Asylum Backlog Reform That Borrows From AI Governance's Worst Habit

On July 28, 2026, the Department of Homeland Security published an interim final rule titled "Affirmative Asylum Referrals Without Interview" under Docket No. USCIS-2026-0199 (RIN 1615-AC95). The rule took effect immediately, with a public comment window running through September 28, 2026.

I've been writing about AI governance for long enough to recognize the move this rule makes. Not because immigration and AI regulation are the same thing — they aren't — but because the structural failure being attempted here is identical to failures I see in model-risk programs, algorithmic-audit boards, and AI incident-review processes all the time. The rule changes who reviews a claim before it moves through the system, without changing the system's capacity to handle what it receives. The backlog doesn't shrink. It relocates.

Richard T. Herman, an immigration attorney, made this argument directly in a piece for The Regulatory Review published August 24, 2026. He's right, and I want to work through why — both for what it reveals about this specific rule and for what it says about a governance failure mode that shows up across regulatory domains, including the one this site covers.

What the Rule Changes, Precisely

Affirmative asylum is available to someone who isn't already in removal proceedings. They file with U.S. Citizenship and Immigration Services (USCIS) voluntarily, and historically a trained asylum officer interviews them before deciding whether to grant protection or refer the case to immigration court. That interview has always been the hinge point of the process — a non-adversarial conversation where an officer develops the record, identifies gaps, and hears the applicant describe what happened to them directly.

Under the new rule, a USCIS officer can now skip that conversation and refer a case straight to the Department of Justice's Executive Office for Immigration Review (EOIR) — without ever speaking to the applicant — when the paper record suggests the claim falls into one of the following four categories:

• Barred by the one-year filing deadline
• Barred on other statutory grounds
• Unlikely to merit a discretionary grant
• Unlikely to succeed on the merits as filed

The rule also eliminates the requirement that the referral letter include a credibility assessment of the applicant. That's not a technical footnote. It's the removal of the one formal mechanism that required an officer to document whether the person's account appeared truthful — before the case moved into a context where proving credibility becomes far harder.

DHS projects roughly 132,167 referrals annually once the rule runs at full speed. The agency also identified up to 444,724 pending applications — close to a third of the entire affirmative backlog — that may turn on one-year deadline questions or missing entry-date information, the exact category this rule is built to screen out without a hearing. The rule text and these projections are available in the Federal Register filing under Docket No. USCIS-2026-0199.

Where the Cases Actually Go

A referral doesn't make a claim disappear. It moves the claim from USCIS to immigration court, from an interview to a hearing, and from an officer developing a record to a judge presiding over two sides contesting each other — one of them a government trial attorney whose job is to argue the claim should fail.

As of the end of the second quarter of fiscal year 2026, EOIR was still carrying 3,570,145 pending matters, according to the agency's own adjudication statistics reporting, with no signs of relief in sight. The Board of Immigration Appeals (BIA) — the next stop after a judge's decision, before a case can reach a federal circuit court — had 219,945 cases pending as of the end of the first quarter of the same fiscal year, according to the same EOIR reporting. Meanwhile, more than 100 immigration judges left the bench between January 2025 and the rule's publication date, cutting into the very capacity the rule depends on to absorb new volume.

So the rule doesn't shrink the number of contested claims sitting somewhere in the system. It moves roughly 132,000 cases per year from one overloaded queue into a different, more overloaded queue — at the exact moment that queue is shrinking in capacity, not growing.

The Math Nobody in the Rule Disputes

None of this is contested. DHS's own estimates and EOIR's own caseload reporting tell the same story from opposite ends of the pipeline.

That last row is the key one. The rule's logic depends on immigration courts having room to absorb what USCIS stops interviewing. The judge count is moving in the wrong direction.

What Gets Lost When You Skip the Conversation

An interview is not a procedural formality. It's the one point in the entire affirmative process where a government officer can ask a follow-up question instead of just reading a form. Someone who missed the one-year filing deadline because they were hospitalized, or because a previous attorney mishandled the paperwork, or because they genuinely didn't know the deadline existed, gets to explain that — out loud, to a person trained to hear it. On paper, a missed deadline looks like a missed deadline. In a conversation, it might look like a statutory exception that the law actually allows.

The interview also serves DHS's own interest in a way that pure documentary review doesn't. A trained officer probing a thin-on-paper claim can determine whether it's fraudulent or just poorly documented. Herman's argument in the original piece is that the interview cuts both ways: it protects legitimate claims from misreading and catches weak ones before they occupy a judge's docket. Removing it doesn't just create fairness risk for applicants. It removes a filtering tool that the agency itself relies on.

Once a case lands in immigration court, that same clarifying conversation — if it happens at all — happens for the first time in front of a judge, opposite a government attorney, in a room designed for contest rather than inquiry. And it happens inside a court system where, as the table above shows, the outcome already varies wildly depending on which judge draws the case. A grant-rate range of 4.8% to 97.1% across judges in the same city is not a hypothetical due-process concern. It's a documented adjudicative lottery, and the new rule adds more tickets.

The Governance Pattern This Rule Belongs To

I want to be direct about why this rule matters beyond immigration policy, because that's the thread that connects it to what this site is actually about.

When a system is carrying more work than it can process and the people running it face pressure to show progress, the honest options are few and unglamorous: add capacity, streamline procedure without cutting what actually works, or accept that clearing the backlog will take years. None of those produce a rule you can announce in July. What you can announce in July is a change to who reviews a case before it moves — and if you don't look past the announcement, it reads like decisive action.

I've written elsewhere about Symbol vs. Substance in AI adoption — the tendency of institutions under pressure to change the paperwork trail rather than the underlying capacity, because the paperwork trail is the only thing they can move quickly. The asylum rule is a non-AI example of exactly the same mechanism. The form changes. The queue doesn't.

What makes this version harder to dismiss as pure theater is that AI governance programs do the same thing, and the stakes are just as concrete when they do. A model-risk committee that reviews new AI deployments exists on paper, gets announced in a press release, and then receives 200 deployment requests in a year it was staffed to handle 40. The committee doesn't disappear — it just becomes a rubber stamp, because the alternative is delay at scale that the business won't tolerate. The review happens, technically. The reviewing doesn't. Institutions under pressure often reach for the version of reform that renames the problem rather than solves it, because renaming is faster.

What AI Governance Teams Can Take From This

If you run or sit on an AI oversight body, the asylum rule is a useful mirror. The failure mode is structural, not moral. Well-intentioned people built a rule that relocates rather than resolves, because the honest fix was not available in the time they had. The same conditions exist inside most large AI governance programs right now.

A few questions worth asking of your own program:

• Referral without review. Does your committee or review board receive more cases than it has the hours to examine properly? If so, what is actually happening to the cases at the bottom of the queue — are they being substantively reviewed, or referred onward with a procedural stamp?
• Capacity versus volume. When your program was announced or last expanded, was the projected caseload matched against actual reviewer capacity — hours per case, not cases per year?
• The credibility-assessment gap. The asylum rule eliminates the requirement to document whether an applicant's account appears credible before referring. What is the analogue in your AI review process? Is there a documented record of whether a model's claimed safeguards were actually tested, or does the review record reflect only that the form was submitted?
• Downstream tracking. Do you know what happens to cases your board doesn't resolve — whether they get resolved at all, or simply pile up somewhere else in the organization, renamed as a different team's problem?

There is no shame in answering these questions honestly and finding that the program is doing some version of what DHS just formalized. The shame is in not asking them.

What to Watch Before September 28

The public comment period on DHS Docket No. USCIS-2026-0199 (RIN 1615-AC95) runs through September 28, 2026. The version of the rule currently in effect is not necessarily final. Three things worth tracking as that window closes:

• The credibility-assessment omission. This is a narrow, specific cut that immigration attorneys are likely to target in comments, and it's the kind of targeted revision DHS could make without retreating from the rule's larger structure. Watch whether any final rule restores a credibility-documentation requirement for referrals.
• Judicial appointment volume. The rule's logic requires courts to absorb roughly 132,167 additional cases per year. Whether EOIR receives a corresponding increase in judges, not just in cases, will determine whether this rule reduces the affirmative backlog or transfers it to EOIR's count.
• The 444,724-case denominator. DHS flagged this number as the population of deadline-related or missing-entry-date cases the rule is built to address. If those cases drop from USCIS's books and reappear, over the following year or two, in EOIR's pending count, the rule will have done exactly what its critics said: moved the problem, not solved it. That's the number to watch.

The deeper question, and the one I think the immigration debate is unlikely to settle cleanly, is whether an institution under genuine resource pressure can reform its way to capacity — or whether the only path is the unglamorous one of building more of what the work actually requires.

In AI governance, we tend to be optimistic that the right framework will close the gap between the workload and the workforce. The asylum case is a useful reminder that frameworks don't do the reviewing. People do.

Last updated: 2026-09-16

Frequently Asked Questions

What does DHS's 'Affirmative Asylum Referrals Without Interview' rule actually do?

It allows a USCIS asylum officer to refer certain affirmative asylum applications directly to immigration court — without interviewing the applicant — when the paper record suggests the claim is likely barred by the one-year filing deadline, barred on other statutory grounds, unlikely to merit a discretionary grant, or unlikely to succeed on the merits. The rule took effect July 28, 2026, under Docket No. USCIS-2026-0199 (RIN 1615-AC95), with a comment period running through September 28, 2026.

Does skipping asylum interviews actually reduce the backlog?

Not by DHS's own numbers. The agency projects roughly 132,167 referrals per year once the rule is fully implemented — cases that move from USCIS's affirmative queue into immigration courts that were already carrying 3,570,145 pending matters as of the end of Q2 FY2026, per EOIR's published caseload statistics. The total number of unresolved claims in the system does not decrease; it relocates.

Why does the interview matter — can't officers just review the paperwork?

The interview is the one point in the affirmative process where a USCIS officer can ask follow-up questions, request missing documentation, or hear an applicant explain why a filing deadline was missed. A missed deadline on paper can reflect a statutory exception the law permits; without the conversation, that exception may never be identified. The interview also serves as a filtering mechanism for weak or fraudulent claims before they reach a judge.

What is the AI governance parallel to the asylum backlog problem?

Both involve institutions announcing a new review structure — a rule, a committee, a framework — without matching it to actual reviewer capacity. The result is that cases or AI deployment requests move from one queue to another, get processed procedurally but not substantively, or accumulate in a different part of the organization under a different name. The structural failure is the same: the form of oversight changes while the workload-to-capacity ratio does not.

How much outcome variation exists across immigration judges?

TRAC Immigration's judge-level data shows asylum grant rates ranging from 4.8% to 97.1% among judges in San Francisco, and from 2.6% to 92.4% among judges in New York City. The same underlying fact pattern can produce dramatically different outcomes depending on which judge is assigned — a documented variability problem that the new rule makes more consequential by routing more cases into the court system.