AI-Drafted EB-1A Expert Letters: What Works and What USCIS Flags

An immigration attorney posted on LinkedIn in July 2025 that she had reviewed a denied EB-1A petition with recommendation letters that AI had clearly written. "They were beautifully formatted, polished," she wrote. "But they all read like they were written by the same person. Because they were."

That describes the central failure mode. The problem was not that AI drafted the letters. The problem was that the AI draft was filed without the attorney review step that catches the four specific errors AI reliably produces in expert letters.

Used correctly, AI turns a 90-minute writing task (briefing the expert and drafting the letter) into a 15-minute editing task. The same AI that produces errors also catches them, if you know which errors to look for.

Briefing vs. Drafting: Two Different AI Tasks

Before using any prompt, understand which task you are asking AI to do. They require different inputs and carry different risk profiles.

Briefing the expert means producing the memo you send to the expert before they write. A briefing memo explains the legal standard, names the contribution to address, and specifies what the letter should and should not include. AI for briefing translates legal framing into plain language the expert understands without attorney supervision. If the briefing prompt produces something imprecise, the expert will write a different letter anyway. The error does not propagate to the petition.

Drafting the letter means producing a complete letter the expert reviews and signs. This is higher-stakes: an error in the AI draft may survive unchanged if the expert does not read carefully. The letters in the LinkedIn attorney's failed petition were likely AI drafts the experts rubber-stamped.

The practical rule: use AI for both tasks, but apply your most careful review to the draft, not the briefing. For specific prompts for each task, see 30 AI Prompts for Immigration Attorneys, Category 3.

What a Strong Expert Letter Must Do

Before examining what AI gets wrong, you need the standard it should meet.

An expert letter for Criterion 5 under 8 CFR 204.5(h)(3)(v) must establish that the contribution is of major significance to the field. Not to the beneficiary's career. Not to their employer's product. To the field.

The USCIS Policy Manual, Vol. 6, Part F, Chapter 2 establishes the adjudication standard: expert letters must come from independent experts who can assess field-level impact. The USCIS EB-1A program page summarizes the threshold: sustained national or international acclaim placing the beneficiary at the very top of the field. The letter must do three things:

1. Establish the expert's independent standing to evaluate this contribution (not employer, not close collaborator)
2. Name the specific contribution and describe the state of the field before it existed
3. Argue field-level impact: how others in the field changed their work because of this contribution

A letter that praises the beneficiary as talented, hardworking, or accomplished without making the field-significance argument is not a Criterion 5 letter. It is a character reference. USCIS will note it as such.

The Four AI Errors in Expert Letters

These four errors appear consistently across AI-drafted expert letters, regardless of which model generated them. Every draft should be checked for each one before the expert receives it.

Error 1: Credential Bleed

Credential bleed happens when the AI mixes the expert's credentials into the factual section about the beneficiary.

What it looks like: "Dr. Smith, with 30 years of experience in computational biology and author of over 200 peer-reviewed papers, has observed that [beneficiary]'s work on protein folding prediction has advanced the field significantly."

Why it fails: The 30 years and 200 papers establish the expert's standing. That belongs in the opening paragraph of the bio section. When those credentials appear as supporting context for the claim about the beneficiary, the letter reads as an AI composition: it needed the credential data to establish tone and inserted it where the argument happened to be.

Fix: Separate credential establishment (opening paragraph) from field-impact argument (body). The credential paragraph exists once. It does not reappear to support individual claims.

Error 2: Generic Superlatives Without Comparative Data

What it looks like: "In my professional opinion, [beneficiary] is among the leading researchers in the field of machine learning." / "The contributions have been extraordinary and have been recognized widely by the research community."

Why it fails: "Among the leading" without a denominator is not an expert opinion. It is a sentiment. USCIS has discounted letters in AAO decisions when they lack specific comparative claims. The standard is "risen to the very top": the expert must say the top of what and relative to whom.

Fix: Replace every superlative with a specific claim. "Among the leading" becomes "in the top 5% of citation impact among machine learning researchers publishing in NeurIPS since 2020, based on Google Scholar data as of January 2026." That is an expert opinion. The generic version is not.

Error 3: Hedge Closers

What it looks like: "I believe this work may have been influential in advancing the field." / "These contributions could potentially represent a significant development in the area."

Why it fails: "May have been," "could potentially," "I believe": these are hedges. The letter's job is to provide a definitive expert opinion. An expert who ends with a hedge has not given one. USCIS needs the expert to say it did, not that it might have.

Fix: The closing opinion paragraph must use declarative language. "This contribution has directly influenced [specific subsequent work]. In my professional judgment, it represents an original contribution of major significance to the field." No hedges. If the expert cannot say it declaratively, they should not sign the letter.

Error 4: Vague Field Definitions

What it looks like: "The work has advanced the field of computer science and contributed to the broader technological landscape." / "This research has implications across multiple domains of scientific inquiry."

Why it fails: Field definitions this broad make the Step 2 final merits argument impossible. If the field is "computer science," the population of people who must be surpassed to reach the "very top" is enormous. USCIS uses the field definition the petition establishes. A vague field definition in the expert letter undermines the comparative argument in the cover letter.

Fix: The field definition should match the petition's field of endeavor statement. "The field of transformer-based interpretability methods applied to large language models" is a field definition. "Computer science" is not.

The Review Checklist

Before sending any AI-drafted letter to the expert, run through this checklist.

Credential bleed:

• Expert credentials appear once, in the opening paragraph
• No mention of the expert's publications, years of experience, or titles appears in the body sections about the beneficiary's contributions

Superlative claims:

• Every "leading," "top," "foremost," or "one of the best" is followed by a specific comparative datum (citation percentile, award selectivity, field population)
• No claim is comparative without a comparator

Closing opinion:

• The final paragraph states a definitive expert opinion in declarative language
• No hedges: "may have," "could potentially," "I believe," "seems to"
• The opinion specifically addresses major significance to the field, not general praise for the beneficiary

Field definition:

• The field named in the letter matches the petition's field of endeavor statement
• The field is specific enough to support a comparative argument but not so narrow that the "top" population is trivially small

Independence:

• The letter explains the expert's independent basis for evaluating this work (not as employer, supervisor, or close collaborator)
• If the expert is a collaborator, the relationship is disclosed, not omitted

Legal standard:

• The letter argues field-level impact, not personal praise
• For C5: the contribution changed how others in the field work, with specific evidence
• For C8 under 8 CFR 204.5(h)(3)(viii): two-prong argument present (organization distinction + beneficiary's critical role)

Worked Example: C5 Brief to Attorney-Ready Draft

Input facts for the briefing prompt:

Beneficiary is a machine learning researcher. Key contribution: a training efficiency method for large language models published in 2022. It was cited 214 times in the subsequent two years, including citations by researchers at four major AI laboratories. Three subsequent papers explicitly built on the method. The expert is a professor at a different institution who has cited the work independently.

What AI produces (before fixes):

Professor Chen, a leading authority in AI systems with 20 years of experience and over 150 publications, writes to support [beneficiary]'s petition. [Beneficiary] has made important contributions to the field of AI. His training efficiency work, published in 2022, has been widely recognized by the research community. The paper has received significant attention and has been influential in advancing machine learning research. I believe this work represents a meaningful contribution to the field.

What the checklist catches:

• Credential bleed: "20 years of experience and over 150 publications" appears in the wrong section
• Generic superlative: "leading authority" without data; "widely recognized" without numbers
• Hedge closer: "I believe this work represents a meaningful contribution"
• Vague field: "AI" and "machine learning research"

Attorney-edited version (post-review):

Professor Chen holds an associate professorship in Computer Science at [Institution] and conducts independent research in large language model training. She has no employment or supervisory relationship with [beneficiary].

Prior to [beneficiary]'s 2022 publication in [Journal], training efficiency methods for large language models required [X computational cost per training run], which limited research groups without industrial-scale GPU access. [Beneficiary]'s method reduced this cost by approximately 40%, based on the paper's benchmarks.

Since publication, the paper has been cited 214 times, including citations in work from [Lab 1], [Lab 2], [Lab 3], and [Lab 4]. Three independent research groups (see Exhibits 7, 8, 9) explicitly adopted the method as a component of their own published systems. This citation pattern indicates that independent researchers in the subfield of LLM training efficiency found the contribution sufficiently significant to build on in their own work.

In my professional judgment as an independent researcher in this subfield, [beneficiary]'s contribution is an original contribution of major significance to the field of LLM training efficiency research. It changed the methodological practice for researchers in this area who were previously constrained by computational cost.

The second version makes the same claim the first version gestures toward. It is more work. That is the point. AI writes the first version. The attorney produces the second.

For the specific prompts that generate both the briefing memo and the initial draft, see the expert letter category in the AI prompts guide. For the full workflow on gathering, briefing, and reviewing expert letters across all criteria, see How to Get Expert Recommendation Letters That Win EB1A Cases. For patterns from recent AAO decisions on expert letter quality and what reviewers are flagging, see the 2024-2025 decision analysis. For a complete template with failure analysis, see AI-Assisted Expert Recommendation Letters for EB1A Petitions.

Immigration Copilot generates expert letter briefing memos directly from the client knowledge base. After document classification, the system extracts the contribution, citation data, and field impact evidence for each expert letter target, and produces a briefing document the attorney sends to the expert. The attorney still reviews and edits. The blank-page problem is gone.