USCIS AI Adjudication: Preparing EB-1A Petitions in 2026

The EB-1A approval rate in Q1 FY2025 was 74.7%. In Q4 FY2025, it was 53.4%. Filing volume did not change. Denials nearly doubled.

That is not a statistical fluctuation. It is a structural change in the adjudication environment, and the machine-learning tools USCIS deployed during this period are part of the explanation. Not the only part. But a part that is now well-documented by practitioners and that has a direct, actionable implication for how petitions should be prepared.

What the ELIS Evidence Classifier Actually Does

USCIS has been building out AI-assisted adjudication tooling since 2017, but deployment accelerated beginning in FY2024. The most directly relevant system for EB-1A practitioners is the ELIS Evidence Classifier.

The classifier is a machine-learning tool embedded in USCIS's Electronic Immigration System (ELIS). When a petition arrives, the classifier processes the uploaded documents automatically, assigns document-type tags to each exhibit, and creates a prioritized view for the adjudicating officer. The officer sees a pre-organized case file rather than a raw document pile. The system has processed over 24 million page scrolls.

USCIS did not build this to increase denials. The objective was efficiency. Officers at high-volume service centers handle dozens of cases daily; automated pre-sorting reduces the time they spend identifying what a document is before they can evaluate whether it is sufficient. That goal is legitimate.

The problem is accuracy. The classifier performs reasonably well on common document types in common formats. It performs less well on anything unusual: an academic paper formatted as a preprint, an award certificate in a non-standard layout, an expert letter from a foreign institution with atypical letterhead. When the classifier misidentifies a document, the officer's case file presents that document under the wrong criterion heading, or does not present it at all.

The officer does not know the classifier made an error. They see an organized case with an apparent gap. Then the gap becomes an RFE.

The Four Mechanisms Generating More RFEs

An April 2026 client alert from Cozen O'Connor's immigration group, authored by Immigration Practice Chair Scott Bettridge and David Adams, documented four specific ways USCIS AI tools are producing higher RFE and denial rates. It is the clearest practitioner-facing account of the mechanisms available.

Misclassified evidence. The classifier auto-tags documents by type. When it misidentifies a document, the officer's case view shows it under the wrong criterion or not at all. The result: an RFE requesting documents already in the record. Practitioners have now documented multiple instances where attorneys received RFEs for evidence that was physically present in the filing.

Cross-document discrepancy detection. The Verification Match Model cross-references names, dates of birth, and employment identifiers across the E-Verify and SAVE databases. Minor inconsistencies (a middle name missing in one document, a slightly different employer abbreviation between exhibits) trigger automated flags that reach the adjudicating officer as apparent inconsistencies requiring explanation. These are not new inconsistencies. They existed in petitions approved two years ago. The automated cross-referencing now surfaces them systematically.

Hidden PDF text layer detection. When a document is scanned and then run through optical character recognition, the resulting PDF has two layers: the visible image and the OCR text. If the OCR was imperfect, the text layer may say something different from what the image shows. USCIS AI tools can detect and read the text layer. A scanned expert letter with a poorly OCR'd passage about the beneficiary's contribution can register differently in the classifier than what the human reader sees.

AI-generated boilerplate RFEs. Practitioners report that recent RFEs increasingly contain templated language, factual errors, and vague requests that appear to be AI-generated or AI-assisted. Some denial notices repeat language from the preceding RFE almost verbatim. The quality-control failure here runs in both directions: USCIS AI tools are both creating the gaps that trigger RFEs and helping draft the RFEs themselves.

The Denial Rate Collapse: What the Data Shows

The quarterly EB-1A numbers from official USCIS Form I-140 data are the clearest picture of what changed.

Quarter	Period	Approval Rate	Denials
Q1	Oct–Dec 2024	74.7%	1,091
Q2	Jan–Mar 2025	72.7%	1,276
Q3	Apr–Jun 2025	66.5%	1,765
Q4	Jul–Sep 2025	53.4%	2,033

Source: USCIS Form I-140 FY2025 Q4. Approval rate = approvals ÷ (approvals + denials).

Three things stand out. First: filing volume was essentially flat across all four quarters, averaging 7,400 petitions per quarter. The applicant pool did not get worse. Second: denials went from 1,091 in Q1 to 2,033 in Q4, an 86% increase in one year. Third: the decline is monotonic. No quarter reversed the trend.

The same pattern shows up in EB-2 NIW. The NIW approval rate in Q4 FY2025 was 35.7%, the lowest quarterly figure on record. Fewer than one in three NIW petitions were approved that quarter. NIW and EB-1A are processed by different officers applying different legal standards. A shared decline of this magnitude, in the same quarters, using the same underlying filing system, is not coincidental.

It also shows up in the student termination episode. Over 1,200 international students had their SEVIS records automatically terminated due to AI-driven database mismatches. Courts issued nationwide injunctions and the terminations were reversed. The incident established clearly that USCIS AI tools are capable of cascading errors at scale.

The full FY2025 quarterly data analysis provides the complete breakdown, including country-level approval rates and the pending backlog figures. The approval rate decline article covers the adjudication standard implications in detail. This article focuses on the AI-specific preparation adjustments.

What AI Preprocessing Sees (And What It Misses)

Understanding the classifier's architecture helps predict where it fails. Machine-learning document classifiers are trained on labeled examples. They perform well on the document types most common in their training set, in the formats most common in those examples. They fail in predictable directions.

Format-dependent failures. A standard expert letter from a U.S. university on institutional letterhead, in 12-point Times New Roman, will classify correctly almost every time. The same letter on a foreign institution's letterhead with non-English formatting cues may not. A citation database printout in the standard Web of Science format classifies cleanly. A custom citation analysis prepared by the attorney does not.

Text-layer-dependent failures. The classifier reads text, not images. A scanned document with no OCR layer is opaque to it. A scanned document with a poor-quality OCR layer presents corrupted text that may misidentify the subject, the date, or the institution. Native-digital PDFs, exported directly from word processors, classify most accurately.

Cross-document reliance. The Verification Match Model's job is specifically to find inconsistencies across documents. It will find them. Every name variation ("Yuri" in the passport versus "Yuriy" on a publication, "MIT" on an award certificate versus "Massachusetts Institute of Technology" on the expert letter) is a potential flag. These variations are common. They never mattered when humans were doing the initial scan. They matter now.

What the classifier cannot do is evaluate the argument. It cannot determine whether the evidence is legally sufficient for Criterion 5 under 8 CFR 204.5(h)(3)(v). It cannot apply the Kazarian final merits analysis. Those remain human determinations. But if the classifier mishandles the preprocessing, the human gets a distorted picture of the case before applying legal judgment.

How to Prepare Your Petition for AI Review

The response to this environment is not abstract. These are specific preparation steps with direct correspondence to the four documented failure modes.

Use native-digital PDFs wherever possible. If a client's award certificate or letter was issued in paper-only form, scan it at 300 dpi or higher and run it through a high-quality OCR tool (Adobe Acrobat, ABBYY FineReader, or equivalent). Verify the OCR output against the visual before filing. Do not trust the automatic OCR that comes with basic scan-to-email.

Standardize names and identifiers across every exhibit. Before filing, audit all documents against the beneficiary's legal name exactly as it appears on their passport. An exhibit-by-exhibit consistency check is not optional anymore. Table it: every document, the name as it appears, the date, the institution. Flag every variation. Address it proactively in the petition narrative.

Label exhibits explicitly, not descriptively. "Exhibit 4" tells the classifier nothing. "Exhibit 4: Expert Letter from Professor Jana Novak, MIT, re Criterion 5 (Original Contributions)" tells it what type of document this is, who issued it, what institution, and which criterion it addresses. You are training the classifier, not fighting it.

Cross-reference exhibits in the petition letter, by exhibit number and criterion, in every section. This is the highest-impact single change. If the classifier mis-tags Exhibit 12 as a general support letter rather than a Criterion 8 critical-role letter, the officer's case view will show a gap at Criterion 8. A petition letter that explicitly states "As established by the critical-role letter from [Institution] (Exhibit 12), the beneficiary directed the largest..." creates a textual anchor that survives misclassification. The officer may still read the exhibit. They will definitely read the petition letter.

What AI Review Means for Expert Letters

Expert letters deserve specific attention. They are the document type most likely to be misclassified because they vary most in format, author style, institution, and country of origin. Misclassification here has the highest downstream consequence: expert letters are load-bearing evidence for Criterion 5 and Criterion 8, the two most challenged criteria in the 2024–2025 AAO decision record.

Several adjustments are worth making with the AI preprocessing environment in mind.

The subject line of the letter should state explicitly which criterion or criteria it addresses. Not a legal conclusion (the letter-writer is not an attorney) but a simple subject heading: "Re: EB-1A Petition for [Name], Extraordinary Ability in [Field]: Evidence of Original Contributions of Major Significance (Criterion 5)." That heading appears in the text layer. The classifier reads it.

Each substantive claim in the letter should name the exhibit that supports it. The classifier and the reviewing officer will both see the connection. "The beneficiary's 2023 paper on transformer interpretability (see Exhibit 7) has been cited 214 times, including by three independent research groups at major universities" does more work than "the beneficiary's research has been widely cited."

Vague praise is the enemy. "Dr. [Name] is one of the most talented researchers I have encountered in my career" is unverifiable, non-specific, and structurally indistinguishable from a form letter in any automated analysis. "Dr. [Name]'s 2022 architecture directly reduced inference latency by 31% in three independent replications (Exhibits 8, 9, and 10) and has been adopted in production by two Fortune 500 companies" is not.

The AI-generated boilerplate RFE problem connects directly here. RFEs that contain templated language often do so because the classifier could not surface specific, verifiable claims in the supporting documentation. The template fills in where specific content was missing. Better expert letters produce more specific AI-extracted summaries. More specific summaries reduce the probability of a gap-filling boilerplate RFE.

For more on the quality standard that currently applies to expert letters, see the AAO patterns analysis and the Kazarian Step 2 denial discussion.

The Practitioner's Pre-Filing Checklist

These are the steps that correspond directly to documented AI-driven failure modes.

Exhibit audit (before final assembly):

• Every exhibit is a native-digital PDF or a high-resolution scan with verified OCR
• Every exhibit has an explicit, criterion-specific label in the file name and first page
• The beneficiary's name appears identically across every exhibit (match passport exactly)
• All dates are consistent across documents referencing the same events
• No document contains a hidden OCR layer that conflicts with visible text

Petition letter audit:

• Every exhibit is cited by number in the criterion section where it serves as evidence
• Each expert letter is cross-referenced at least twice: once in the criterion section, once in the Final Merits argument
• The cover letter or exhibit index includes a criterion-to-exhibit mapping table
• No section relies on unstated evidence: if you are arguing it, cite the exhibit number

Document organization:

• Table of contents submitted with the petition lists each exhibit, its type, and its criterion
• Documents supporting each criterion are grouped consecutively, not scattered across the filing
• Any foreign-language document is immediately followed by the certified translation in the same exhibit slot

Implications for Attorneys

The threshold question for any EB-1A petition has always been evidentiary merit: does the record, taken as a whole, establish extraordinary ability? That has not changed. A weak petition will fail regardless of how well it is organized, and a genuinely strong petition should survive some degree of AI-preprocessing noise.

The change is in the error distribution. Before automated preprocessing, a disorganized but meritorious petition might be evaluated in full by a thorough officer. Now it might receive an RFE that the attorney then spends three months responding to, with the beneficiary waiting. The merit survived the process, but the RFE was preventable.

Over 30% of EB-1A approvals in 2025 came after RFEs or NOIDs requiring additional documentation, by practitioner estimates. Some of those RFEs had merit. Others were artifacts of the preprocessing layer: a three-month delay and a client in status limbo for a gap that better document labeling would have prevented.

The AI adjudication environment also creates a direct argument for using AI in petition preparation. If the system reviewing your filing is performing automated document classification, evidence extraction, and cross-document consistency analysis, then preparing with tools that do the same thing (on your side, before filing) closes the information gap. What USCIS's classifier sees should not be a surprise after the RFE arrives.

Immigration Copilot builds this preparation layer into the petition workflow: document classification against EB-1A criteria, cross-exhibit consistency checking, and criterion-specific evidence gap analysis before the filing goes out the door. The tools exist. The question is whether your current preparation workflow uses them.

For the foundational question of whether AI tools are safe to use in your practice, see Is AI Safe for Your Immigration Practice?. For the full FY2025 data on EB-1A approval rates and the denial acceleration, see EB-1A Approval Rate Fell from 74% to 53%: What Attorneys Must Do Now.