reason why judicial review would inadequately ensure transparency and accountability in the AI mechanism is that even if the model were judicially reviewed, reviewing courts would run into serious informational and competency problems. Because the model’s decisions would be embedded in code, litigants and judges alike may be unable to understand how the algorithm works in practice, frustrating effective review.218 Further, because the model’s operations “may only become intelligible” with the underlying data from which it was trained,219 there may be issues in gaining the decisional input data. And this assumes the model and the data from which it is learning remain static. If the algorithm were one that continued to learn from new decisions, the challenged model could be different from the one that the court ultimately scrutinizes months, if not years, later.220 Lastly, even if courts had perfect access to the algorithm and data, the model’s sophistication often makes it a black box such that even the engineers who programmed it often cannot understand its decision-making—and generalist judges are likely to understand it even less.221  

2.  Notice-and-comment procedures.

Under current doctrine, it is unclear if the reprioritization mechanism would need to go through notice-and-comment rulemaking or if it is exempt from such procedures for being a rule of agency procedure. Either result necessarily entails its own costs, benefits, and complications. The following discussion first explores whether the mechanism must legally go through notice-and-comment. It then describes the potential benefits and drawbacks of either outcome before suggesting a plausible middle path.

Agencies use authority delegated through statutes to create rules related to a statute’s enforcement,222 and in that broad category of “rule,” those that “carry the ‘force and effect of law’” are known as legislative rules.223 Legislative rules, under which the “agency and regulated parties are bound,” must be made pursuant to the procedures set forth in the APA.224 Notice-and-comment rulemaking is the most common type of rulemaking in the modern administrative state.225 Notice-and-comment procedures require (1) a notice of proposed rulemaking in the Federal Register containing certain information;226 (2) opportunity for interested members of the public to submit data, views, arguments, or comments;227 and (3) that the agency consider and respond to relevant submitted materials in the process of promulgating the final rule.228 To rescind a legislative rule created through notice-and-comment, an agency generally must go through the same process again.229

For many, the opportunity for comment and the obligation that agencies review and at least respond to relevant comments are seen as a plus in so far as it reintroduces an element of democratic public participation in the administrative process.230 Yet, given how extensive procedures may be in practice as a result of the increasing complexity of modern governance, the burdensome nature of notice-and-comment procedures and their potential to ossify the law must be recognized.231

However, some rules are nonlegislative, and thus their promulgation does not require adherence to the rulemaking procedures put forth in the APA.232 The APA delineates certain exempt categories, and if a particular rule the agency wants to issue falls within an exception, then it could be promulgated without the need to receive and consider public comments.233 With the opportunity to avoid costly notice-and-comment procedures, agencies have an incentive to try and twist substantive legislative rules into the visage of a nonlegislative rule.234 Moreover, an agency can rescind the rule without going through rulemaking procedures if the rule qualifies for an exception. The exception into which the reprioritization may most likely qualify is the exception for “rules of agency organization, procedure, or practice.”235

Given that the distinction between procedure and substance is not clear, the doctrines developed by lower courts to distinguish the two are “enshrouded in considerable smog.”236 Of the various formulations courts have concocted, the two most relevant are broadly distilled to: (1) whether the rule diminishes discretion and has a binding effect on the agency,237 and (2) whether the rule substantially alters the rights or interests of regulated parties.238

Under the first test, the reprioritization mechanism would seem to be a legislative rule because it—in a sense—curtails USCIS’s discretion. Because the model would make a yes/no decision and flag the underlying I-589 for reprioritization if the petitioner were successful, USCIS’s discretion over each petition would be diminished, unlike expedite request procedures in which a human decides each request. It should be noted that the SSA’s QDD system did go through notice-and-comment procedures, but that was because it required the amendment of existing rules.239 Yet the QDD system might be distinguished as having more human involvement and discretion throughout the process240 whereas the only human involvement in the reprioritization mechanism for most petitions would be during the final decision on the merits of the underlying asylum claim.

One could argue that under the second test the reprioritization mechanism would substantially alter regulated parties’ rights in that many individuals waiting in the queue—but outside the specified subsets—would have to wait longer for a decision because successful petitioners get to cut in line. As discussed in Part I.B, although the 1995 reforms to the asylum system were effective in reducing the backlog, they also disproportionately and  negatively affected certain applicants—namely applicants who filed earlier in time. We might be similarly concerned here about creating a system that not only would cause groups of asylum seekers in the queue to wait longer but also would selectively confer benefits at the expense of others rather than being a bright-line rule like LIFO. This more discriminatory effect may be sufficient for some courts to conclude that the rule would substantially affect the rights of regulated parties and thus should go through notice-and-comment. On the other hand, it does not seem as though the expedite request system was instituted through informal rulemaking, as I can find no final rules or notices of rulemaking in the Code of Federal Regulations or Federal Register.241 The expedite request process poses the same problem of longer wait times because of queue-jumping to other applicants, but perhaps the fact that USCIS still retains human discretion over individual requests cuts in favor of the rule being considered procedural in nature. LIFO, too, forces many applicants to wait considerably longer, but again, USCIS retains discretion to modify it, which makes it seem more like it alters not the rights of the parties but “the manner in which the parties present themselves . . . to the agency.”242

Closely related to the doctrinal question of whether the proposed mechanism would be a legislative rule or a procedural nonlegislative rule are the policy reasons why one set of restrictions and benefits are better than the other.

Having the mechanism go through notice and comment, even if such a process isn’t legally required, could produce two important benefits. First, it would entrench the mechanism because its alteration or repeal would also need to also go through notice and comment, which is a costly process. Though USCIS is traditionally a service provider rather than an enforcement agency,243 policies of recent presidential administrations—such as “metering,”244 the Title 42 border expulsion policy,245 the “public charge” rule,246 or redirecting asylum officers to other tasks247 —demonstrate the executive branch’s ability to willfully thwart the humanitarian aspects of the immigration system. Implementing the AI reprioritization mechanism through notice and comment makes it much more difficult for hostile administrations to arbitrarily tamper with it because recission or modification through notice and comment are costly248 and poorly reasoned rules are more easily challenged.249

Moreover, there is a second important benefit. A “major purpose[ ] served by the notice and comment procedure is to assist the agency in obtaining information” from the “important, perhaps unique” sources of information who are affected parties.250 Accordingly, an opportunity for public comment may very well lead to a much better mechanism, particularly with feedback from advocates in the fields of refugee or asylum law, scholars, or other private entities who may be interested in promoting the development of AI technology in governance. Legal advocates would be better positioned than I am to provide more accurate information on the scope of the problem, evidence that can or cannot be reasonably provided for the reprioritization determination, and how to best appraise asylum seekers of the mechanism.

However, the fact that changing the rule is so costly may be a silver bullet where AI models are involved. Because the model is dynamic and continually “fed” data points from which to learn,251 it is in a constant state of modification. It is beyond the pale to think each change to the rule being implemented (the model) would really trigger a new round of notice and comment. Further, USCIS may want a lot of flexibility to tinker with the model and its implementation in the beginning to work through kinks and complications. These considerations strongly favor viewing the mechanism as a procedural rule—one concerning “the manner in which the parties present themselves . . . to the agency”252 —and avoiding the notice-and-comment debacle in implementation, lest refinement of this mechanism and future experimentation with AI by USCIS be stymied by repeated notice-and-comment procedures.

There may, however, be a middle path that reaps some of the benefits of notice and comment without running into the undue restrictions on USCIS’s ability to adapt: the creation of a new adjudicative track for the identified subsets of asylum seekers at a high level could go through notice and comment, while the way that the mechanism is executed—with AI technology—could be excepted as a nonlegislative rule of agency procedure.253 This reaps the benefits of public participation and large-scale entrenchment without compromising the latitude USCIS will need in order to develop sophisticated AI technology in a pilot program. There is a loss, however, in potential informational benefits from comments from AI scholars and experts that would not come before the agency if the AI component is treated as an internal procedural matter after the rulemaking period.

3.   “Benchmarking”: keeping artificial intelligence accountable.

In a 2020 paper on AI and administrative law, Professors David Engstrom and Daniel Ho sketched a novel approach to maintaining transparency and accountability in agency deployment of AI. Recognizing the deficiencies of current legal tools, Freeman and Ho proposed “prospective benchmarking”—the process of using sample petitions, adjudicated by humans, as a “control group” against which AI-decided petitions could be measured—as a viable answer. This Section concludes by arguing that benchmarking would serve USCIS well in implementing the reprioritization mechanism and should therefore be implemented voluntarily by USCIS.

Benchmarking would entail setting aside a randomly selected sample of petitions on which “conventional human decision-making would be deployed” to serve as a comparison group to the rest, and likely majority, of petitions that are adjudicated by AI.254 Because adjudication for reprioritization is similar to a lot of adjudication that USCIS already does, mandating that a small group of petitions be benchmarked would pose very small personnel and training costs in comparison to the potential benefits.

Benchmarking would facilitate the validation and refinement of the reprioritization model. Because comparator data points would be generated consistently over time, USCIS could both better judge the model’s accuracy over time and recalibrate it to reflect changes in policy or real-world circumstances.255 Also, because USCIS would have comparator information across time and recalibrations, this would make future implementation of AI in other areas of agency activity easier and more likely to be effective quicker. Benchmarking would also create new data points from which the model could continue updating or “learning.” As I mentioned in Part IV.B.2, the model would originally learn from expedite request decisions, which are not a perfect fit for the reprioritization system; having human reprioritization decisions as input data would make the model’s predictions more accurate and responsive to changes in the state of real-world affairs.256 Lastly, benchmarking would more quickly alert USCIS to inaccuracies or biases in the model’s predictions,257 which is particularly important when the difference between reprioritization or not could be years for an applicant who doesn’t have years to spare.

The principal goal of this intervention is to propose a viable method by which certain especially vulnerable asylum seekers may gain relief from long wait times created by the current backlog. To meet this goal, it is important to ensure its efficacy, its accuracy and continued development, and its conformity with prevailing administrative law principles of transparency and accountability. Part IV.C.1 strongly concludes that ex post judicial review is inadequate. Part IV.C.2 suggests that notice and comment may be better in this regard, but that the law provides few clear answers in this area, and that any plausible choice requires balancing the values of additional information and flexibility. Part IV.C.3 argues that prospective benchmarking is a sensible way not only of ensuring the proper development and functioning of the model but also of reinserting transparency and human accountability back into the process.

D.   Comparative Advantages of a Centralized Agency-Side Remedy to Federal Court Litigation

Thus far, this Part endeavored to demonstrate how a viable path to relief may be achieved through an agency-side solution that incorporates AI technology, and in doing so, Part IV.C addressed how existing administrative law doctrines can, or cannot, help ensure a fair and functioning mechanism. This Section returns to the claim made in Part III.B that an agency-side solution is comparatively superior to federal court litigation—even considering the difficulties discussed in Part IV.B.2 and C.

Despite the upfront costs associated with the proposed system,258 the reprioritization mechanism should eventually create a consistent and lower-cost forum for addressing this problem than litigation in federal court. First, the mechanism would allow USCIS to continue its policy of deterring frivolous applications through LIFO while still giving due attention to asylum seekers’ unique humanitarian circumstances. It would emulate adjudication procedures at which USCIS is already adept and use information to which USCIS already has access. Additionally, the “stakes” of adjudication would be lower because an immigration benefit is not what would be granted. Because there is a good chance that the mechanism would not need to go through notice-and-comment, and because the agency would use benchmarking, an agency-side solution would give USCIS flexibility to adjust as it learns more about the process in practice and the underlying problem.

Second, the costs of using the reprioritization mechanism would be lower for asylum seekers and legal service providers compared to using litigation. Federal court litigation is by no means a cheap enterprise.259 The costs of finding and hiring a competent attorney are great for disadvantaged communities, there is a nationwide dearth of counsel available to represent impoverished persons,260 and immigrants generally face higher barriers to accessing adequate representation and justice.261 Many of these costs can be avoided if asylum seekers are able to seek relief through USCIS, an entity with which they are already familiar by having filed their original I-589. For instance, a fee payment is not required to file an I-589,262 and this proposal advocates that this also be the case for the reprioritization mechanism. That said, there are real expenses that asylum seekers would need to incur to gather the necessary corroborating evidence, but those same costs would also be incurred in litigation.

Third, the reprioritization mechanism would do away with the need for asylum seekers to either seek litigation attorneys (many low-cost or free legal service providers for immigrants, refugees, and asylum seekers don’t do federal court litigation) or attempt to litigate pro se. Instead, the mechanism would mimic those USCIS processes that most low-cost and free legal-service providers already deal with in the course of servicing asylum-seeking populations.263 Thus, such legal-service providers would be able to help clients with reprioritization without extensive additional training. Further, the ease with which existing legal-services providers could adapt to the new mechanism would lead to representation for reprioritization likely being more widely available and accessible than representation for litigation. Lastly, even if an applicant were to petition unrepresented, she would be more likely to get relief through the reprioritization mechanism than litigation due to the variety of self-help or pro se manuals from different organizations that are available to all persons and could be modified to account for the new mechanism.264

Conclusion

The U.S. asylum process “is in tatters.”265 Although USCIS and Congress have taken various steps to adjust the system periodically, the backlog still exists; individuals wait for years for an interview, let alone a decision, and suffer as a result. To address the system’s overarching problems and provide relief to the many asylum seekers outside the subsets I have identified, comprehensive reform by the political branches and diligence by the courts will be necessary.

However, to sit around and wait for this legislative panacea to come is a fool’s errand. Many asylum seekers face immediate harm and cannot wait for legislative salvation. This intervention has attempted to show that continued reliance on federal court litigation will not properly address these harms because resolution of those harms is not a task for which courts are competent. But I have also tried to show that the adaptability of administrative agencies and the recent strides in artificial intelligence technology behoove us to not yet despair. Beyond simply outlining a specialized form of relief to address the unique humanitarian concerns of certain asylum seekers, my proposal endeavors to kickstart a larger conversation between agencies, academics, and lawyers about how new technology may be used to start solving the problems endemic to our immigration system.

• 1U.N. High Comm’r for Refugees, Mid-Year Trends 2021: Forcibly Displaced Population 15.
• 2See, e.g., Sabrineh Ardalan, Access to Justice for Asylum Seekers: Developing an Effective Model of Holistic Asylum Representation, 48 U. Mich. J.L. Reform 1001, 1001–02 (2015); Tania Guerrero, “Out of Sight, Out of Mind:” Six Stories of Asylum Seekers and Migrants Under MPP, CLINIC (Jan. 28, 2020), https://perma.cc/3R4V-QVFM. Content warning: the stories in this report include descriptions of sexual assault.
• 3U.N. High Comm’r for Refugees, Handbook on Procedures and Criteria for Determining Refugee Status and Guidelines on International Protection ¶ 196 (2019).
• 48 U.S.C. § 1158(a) provides for “affirmative” asylum claims, and 8 U.S.C. § 1225(b)(1)(A)(ii) provides for “defensive” asylum claims. The Immigration and Nationality Act (INA), Pub. L. No. 82-414, 66 Stat. 163 (1952) (codified as amended at 8 U.S.C. §§ 1101–1537), does not use the terms “affirmative” or “defensive”; these are labels used by practitioners, scholars, and agencies. An individual can invoke a defensive asylum claim in removal proceedings as a defense to deportation. Defensive asylum claims are filed with the individual’s assigned immigration judge. Further, defensive asylum is the business of the Executive Office for Immigration Review (EOIR) in the Department of Justice. Alternatively, an individual who is not in removal proceedings can apply to the U.S. Citizenship and Immigration Services (USCIS) with an affirmative asylum claim within one year of their most recent entry into the country, absent exceptional circumstances. See Asylum in the United States, AIC (June 11, 2020), https://perma.cc/UL6K-BT7T [hereinafter Asylum in the United States]. But see infra Part III.A.1 (discussing a new regulation that would require the USCIS to adjudicate defensive asylum claims in the first instance). This Comment is principally concerned with affirmative asylum, which is discussed in further detail in Part I.A.
• 5Dep’t of Homeland Sec., Yearbook of Immigration Statistics 2020, Table 16 (Apr. 6, 2022), https://www.dhs.gov/immigration-statistics/yearbook/2020.
• 6Id. at Table 17d.
• 7Exec. Off. for Immigr. Rev., Executive Office for Immigration Review Adjudication Statistics: Affirmative Asylum Applications, Dep’t of Just. (Apr. 18, 2022), https://perma.cc/9T2D-Q3MZ [hereinafter Adjudication Statistics].
• 8Hum. Rts. First, Protection Postponed: Asylum Office Backlogs Cause Suffering, Separate Families, and Undermine Integration 5 (2021) [hereinafter Protection Postponed].
• 9Id. at 6. As noted in the report, “Aaron” is a pseudonym used to protect the individual’s privacy and safety. Id. at 6 n.1.
• 10Id. 
• 11Id. The report was published in April 2021, so it is unknown whether Ibrahim has received his interview yet or if he is now in his seventh year of waiting. Like “Aaron,” “Ibrahim” is a pseudonym. Id. at 6 n.1.
• 12See id.
• 13I use the term “foreign national” as synonymous with “alien” as defined in 8 U.S.C. § 1101(a)(3). When quoting sources, I use “alien” as the text itself reads.
• 14See Asylum, USCIS (Nov. 29, 2021), https://perma.cc/3548-DBVF.
• 15But see Protection Postponed, supra note 8, at 5–10 (collecting stories of the harms individual asylum seekers have suffered from prolonged wait times).
• 16This process is known as “derivative asylum” because the family members do not need to have an independent asylum claim that must be adjudicated on the merits—rather, their asylum status is derived from the principal asylee’s and remains so as long as the principal asylee maintains her status. See Family of Refugees and Asylees, USCIS (Sept. 25, 2017), https://perma.cc/J65Z-Q449. Qualifying immediate relatives are (1) spouses and (2) unmarried children who were under the age of twenty-one when the principal asylee first applied for asylum. See id.; I-730, Refugee/Asylee Relative Petition, USCIS (Apr. 29, 2022), https://perma.cc/UQ4T-HKYZ. Exploration of the family reunification process is outside this Comment’s scope, but it is worth noting that the derivative asylum petition process is both generally easier—because the petitioner need only prove the relationship, not persecution—and faster than a principal’s claim for asylum. Cf. Check Case Processing Times, USCIS, https://perma.cc/LF3W-V6LJ (noting that, on a first-come-first-processed basis, it takes about twenty-four months for the Texas Service Center and 12.5 months for the Nebraska Service Center, as of May 24, 2020, to process I-730 Refugee/Asylee Relative Petitions).
• 17Cf. Humanitarian or Significant Public Benefit Parole for Individuals Outside the United States, USCIS (Nov. 12, 2021), https://perma.cc/WL32-KPPG (noting that urgent humanitarian reasons for parole may include “time-sensitive action” like “critical medical treatment”); Guidance on Evidence for Certain Types of Humanitarian or Significant Public Benefit Parole Requests, USCIS (Dec. 8, 2021), https://perma.cc/H4PW-BNKY (discussing evidence relevant to supporting a claim of urgent humanitarian reasons for receiving medical treatment in the United States); Refugee, Asylum, and International Operations Directorate, Asylum Division: Affirmative Asylum Procedures Manual, USCIS (May 2016), https://perma.cc/SZ42-57R6 (noting that expedition of asylum applications may be appropriate when USCIS determines the “spouse or child of an asylum applicant is in danger of harm in the country of claimed persecution”).
• 18Pub. L. No. 79-404, 60 Stat. 237 (1946) (codified as amended in scattered sections of 5 U.S.C.).
• 19See infra Part III.B.
• 20See Jacob B. Natwick, Unreasonable Delay at the VA: Why Federal District Courts Should Intervene and Remedy Five-Year Delays in Veterans’ Mental-Health Benefits Appeals, 95 Iowa L. Rev. 723, 729–45 (2010); Lauren E. Sasser, Waiting in Immigration Limbo: The Federal Court Split over Suits to Compel Action on Stalled Adjustment of Status Applications, 76 Fordham L. Rev. 2511, 2542–52 (2008); Amber Pershon, Processing Citizenship: Jurisdictional Issues in the Unreasonable Delay of Adjudication of Naturalization Applications, 5 Phx. L. Rev. 259, 293–300 (2011).
• 21Pub. L. No. 96-212, 94 Stat. 102 (codified in scattered sections of 8 U.S.C.).
• 22Pub. L. No. 82-414, 66 Stat. 163 (codified as amended at 8 U.S.C. §§ 1101–1537 (2000)).
• 23Edward M. Kennedy, Refugee Act of 1980, 15 Int’l Migration Rev. 141, 143 (1981).
• 24See 8 U.S.C. § 1101(a)(42).
• 25See 8 U.S.C. §§ 1158(b)(1)(A), 1225(b)(1)(A)(i)–(ii).
• 26For persons without a nationality, the relevant country is that of last habitual residence. See 8 U.S.C. § 1101(a)(42).
• 278 U.S.C. § 1101(a)(42). The persecution must be on the basis of at least one of five protected grounds: (1) race; (2) religion; (3) nationality; (4) political opinion; and (5) membership in a particular social group. Id.
• 28See Asylum in the United States, supra note 4. “Affirmative” and “defensive” are terms used by agencies, practitioners, and scholars; they are not used by the INA.
• 29Id.
• 30See Obtaining Asylum in the United States, USCIS (Mar. 15, 2022), https://perma.cc/4CRJ-RMDR [hereinafter Obtaining Asylum in the United States].
• 31See Asylum in the United States, supra note 4. Unlike defensive asylum claims, an affirmative asylum claim need not be raised when an individual is in the country illegally and thus subject to removal. For instance, an individual lawfully present in the United States on a student visa may apply for affirmative asylum, or an individual may approach a port of entry and make an affirmative request for asylum to a border control agent.
• 32Obtaining Asylum in the United States, supra note 30.
• 33Id.
• 348 U.S.C. § 1158(b)(1)(A).
• 35Some important procedural regulations issued by DHS include 8 C.F.R. §§ 208.1–208.24 (the regulations promulgated specifically for asylum and withholding of removal procedures), 8 C.F.R. § 103.2(a)(1) and (b)(9) (describing the procedures for submitting benefit requests and appearing for interview and biometric procedures), and 8 C.F.R. §§ 1208.3 and 1208.4 (outlining the requirements for completing and filing an asylum petition).
• 36Form I-589, Application for Asylum and Withholding of Removal, USCIS (Aug. 25, 2020), https://perma.cc/RA2L-6MCD.
• 37The Affirmative Asylum Process, USCIS (Sept. 16, 2021), https://perma.cc/2ZNV
  -ZMAS.
• 388 U.S.C. § 1158(d)(5)(A)(ii)–(iii).
• 398 U.S.C. § 1158(d)(7) (“Nothing in this subsection shall be construed to create any substantive or procedural right or benefit that is legally enforceable by any party against the United States or its agencies or officers or any other person.”).
• 40See Protection Postponed, supra note 8, at 2, 4, 13.
• 41Cf. 8 C.F.R. § 208.7.
• 42See 8 U.S.C. § 1158(c)(1)(B) (providing employment authorization for asylees given certain conditions); 8 U.S.C. § 1158(c)(1)(C) (allowing asylees to travel abroad with the consent of the Attorney General); Off. of Refugee Resettlement, What We Do, U.S. Dep’t of Health & Hum. Servs. (Feb. 18, 2021), https://perma.cc/NB5F-HZ42 (discussing the availability of certain federal benefits and services to asylees); Form I-730 Instructions, USCIS (Sept. 17, 2019), https://perma.cc/K3ZB-GQME (providing asylees with the ability to petition to have certain family members reunited in the United States).
• 43Cf. Adjustment of Status, USCIS (Sept. 25, 2020), https://perma.cc/3WWK-KRQG [hereinafter Adjustment of Status].
• 44See Become a U.S. Citizen Through Naturalization, USCIS (June 23, 2021), https://perma.cc/ZZ4K-AF72 (describing naturalization eligibility requirements).
• 45A former Department of Justice agency, INS was the predecessor to USCIS. In 2003, INS was dissolved, and its functions were transferred to three new agencies: USCIS, U.S. Immigration and Customs Enforcement (ICE), and U.S. Customs and Border Protection (CBP). See Overview of INS History, USCIS 11 (2012), https://perma.cc/9GEW
  -YU45.
• 46Rebecca Hamlin, Ideology, International Law, and the INS: The Development of American Asylum Politics 1948–Present, 47 Polity 320, 326 (2015); see also Andorra Bruno, Cong. Rsch. Serv., R45539, Immigration: U.S. Asylum Policy 9 (2019) (citing How Do We Determine Who is Entitled to Asylum in the United States and Who Is Not?: Hearing Before the Subcomm. on Immigration and Refugee of the S. Comm. on the Judiciary, 97th Cong., 1st sess., at 6, (1981) (testimony of Doris Meissner, INS Acting Commissioner)).
• 47See Hamlin, supra note 46, at 328; Claire P. Gutekunst, Interdiction of Haitian Migrants on the High Seas: A Legal and Policy Analysis, 10 Yale J. Int’l L. 151, 154–58 (1984).
• 48Cf. Susan Bibler Coutin, Falling Outside: Excavating the History of Central American Asylum Seekers, 36 L. & Soc. Inquiry 569, 575–76 (2011) (describing drastic increases in the late 1980s through early 1990s in migration flows from Central America— specifically from El Salvador and Guatemala, where civil wars raged until 1992 and 1996, respectively); Idean Salehyan & Marc R. Rosenblum, International Relations, Domestic Politics, and Asylum Admissions in the United States, 61 Pol. Rsch. Q. 104, 104 (2008).
• 49Salehyan & Rosenblum, supra note 48, at 104.
• 50Asylum Reform: Five Years Later, 1 (Feb. 1, 2000), https://perma.cc/79D3-WSKZ [hereinafter Asylum Reform].
• 51Doris Meissner, Faye Hipsman & T. Alexander Aleinikoff, The U.S. Asylum System in Crisis: Charting a Way Forward 24 (2018).
• 52Asylum Reform, supra note 50, at 2. This period was extended to 180 days when Congress codified the policy change in the Illegal Immigration Reform and Immigrant Responsibility Act of 1996 (IIRIRA), Pub. L. No. 104-208, 110 Stat. 3009, § 203(d)(2) (codified at 8 U.S.C. § 1158(d)(2)).
• 53Cf. Meissner et al., supra note 51, at 7.
• 54Before the reforms, asylum seekers were eligible to apply for work authorization immediately upon filing their application. See Asylum Reform, supra note 50, at 1; see also Diane Uchimiya, A Blackstone’s Ratio for Asylum: Fighting Fraud While Preserving Procedural Due Process for Asylum Seekers, 26 Penn. St. Int’l L. Rev. 383, 383 n.1 (2007).
• 55Asylum Reform, supra note 50, at 2; Pub. L. No. 104-208, 110 Stat. 3009, § 605.
• 56Meissner et al., supra note 51, at 13 fig.3.
• 57Id. at 9.
• 58Id.
• 59See Asylum Reform, supra note 50, at 1; see also David A. Martin, The 1995 Asylum Reforms, Ctr. for Immigr. St. (May 1, 2000), https://perma.cc/PNH8-3PWN.
• 60Backlog Elimination Plan, USCIS 4 (June 16, 2004) (on file with author).
• 61Meissner et al., supra note 51, at 13 fig.3.
• 62Letter from USCIS Acting Director Tracy Renaud to Representative Gerald Connolly 3 (July 29, 2021), https://perma.cc/5U6H-AV9R [hereinafter Letter from USCIS].
• 63Cf. Meissner et al., supra note 51, at 24.
• 64Adjudication Statistics, supra note 7.
• 65The current LIFO schedule has four levels of priority. The order of priority, from first to fourth, is: (1) rescheduled interviews; (2) applications that have been pending for twenty-one days or fewer; (3) all other pending affirmative asylum applications, which are scheduled for interviews starting with new filings and working back toward older filings; and (4) all pending affirmative asylum applications that are over one hundred days old. USCIS Response to the Citizenship and Immigration Services Ombudsman’s (CISOMB) 2020 Annual Report to Congress, Dep’t of Homeland Sec. 10 (Dec. 4, 2020), https://perma.cc/6A78-TY9V.
• 66Letter from USCIS, supra note 62, at 1, 3.
• 67Id. at 1.
• 68See infra Part IV.C.2.
• 69See Adjudication Statistics, supra note 7.
• 705 U.S.C. § 555(b) (emphasis added).
• 715 U.S.C. § 702.
• 72Lujan v. Nat’l Wildlife Fed., 497 U.S. 871, 883 (1990) (quoting Clarke v. Sec. Indus. Ass’n, 479 U.S. 388, 396–97 (1987)).
• 735 U.S.C. § 706(1).
• 74See 5 U.S.C. § 701(a)(1)–(2).
• 75Norton v. S. Utah Wilderness All. (SUWA), 542 U.S. 55, 62–63 (2004).
• 76Lujan, 497 U.S. at 882 (quoting 5 U.S.C. § 704).
• 77SUWA, 542 U.S. at 63.
• 78Cf. Pub. Citizen Health Rsch. Grp. v. FDA, 740 F.2d 21, 31–32 (D.C. Cir. 1984).
• 79750 F.2d 70 (D.C. Cir. 1984).
• 80Id. at 80.
• 81See, e.g., Barrios Garcia v. U.S. Dep’t of Homeland Sec., 14 F.4th 462, 485 (6th Cir. 2021) (“When resolving whether an agency action has been unreasonably delayed, the federal courts consider [the TRAC factors].”), amended by 25 F.4th 430 (6th Cir. 2022); In re Pesticide Action Network N. Am., NRDC, Inc., 798 F.3d 809, 813 (9th Cir. 2015); Mote v. Wilkie, 976 F.3d 1337, 1343–44 (Fed. Cir. 2020); Gonzalez v. Cuccinelli, 985 F.3d 357, 375 (4th Cir. 2021) (noting that many lower courts use the TRAC factors because they “offer helpful guidance in [the] inquiry”); Towns of Wellesley, Concord & Norwood v. FERC, 829 F.2d 275, 277 (1st Cir. 1987).
• 82TRAC, 750 F.2d at 80.
• 83Mashpee Wampanoag Tribal Council, Inc. v. Norton, 336 F.3d 1094, 1100–03 (D.C. Cir. 2003).
• 84Espin v. Gantner, 381 F. Supp. 2d 261, 266 (S.D.N.Y. 2005) (citing INS v. Miranda, 459 U.S. 14, 19 (1982)).
• 85See In re Am. Rivers & Idaho Rivers United, 372 F.3d 413, 419 (D.C. Cir. 2004) (noting that “[t]here is ‘no per se rule as to how long is too long’ to wait for agency action” (quoting In re Int’l Chem. Workers Union, 958 F.2d 1144, 1149 (D.C. Cir. 1992))).
• 86See Daniel T. Shedd, Cong. Rsch. Serv., R43013, Administrative Agencies and Claims of Unreasonable Delay: Analysis of Court Treatment 10 (2013); cf., e.g., Yu v. Brown, 36 F. Supp. 2d 922, 935 (D.N.M. 1999) (“What constitutes an unreasonable delay in the context of immigration applications depends to a great extent on the facts of the particular case.” (citing Fraga ex rel. Fraga v. Smith, 607 F. Supp. 517, 522 (D. Or. 1985))).
• 87One such provision is 8 U.S.C. § 1158, as discussed in Part I.A. The INA is notorious for having a “labyrinthine statutory regime governing judicial review” with many, and sometimes convoluted, provisions limiting courts’ ability to review agency decisions. See Yule Kim, Cong. Rsch. Serv., RL34444, Judicial Review of Removal Orders 1, 1 (2008); see also id. at 5–10 (discussing the many jurisdictional bars on judicial review for removal orders alone).
• 88See, e.g., Yu Liu v. Wolf, 19 Civ. 410, 2020 WL 2836426, at *7 (S.D.N.Y. May 30, 2020) (discussing cases). Section 706(1) of the APA provides the cause of action, and 28 U.S.C. § 1331—the federal question statute—provides jurisdiction.
• 89For the statute that typically authorizes a writ of mandamus, see 28 U.S.C. § 1361.
• 90See 8 U.S.C. § 1158(d)(7).
• 91See, e.g., Yu Liu, 2020 WL 2836426, at *7; Fangfang Xu v. Cissna, 434 F. Supp. 3d 43, 51–52 (S.D.N.Y. 2020).
• 92See, e.g., Yu Liu, 2020 WL 2836426, at *6 (collecting cases).
• 93But see Yahya v. Barr, No. 1:20-cv-01150, 2021 WL 798873 (E.D. Va. Jan. 19, 2021). In Yahya, the court extended to the asylum context Safadi v. Howard, 466 F. Supp. 2d 696 (E.D. Va. 2006), which had held that the time to process an individual’s adjustment of status application was committed to agency discretion. Yahya, 2021 WL 798873, at *2. Although other courts have not directly addressed this decision, they generally provide three rebuttals against government arguments of this flavor: (1) the plain text of § 1158 says that agencies “shall” adjudicate the application; (2) allowing such conduct to be wholly within the agency’s discretion, and thus potentially left indefinitely delayed, would contravene Congress’s purpose in enacting § 1158; and (3) USCIS “does not possess unfettered discretion to relegate aliens to a state of ‘limbo,’ leaving them to languish there indefinitely.” Kim v. Ashcroft, 340 F. Supp. 2d 384, 393 (S.D.N.Y. 2004).
• 94But see Hui Dong v. Cuccinelli, No. 20-10030, 2021 WL 1214512 (C.D. Cal. Mar. 2, 2021) (order denying motion to dismiss). The court denied the government’s motion to dismiss and, eight days later, granted the government’s stipulation to stay pending adjudication of the plaintiffs’ application. See id.; Hui Dong, No. 20-10030, (C.D. Cal. Mar. 10, 2021) (order approving stipulation to stay the case). On June 11, 2021, the plaintiff gave notice of a voluntary dismissal. Notice of Voluntary Dismissal, Hui Dong, 2021 WL 1214512 (C.D. Cal. Mar. 2, 2021) (order approving stipulation to dismiss the case). The Hui Dong litigation is discussed further in Part III.B.
• 95But see id. at *4 (refusing to apply the TRAC analysis at the pleading stage because of its fact-specific nature).
• 96Xu, 434 F. Supp. 3d at 53; see also Varol v. Radel, 420 F. Supp. 3d 1089, 1097 (S.D. Cal. 2019). But see Ruan v. Wolf, No. 19-cv-4063, 2020 WL 639127, at *6 (E.D.N.Y. Feb. 11, 2020) (“I have serious doubts that a rule that leaves asylum seekers waiting eight years to have their applications adjudicated can be said to be governed by a rule of
  reason.”).
• 97Xu, 434 F. Supp. 3d at 53; see also Baisheng Chen v. Wolf, No. 19-cv-9951, 2020 WL 6825681, at *5 (S.D.N.Y. Nov. 20, 2020) (agreeing with the Xu court’s analysis).
• 98Xu, 434 F. Supp. 3d at 53.
• 99Zhang v. Wolf, No. 19-cv-5370, 2020 WL 5878255, at *5 (E.D.N.Y. Sept. 30, 2020).
• 100Xu, 434 F. Supp. 3d at 51.
• 101See Zhu v. Cissna, No. CV 18-9698 PA, 2019 WL 3064458, at *4 n.5 (C.D. Cal. Apr. 22, 2019) (considering such risks to be “inherent in the process of seeking asylum and would not change this Court’s analysis”); see also Xu, 434 F. Supp. 3d at 54 n.5 (citing Zhu for the same proposition).
• 102Ruan, 2020 WL 639127, at *6–7 (quoting Pesantez v. Johnson, 15 Civ 1155, 2015 WL 5475655, at *5 (E.D.N.Y. Sept. 17, 2015)).
• 103No. CV 18-9698 PA, 2019 WL 3064458 (C.D. Cal. Apr. 22, 2019).
• 104434 F. Supp. 3d 43, 51–52 (S.D.N.Y. 2020).
• 105Zhu, 2019 WL 3064458, at *4 n.5.
• 106Yu Liu, 2020 WL 2836426, at *9.
• 107Mashpee, 336 F.3d at 1100 (quoting In re Barr Lab’ys, Inc., 930 F.2d 72, 75 (D.C. Cir. 1991)).
• 108See supra note 102 and accompanying text; see also infra Part III.B (discussing the Hui Dong litigation).
• 109See, e.g., Adrian Vermuele, Judging Under Uncertainty: An Institutional Theory of Legal Interpretation 63–85 (2006); Eric Berger, Comparative Capacity and Competence, 2020 Wis. L. Rev. 215, 217 (2020); Andrew Coan, Rationing the Constitution: How Judicial Capacity Shapes Supreme Court Decision-Making 179–89 (2019). See also generally, e.g., Neil K. Komesar, Imperfect Alternatives: Choosing Institutions in Law, Economics, and Public Policy (1994).
• 110See supra note 109.
• 111See generally Virginia Robinson, State Policy in Federal Courts: Stabilizing the Burford Abstention Doctrine, 89 U. Chi. L. Rev. 2109 (2022).
• 112See Murray L. Schwartz, The Other Things that Courts Do, 28 UCLA L. Rev. 438, 453 (1981).
• 113Neil K. Komesar, Taking Institutions Seriously: Introduction to a Strategy for Constitutional Analysis, 51 U. Chi. L. Rev. 366, 379 (1984) (“Where the legislation under review affects judicial procedure or quasi-judicial, administrative procedure, judges feel more confident, probably with justification, in their ability to dispose accurately and efficiently of the issue.”).
• 114See Schwartz, supra note 112, at 452, 454. The corollary to such insulation is that courts then suffer from severe information deficits, see Vermuele, supra note 109, at 65, particularly when the question is one of politics or policy, cf. Komesar, supra note 113, at 379 (“[J]udges have few formal channels for independent investigation. More important, the judicial system is poorly placed to receive information on the desires and preferences of the public or any given part of it.”).
• 115Jeff King, The Pervasiveness of Polycentricity, 2008 Pub. L. 101, 103 (2008).
• 116Donald L. Horowitz, Courts and Social Policy 59 (1977).
• 117470 U.S. 821 (1985).
• 118Id. at 837–38.
• 119Id. at 831.
• 120Id.
• 121Id.; see also id. at 843 (Marshall, J., concurring in the judgment) (“As long as the agency is choosing how to allocate finite enforcement resources, the agency’s choice will be entitled to substantial deference.”).
• 122508 U.S. 182 (1993).
• 123Id. at 193 (quoting Heckler, 470 U.S. at 831–32).
• 1245 U.S.C. § 706(1).
• 125See, e.g., Pesantez, 2015 WL 5475655, at *6 (“[A]lthough the analogy to the Supreme Court’s enforcement-discretion cases is imperfect, the policy considerations that they raise are instructive.”).
• 126930 F.2d 72 (D.C. Cir. 1991).
• 127Id. at 73. Although the claim was brought under 28 U.S.C. § 1361, the request that the court compel delayed agency adjudication by issuing a writ of mandamus is analytically similar to APA unreasonable delay claims. In fact, the D.C. Circuit applied the TRAC test to analyze the mandamus claim. Id. at 74–75.
• 128Id. at 76.
• 129362 F.3d 272 (4th Cir. 2004).
• 130Id. at 275–76.
• 131Id. at 280.
• 132Cf. Am. Immigr. Law. Ass’n, Featured Issue: USCIS Budget Shortfall and Furloughs, AILA (Jan. 11, 2022), https://perma.cc/9TLP-NMB5. See generally Sarah Pierce & Doris Meissner, USCIS Budget Implosion Owes to Far More than the
  Pandemic (2020).
• 133See Ruth Ellen Wasem, Cong. Rsch. Serv., R41753, Asylum and “Credible Fear” Issues in U.S. Immigration Policy 4, 9 (2011).
• 134Procedures for Credible Fear Screening and Consideration of Asylum, Withholding of Removal, and CAT Protection Claims by Asylum Officers, 87 Fed. Reg. 18078, 18086 (May 31, 2022) (to be codified in scattered sections of 8 C.F.R.). The rationale is that this process will be more efficient because asylum officers have more expertise in these protection claims, can adjudicate the claims more expeditiously, and can use documentation from the foreign nationals’ credible-fear screenings to adjudicate the asylum claims in lieu of asking the individuals to prepare new applications. Id. at 18085.
• 135Id. at 18086.
• 136Id.
• 137Cf. All Forms, USCIS, https://perma.cc/682D-TD98 (listing the forms for the many benefits USCIS administers in addition to those enumerated in the main text).
• 138See notes 117–22 and accompanying text.
• 139See infra notes 148–63 and accompanying text.
• 140“Adjustment of status” is the legal process by which individuals with various immigration statuses can apply to become legal permanent residents, otherwise known as “green-card holders.” See Adjustment of Status, supra note 43.
• 141The comparison is given additional force by the fact that courts deciding unreasonable delay cases in the asylum context often look to AOS cases as guideposts. See Xu, 434 F. Supp. 3d at 55 (citing several AOS cases in concluding that plaintiff had failed to state a claim of unreasonable delay); Hu v. Sessions, No. 18-CV-2248, 2020 WL 7486681, at *3 (E.D.N.Y. Dec. 17, 2020) (discussing several AOS cases when analyzing the first TRAC factor); Zhang, 2020 WL 5878255, at *5 (same). Below I argue that such comparisons misapprehend the respective natures of delay.
• 142A “Tier III” organization is “a group of two or more individuals, whether organized or not, which engages in [certain terrorist activities defined in 8 U.S.C. § 1182(a)(3)(B)(iv)].” 8 U.S.C. § 1182(a)(3)(B)(vi)(III).
• 143Latifi v. Neufeld, No. 13-cv-05337, 2015 WL 3657860, at *2 (N.D. Cal. June 12, 2015) (citing 8 U.S.C. § 1182(d)(3)(B)(i)).
• 144See, e.g., Islam v. Heinauer, 32 F. Supp. 3d 1063, 1067 (N.D. Cal. 2014).
• 145Mugomoke v. Curda, No. 2:10-CV-02166, 2012 WL 113800, at *7 (E.D. Cal. Jan. 13, 2012).
• 146See, e.g., Islam, 32 F. Supp. 3d at 1071–74. Some courts in the AOS context have found that harms like psychological stress, inability to accrue time toward gaining citizenship, and bureaucratic or financial hassles of applying for employment and travel authorization tilt the third and fifth TRAC factors in the plaintiff’s favor. See, e.g., Al Karim v. Holder, No. 08-cv-00671, 2010 WL 1254840, at *4 (D. Col. Mar. 29, 2010); Kashkool v. Chertoff, 553 F. Supp. 2d 1131, 1145 (D. Ariz. 2008); Boussana v. Johnson, No. 14-cv-3757, 2015 WL 3651329, at *9 (S.D.N.Y. June 11, 2015); Latifi, 2015 WL 3657860, at *7. However, not all holdings in favor of the plaintiff required the third and fifth TRAC factors be in their favor. Islam is an example where the court concluded “these factors do not weigh [ ] in either party’s favor,” 32 F. Supp. 3d at 1073, which is more similar to how some asylum-context courts have viewed them. Cf. supra notes 101–04 and accompanying text.
• 147Cf. Faith Proper, Mission (Im)possibility: Determining When Mandamus is an Appropriate Remedy to Address Agency Delay or Inaction, 72 Fla. L. Rev. 933, 950 (2020) (suggesting that mandamus should issue to compel adjudication if the problem is “agency drift” but not if the problem is inadequate resources).
• 148Cf. USCIS Announces New Actions to Reduce Backlogs, Expand Premium Processing, and Provide Relief to Work Permit Holders, USCIS (Mar. 29, 2022), https://perma.cc/SW6E-WW3N.
• 149See Pesantez, 2015 WL 5475655, at *6.
• 150Cumberland Cnty. Hosp. Sys., Inc. v. Burwell, 816 F.3d 48, 56 (4th Cir. 2016); see also Pesantez, 2015 WL 5475655, at *6.
• 151Abner J. Mikva, Agency Inaction, 4 J. Nat’l Ass’n Admin. L. Judges 5, 7 (1984).
• 152Cf. Kent Roach, Polycentricity and Queue Jumping in Public Law Remedies, 66 U. Toronto L.J. 3, 16 (2016) (discussing scholarship warning that “managerial judging” may exacerbate concerns about the legitimacy of courts and produce counterproductive consequences (quoting Katharine G. Young, Constituting Economic and Social Rights 161 (2012))).
• 153No. 20-10030, 2021 WL 1214512 (C.D. Cal. Mar. 2, 2021).
• 154See id. at *1.
• 155Defendants’ Motion to Dismiss, Hui Dong, 2021 WL 1214512 (C.D. Cal. Mar. 2, 2021) (No. 20-10030).
• 156Hui Dong, 2021 WL 1214512, at *4 (first citing Gelfer v. Chertoff, No. C 06-06724, 2007 WL 902382, at *2 (N.D. Cal. Mar. 22, 2007); then citing Garcia v. Johnson, No. 14-cv-01775, 2014 WL 665759, at *12 (N.D. Cal. Nov. 21, 2014); and then citing Indep. Mining Co. v. Babbitt, 105 F.3d 502 (9th Cir. 1997)).
• 157See id. at *5.
• 158See Stipulation to Stay Case Pending Adjudication of Application Filed by Defendants, Hui Dong, 2021 WL 1214512 (C.D. Cal. Mar. 2, 2021) (No. 20-10030).
• 159See id.
• 160Pesantez, 2015 WL 5475655, at *4; see also Mikva, supra note 151, at 7.
• 161See Proper, supra note 147, at 950; Heckler, 470 U.S. at 839 (Brennan, J., concurring); id. at 842–43 (Marshall, J., concurring in the judgment) (reasoning that agency non-enforcement decisions that are based on sham resource-allocation rationales, on vindictive or personal motives, or on “simply ignor[ing] the request” should be subject to more exacting review than legitimate resource-allocation decisions).
• 162See 8 U.S.C. § 1158(b)(1)(A) (“The Secretary of Homeland Security . . . may grant asylum to an alien who has applied for asylum in accordance with the requirements and procedures established by the Secretary . . . if [the Secretary] determines that such alien is a refugee.”); 8 U.S.C. § 1158(d)(1) (“The Attorney General shall establish a procedure for the consideration of asylum applications.” (emphasis added)); 8 U.S.C. § 1158(d)(5)(B) (“The Attorney General may provide by regulation for any other conditions or limitations on the consideration of an application for asylum not inconsistent with this chapter.” (emphasis added)).
• 163U.S. Const. art. II, § 3.
• 164Cf. Neil Komesar, In Search of a General Approach to Legal Analysis: A Comparative Institutional Alternative, 79 Mich. L. Rev. 1350, 1352, 1359 (1981).
• 165See supra Part III.A.1.
• 166David Freeman Engstrom, Daniel E. Ho, Catherine M. Sharkey & Mariano-Florentino Cuéllar, Government by Algorithm: Artificial Intelligence in Federal Administrative Agencies 12 (2020) [hereinafter Government by Algorithm].
• 167Cary Coglianese & Lavi Ben Dor, AI in Adjudication and Administration, 86 Brooklyn L. Rev. 791, 791 (2021).
• 168Government by Algorithm, supra note 166, at 6.
• 169Id. at 12.
• 170See id. at 7; David Freeman Engstrom & Daniel E. Ho, Algorithmic Accountability in the Administrative State, 37 Yale J. on Reg. 800, 803–04 (2020) (describing how governmental attempts to innovate technologically have floundered compared to those in the private sector).
• 171For broader and comprehensive surveys of case studies of state and federal government use of AI in a variety of manners, see generally Coglianese & Ben Dor, supra note 167, Government by Algorithm, supra note 166, and Engstrom & Ho, supra note 170.
• 172Government by Algorithm, supra note 166, at 38 (quoting Barnhart v. Thomas, 540 U.S. 20, 28–29 (2003)).
• 173Id.
• 174Id. at 39–40.
• 175Id. at 40; Engstrom & Ho, supra note 170, at 811.
• 176Government by Algorithm, supra note 166, at 40; Engstrom & Ho, supra note 170, at 811.
• 177Government by Algorithm, supra note 166, at 40; Engstrom & Ho, supra note 170, at 811. The human examiner may then grant the claim, see 20 C.F.R. §§ 404.1615(c)(3), 404.1619, but if she cannot make a favorable determination after a medical consultation, the claim is adjudicated using regularly applicable non-QDD procedures, see 20 C.F.R. § 404.1619(c).
• 178Government by Algorithm, supra note 166, at 40 (quoting Felix F. Bajandas & Gerald K. Ray, Admin. Conference of the U.S., Implementation and Use of Electronic Case Management Systems in Federal Agency Adjudication 49 (2018)).
• 179I have found relatively little empirical data on the successes of these two AI mechanisms, but the existing evidence suggests that there are benefits to the QDD system. About a year after its implementation, the SSA announced that the QDD model had increased the number of disability claimants receiving expedited approvals “to about 4 percent of all disability cases,” which helped the agency cope with the sharp rise in claims resulting from the economic recession. Carolyn Puckett, Administering Social Security: Challenges Yesterday and Today, 70 Soc. Sec. Bull., no. 3, 2010, at 27, 67.
• 180See Letter from USCIS, supra note 62.
• 181Although I use the categories of the current LIFO system, the proposed mechanism would be applicable to any other prioritization schedule. For each, there would just need to be a determination of which higher, or more urgent, category applications should be moved up if the applicant receives a favorable reprioritization decision. This would depend on the specifics of whichever schedule is in place at the time.
• 182See How to Make an Expedite Request, USCIS (Dec. 28, 2021), https://perma.cc/8BML-CNXT [hereinafter How to Make an Expedite Request].
• 183For instance, I once worked on an expedite request for a family reunification petition in which the applicant’s spouse was still in the country of origin where an extreme wave of violence had broken out against the minority group to which the family belonged. The “request” included a letter brief in which I laid out the family’s factual circumstances, marshalled substantial evidence of the erupting violence, and argued that the spouse was in specific and immediate danger and thus met the criteria set out in the agency’s policy manual. For the specific criteria, see Policy Manual Vol. 1, Part A, Ch. 5, USCIS (Jan. 26, 2022), https://perma.cc/8HDS-K4F4. The request also contained citations to, and some electronic copies of, different types of evidence on which I relied: State Department human rights reports, recent news articles, statements and reports by U.N. special rapporteurs on human rights, and reporting from local grassroots organizations in the country of origin.
• 184Questions regarding harms attendant to the inability to access federal public medical care benefits may include requests for: (1) descriptions of the symptoms or complications and how they have diminished the applicant’s ability to work, care for dependents, go to school, etc.; (2) descriptions of any steps taken or care sought out for the symptoms or underlying causes as well as the outcomes of said efforts; or (3) financial information indicative of the applicant’s ability to seek care.

  Questions related to harms stemming from the ineligibility to petition to bring to the United States immediate relatives may include requests for: (1) descriptions of the persecution suffered by family members or similarly situated individuals, (2) explanations of the events in the country of origin that lead the applicant to believe her family will be or is very likely to be harmed, or (3) unique circumstances or characteristics that would render the family members particularly vulnerable to persecution as compared to similarly situated individuals.

• 185It is beyond the scope of this Comment to detail what type of ML model should be used or to address the technical details of, among other things, defining the model’s objective function; outlining the data collection, cleaning, and partition; or discussing model training. For an in-depth primer on ML, see generally David Lehr & Paul Ohm, Playing with the Data: What Legal Scholars Should Learn About Machine Learning, 51 U.C.D. L. Rev. 653 (2017).
• 186Cf. Government by Algorithm, supra note 166, at 41.
• 187Coglianese & Ben Dor, supra note 167, at 826.
• 188Although I-589s are not among them, USCIS already has a variety of forms that may be filed online. See Forms Available to File Online, USCIS (Dec. 21, 2021), https://perma.cc/KH86-3QAM. USCIS recently announced a five-year plan in which it hopes to establish fully electronic filing and digital processing. See Section 4103 Plan Pursuant to the Emergency Stopgap USCIS Stabilization Act, DHS (Sept. 7, 2021), https://perma.cc/269Y-BLUT.
• 189Government by Algorithm, supra note 166, at 41.
• 190Of course, requiring electronic submissions does introduce some equity concerns if there are applicants who do not have access to the internet and technology or who are uncomfortable with electronic submissions. One possible way of alleviating these concerns would be to allow applicants to mail in hard copies that would be processed by officers, if necessary. However, this may introduce two distinct problems: (1) it may incentivize overuse of the mail-in option if people believe that having a human reviewer would lead to a higher likelihood of reprioritization; and (2) as discussed in much greater detail in Part IV.C.3, a random sample of petitions would need to be adjudicated by humans to facilitate proper learning and functioning of the model. If these mail-in petitions were to become part of the pool of petitions reviewed by humans, we might fear there would perhaps be characteristics about these applicants that would be different than the “average petitioner,” which would skew the results of the now-not-randomized sample. I don’t have a solution to these problems, but acknowledging them here is necessary for evaluating what would be gained and lost by my proposal.
• 191Cf. Lehr & Ohm, supra note 185, at 675.
• 192See supra note 182 and accompanying text.
• 193How to Make an Expedite Request, supra note 182.
• 194Id.
• 195See supra note 17.
• 196Cf. Government by Algorithm, supra note 166, at 71–74.
• 197Id. at 72–73.
• 198See Dep’t of Homeland Sec., United States Citizenship and Immigration Services Budget Overview: Fiscal Year 2022, at 6.
• 199Cf. Government by Algorithm, supra note 166, at 44.
• 200See Engstrom & Ho, supra note 170, at 803–07, 826 (discussing existing scholarship on AI that focuses on big-picture abstractions and constitutional issues, like procedural due process).
• 201Government by Algorithm, supra note 166, at 75.
• 202See, e.g., Alexander W. Resar, The Parameters of Administrative Reason Giving, 67 U. Kan. L. Rev. 575, 575–76 (2019); Jerry L. Mashaw, Reasoned Administration: The European Union, the United States, and the Project of Democratic Governance, 76 Geo. Wash. L. Rev. 99, 111 (2007) (noting that reason-giving is a “protector of judicial review”).
• 203Policy Manual Vol. 1, Part E, Ch. 8, USCIS (Feb. 23, 2022), https://perma.cc/U45P
  -M6LA.
• 204Id.
• 205Engstrom & Ho, supra note 170, at 828.
• 206The harmless-error rule is used by reviewing courts to distinguish between mistakes in prior proceedings that justify a remand and mistakes that “clearly [have] no bearing on the procedure used or the substance of decision reached.” U.S. Steel Corp. v. EPA, 595 F.2d 207, 215 (5th Cir. 1979) (quoting Braniff Airways v. CAB, 379 F.2d 453, 466 (D.C. App. Ct. 1967)). The rule essentially asks whether the mistake was prejudicial.
  See Sea “B” Mining Co. v. Addison, 831 F.3d 244, 253 (4th Cir. 2016) (citing
  Consolidation Coal Co. v. Williams, 453 F.3d 609, 621–22 (4th Cir. 2006)). The harmless-error rule applies to agency actions, including adjudications, but if the mistake did not affect the ultimate outcome, the action will be upheld because “it would be senseless to vacate and remand for reconsideration.” Id.
• 207No. 3:12-cv-1059-O, 2013 WL 5020495 (N.D. Tex. Sept. 13, 2013).
• 208Id. at *20.
• 209No. 19-17386, 2020 WL 4435405 (D.N.J. Aug. 3, 2020).
• 210Id. at *11.
• 211Id. at *11–12; Webb, 2013 WL 5020495, at *20–21; see also Yaws ex rel. B.D.H. v. Berryhill, No. 4:15-CV-0961-Y, 2017 WL 1026231, at *6–7 (N.D. Tex. Feb. 23, 2017).
• 212Webb, 2013 WL 5020495, at *21.
• 213Policy Manual Vol. 1, Part A, Ch. 5, USCIS (Jan. 26, 2022), https://perma.cc/8HDS
  -K4F4.
• 214Several cases note that the plaintiff requested expedition and was denied, but the plaintiff’s claim was unreasonable delay or a mandamus action, not a challenge to the denial of expedited processing. See, e.g., Muvvala v. Wolf, No. 1:20-cv-02423, 2020 WL 5748104, at *2 (D.D.C. Sept. 25, 2020); Nibber v. USCIS, No. 20-3207, 2020 WL 7360215, at *1, *3 (D.D.C. Dec. 15, 2020).
• 215Ruan, 2020 WL 639127, at *2 (“The website [detailing USCIS’s expedite request procedures] appears to indicate that there is no opportunity to appeal or receive judicial review of denials of expedite requests.”).
• 216Asylum Appeals, KPB Immigration Law Firm PC, https://perma.cc/6DQF-YVVQ.
• 217See 8 C.F.R. § 208.14(c)(1).
• 218Cf. Engstrom & Ho, supra note 170, at 841–42.
• 219Id.
• 220See id. at 842.
• 221See Coglianese & Ben Dor, supra note 167, at 796, 806. But see Lehr & Ohm, supra note 185, at 706 (collecting authorities, suggesting that ML models aren’t completely impenetrable, and offering suggestions on how “explainability” may be achieved);
  Cary Coglianese & David Lehr, Regulating by Robot: Administrative Decision Making in the Machine-Learning Era, 105 Geo. L.J. 1147, 1206–07 (2017).
• 222The APA defines “rule” as including “the whole or a part of an agency statement of general or particular applicability and future effect designed to implement, interpret, or prescribe law or policy or describing the organization, procedure, or practice requirements of an agency.” 5 U.S.C. § 551(4).
• 223Kristin E. Hickman, Unpacking the Force of Law, 66 Vand. L. Rev. 465, 475 (2013) (quoting Chrysler Corp. v. Brown, 441 U.S. 281, 295 (1979)).
• 224Jill E. Family, Administrative Law Through the Lens of Immigration Law, 64 Admin. L. Rev. 565, 569 (2012); see also Phillip M. Kannan, The Logical Outgrowth Doctrine in Rulemaking, 49 Admin. L. Rev. 213, 213–14 (1996) (citing 5 U.S.C. § 553(b), (b)(B), (c), and (d)(3)).
• 225See Sidney A. Shapiro & Richard W. Murphy, Eight Things Americans Can’t Figure Out About Controlling Administrative Power, 61 Admin. L. Rev. 5, 13 (2009).
• 2265 U.S.C. § 553(b).
• 2275 U.S.C. § 553(c).
• 2285 U.S.C. § 553(c); see also United States v. Nova Scotia Food Prod. Corp., 568 F.2d 240, 251 (2d Cir. 1977) (“It is not in keeping with the rational process to leave vital questions, raised by comments which are of cogent materiality, completely unanswered.”).
• 229See Kate R. Bowers & Daniel J. Sheffner, Cong. Rsch. Serv., R46673, Agency Recissions of Legislative Rules 10 (2021) (citing Perez v. Mortg. Bankers Ass’n, 575 U.S. 92, 101 (2015)).
• 230See, e.g., Ernest Gelhorn, Public Participation in Administrative Proceedings, 81 Yale L.J. 359, 369 (1972) (emphasizing the need for public participation in a process so akin to the legislative process); Michael Asimow, Interim-Final Rules: Making Haste Slowly, 51 Admin. L. Rev. 703, 707–08 (1999) (“Public input provides valuable information to rulemaking agencies. . . . Rules adopted with public participation are likely to be more effective and less costly to administer than rules written without such participation.”).
• 231See, e.g., Kevin Hartnett, Jr., An Approach to Improving Judicial Review of the APA’s “Good Cause” Exception to Notice-and-Comment Rulemaking, 68 Buff. L. Rev. 1561, 1562 n.5 (2020) (noting how “burdensome, costly, and time-consuming” notice-and-comment procedures can be and collecting sources on that point).
• 232Of course, an agency’s organic statute can impose procedures in addition to or completely different from those put forth in the APA. The APA serves as a set of default rules when Congress doesn’t say otherwise.
• 233See 5 U.S.C § 553(b)(A)–(B).
• 234Hickman, supra note 223, at 474 (citing Richard J. Pierce, Distinguishing Legislative Rules from Interpretive Rules, 52 Admin. L. Rev. 547, 555 (2000)).
• 2355 U.S.C. § 553(b)(A).
• 236Cmty. Nutrition Inst. v. Young, 818 F.2d 943, 946 (D.C. Cir. 1987) (per curiam) (quoting Noel v. Chapman, 508 F.2d 1023, 1030 (2d Cir. 1975)).
• 237Id. at 947.
• 238Cf. Batterton v. Marshall, 648 F.2d 694, 707 (D.C. Cir. 1980).
• 239Engstrom & Ho, supra note 170, at 837–38.
• 240See id.; see also supra notes 174–77 and accompanying text.
• 241I searched terms like “expedite,” “expedite request,” and “expedited processing” in the Federal Register; filtered by agency (“Homeland Security Department”); and filtered by topic (“Aliens”). Although several rules contained references to the expedite system and citations to USCIS’s expedite request webpage, see supra note 182, none of the rules concerned the creation of the system. Some comments for rules suggested changes to the expedite system. The agency generally declined to follow the suggestions, and it is unclear if the agency thinks it would need to go through APA rulemaking to modify the rules. See Expansion of Provisional Unlawful Presence Waivers of Inadmissibility, 81 Fed. Reg. 50244, 50260 (July 29, 2016) (codified at 8 C.F.R. §§ 103 and 212). Similarly, searching Title 8 of the Code of Federal Regulations, which concerns “aliens and nationality,” for terms like “expedite” or “expedite request” does not turn up any regulations concerning USCIS’s expedite request. The closest match is 8 C.F.R. § 208.5, which provides that expedited consideration “shall be given to applications” of foreign nationals detained by DHS who make a claim for a credible-fear determination. This is still distinct from the general expedite request system that is open to USCIS applicants writ large.
• 242Batterton, 648 F.2d at 707; see also Engstrom & Ho, supra note 170, at 838 (suggesting that the SSA’s QDD system may not be substantive in nature because “there is no alteration of rights in the sense that eligibility criteria are unchanged”).
• 243Cf. Ted Hesson, The Man Behind Trump’s ‘Invisible Wall,’ Politico (Sept. 20, 2018), https://perma.cc/AXY5-D6U9 (noting that USCIS is “typically the quietest of the three agencies that deal with immigration”); Ryan Devereaux, U.S. Citizenship and
  Immigration Services Will Remove “Nation of Immigrants” from Mission Statement, The Intercept (Feb. 22, 2018), https://perma.cc/7FMN-XMZS (quoting a USCIS official regarding the change from INS to USCIS: “We wanted people to feel comfortable with coming to us and know that they could get a fair hearing—that we were different from ICE and CBP”).
• 244See generally Metering and Asylum Turnbacks, AIC (Mar. 2021), https://perma.cc/R7CM-R6FY.
• 245See generally A Guide to Title 42 Expulsions at the Border, AIC (Oct. 2021), https://perma.cc/J4A6-QBCB; Q&A: US Title 42 Policy to Expel Migrants at the Border, Hum. Rts. Watch (Apr. 8, 2021), https://perma.cc/4XBM-X47N.
• 246Public Charge Ground of Inadmissibility, 86 Fed. Reg. 47,025 (proposed Aug. 23, 2021) (to be codified at 8 C.F.R. § 212).
• 247See Protection Postponed, supra note 8, at 13.
• 248See Hartnett, Jr., supra note 231, at 1562 n.5.
• 249See, e.g., Dep’t of Homeland Sec. v. Regents of the U. of California, 140 S. Ct. 1891, 1910–15 (2020) (holding that the Department of Homeland Security’s recission of the
  Deferred Action for Childhood Arrivals (DACA) program was arbitrary and capricious under 5 U.S.C. § 706(2)(A) because the agency failed to consider relevant factors, like reliance interests).
• 250Kannan, supra note 224, at 217.
• 251See supra Part IV.B.2.a.
• 252Batterton, 648 F.2d at 707; see also Engstrom & Ho, supra note 170, at 838.
• 253The SSA took a slightly different approach with the QDD system in that information on the model was included in the proposed rule, but there was concern among scholars that the SSA hadn’t provided sufficient detail. See Engstrom & Ho, supra note 170, at 837–38.
• 254Id. at 849–52.
• 255Id. at 851.
• 256Id. (discussing the phenomenon of “temporal drift” in which a model, trained on a retrospective test sample, does not generalize prospectively because adjudicators, claimants, and circumstances have changed with time).
• 257Id.
• 258See supra Part IV.A.
• 259Cf. District Court Miscellaneous Fee Schedule, U.S. Cts. (Dec. 1, 2020), https://perma.cc/8HNE-GZFT.
• 260See generally Deborah L. Rhode, Access to Justice, 69 Fordham L. Rev. 1785 (2001).
• 261See generally Robyn Meyer-Thompson, Facing High Hurdles: Immigration, Poverty, and Access to Justice, 83 Hennepin L. 28 (2014) (discussing how (1) language barriers, (2) fears of deportation, (3) cultural differences, and (4) problems with access to health care pose additional barriers to immigrants seeking legal representation, particularly for low-income immigrants).
• 262See Form G-1055: Fee Schedule, USCIS 5 (Feb. 9, 2022), https://perma.cc/H64N
  -K9YU.
• 263Filing with USCIS—as well as gathering or submitting medical documentation, country conditions research, affidavits, and other forms of supplementary evidence—is well within the competencies of lawyers who provide services for asylum, refugee resettlement, family reunification, or humanitarian parole.
• 264See, e.g., Pro Se Asylum Manual, Pol. Asylum/Immigr. Representation Project (May 2016), https://perma.cc/MLP9-QHR2; University of Maine School of Law Refugee and Human Rights Clinic, Penn State Law Center for Immigrants’ Rights Clinic, Immigrant Legal Advocacy Project & THRIVE International Programs, Guide to Seeking Asylum, Withholding of Removal, and Convention Against Torture Protection (Jan. 2021), https://perma.cc/6PKH-9CLL.
• 265See Pesantez, 2015 WL 5475655, at *6.