Consider a founder (a composite, but a familiar one) who uses a language model to help generate a draft of his immigration petition and gets a polished draft in about twenty minutes. It is fluent, well organized, and confident. It describes him as having “pioneered” a category, “led” a team of engineers, and “driven” a product to “widespread industry adoption.” It reads like a strong case.
Almost none of it can be proven with what he has in his folder.
That gap, between what an AI draft asserts and what the file can actually support, is among the most important things to check before an AI-assisted immigration petition is filed. The drafting is no longer the hard part. The evidence is.
Why immigration petitions punish overstatement
Most people who have not been through the process assume a visa petition is an argument. It is closer to a trial record. The petitioner bears the burden of proof. The regulations require the filing to include all required initial evidence, and the petitioner must establish that eligibility existed at filing and continues through adjudication. Under the preponderance standard, the petitioner must show that the facts required for eligibility are more likely than not true when the record is considered as a whole.
An officer does not take the petition’s word for it. Each factual statement in the cover letter or personal statement is a claim that either points to a document or does not. In EB-1A extraordinary-ability petitions, USCIS applies a two-step analysis drawn from Kazarian v. USCIS: officers first assess whether the initial evidence meets the specific regulatory criteria, then evaluate the whole record in a final merits determination. In both steps, unsupported adjectives count for nothing. Worse, unsupported or inconsistent claims can undermine the credibility of the filing.
The Board of Immigration Appeals said this plainly decades ago in Matter of Ho: doubt cast on any aspect of the evidence may lead to a reevaluation of the reliability and sufficiency of the remaining evidence. And a petitioner signs under penalty of perjury. A willful, material misrepresentation made to obtain an immigration benefit may also raise inadmissibility issues under federal immigration law, depending on the facts.
That is the environment an AI draft enters. It was not built for it.
Where AI drafts drift from the record
Language models are designed to generate plausible, fluent text based on patterns in their training data and the information provided to them. That can produce confident language that goes beyond what the underlying record actually supports. In AI-assisted drafts, the drift shows up in predictable places.
Role inflation. “Co-founder and technical lead” may become “founder and architect.” A member of a four-person team may become the person who “led” it. The resulting language may effectively “round up” the facts to a more impressive version than the underlying evidence supports. In an L-1A or EB-1A petition, the distinction between managing a function and contributing as part of a team can determine whether the evidence establishes managerial or executive capacity in the L-1A context, or supports the leading or critical role criterion in the EB-1A context.
Invented or unattributed metrics. AI-generated drafts can contain figures the applicant never supplied: user counts, revenue growth, citation numbers, or market share. Sometimes the model extrapolates from a vague statement (“we grew a lot”) into a specific percentage. An unsupported number may weaken the record and can contribute to an RFE, NOID, or denial, and a specific number that turns out to be wrong is a credibility problem.
Misstated legal standards. Models can describe the requirements for a category loosely or incorrectly, or borrow criteria from a neighboring visa. An EB-2 National Interest Waiver draft that argues “extraordinary ability” is arguing the wrong case. A draft that treats a startup’s funding round as satisfying an “awards” criterion is misreading the regulation.
Fabricated citations. This is well documented in litigation, where courts have sanctioned lawyers for filing briefs containing case citations an AI tool invented, and the same risk applies to immigration drafts: references to publications, cases, or press coverage that do not exist, or that exist but say something different.
Generic recommendation letters. AI-assisted drafting can result in support letters that repeat the petition’s claims in similar language and with similar phrasing. Regardless of how a letter is prepared, it should accurately reflect the writer’s own knowledge and views, explain the writer’s basis of knowledge, provide specific facts, and remain consistent with the underlying evidence.
A useful review discipline: every material claim points to supporting evidence
The discipline that fixes all of this is simple to state and tedious to do. Take the draft, highlight every factual assertion that matters to eligibility, and ask one question of each: what evidence supports this?
If the supporting evidence is already in the file, identify and cite it where appropriate. If additional corroborating evidence could reasonably be obtained, consider whether it should be gathered and included. Examples might include an employment verification letter, a cap table, a contract, an analytics export, an org chart, or a published article. If a material claim cannot be adequately supported, it should be removed or rewritten to reflect what the evidence actually establishes.
That last step is where founders resist, because the accurate version is often less flattering. “Managed a team of twelve engineers” becomes “served as one of three engineering leads, with direct responsibility for a team of four.” The second sentence is weaker as marketing. It is stronger as a petition, because it survives contact with the organizational chart.
A few practical checks belong on every review:
- Titles and dates in the narrative should match the employment letters, tax records, and corporate filings exactly.
- Numbers trace to a source that can be produced on request, and the source is included rather than described.
- Ownership and control language (founder, majority owner, executive) matches the incorporation documents and share ledger.
- Third-party recognition (press, awards, citations) is real, reproduced in full, and about this person rather than the company generally.
- The legal standard stated in the draft is the correct one for the category being filed, and each criterion claimed is matched to the specific evidence offered for it.
What meaningful attorney review actually involves
“Attorney reviewed” can describe different levels of attorney involvement. A petitioner should understand what that review actually includes.
Meaningful review is not proofreading. It is the evidence audit described above, done by someone who knows what the adjudicating officer will look for, and it involves three things the attorney must handle with independent professional judgment.
First, judgment about category and strategy. Whether a founder’s profile is better suited to an O-1A, EB-1A, EB-2 NIW, E-2, or L-1 path requires professional judgment based on the individual facts, applicable law, available evidence, and the client’s objectives. Pursuing the wrong immigration category or filing strategy with a polished draft is an expensive way to reach a denial.
Second, an independent reading of the evidence. The attorney should look at the source documents themselves, not the AI’s summary of them, and form their own view of what each one proves. A recommendation letter that the model characterizes as “strong independent support” may, on reading, come from a writer whose basis of knowledge is far narrower than the summary suggests.
Third, responsibility. The attorney of record files a Form G-28 and signs the preparer section of the benefit form, making the representations in that declaration, while the petitioner separately certifies the filing. That accountability changes how carefully the file is read. It is also why the attorney’s review has to happen before filing, not after a Request for Evidence arrives.
The old habit of filing a thin case and fixing it in the RFE response was never wise. A petitioner is not entitled to an RFE. Under 8 CFR 103.2(b)(8), USCIS may deny a filing without first issuing an RFE or NOID where the record establishes ineligibility, and where required initial evidence is missing the agency may deny rather than ask for it. The evidence audit has to be finished before the file leaves the office.
Using AI responsibly
None of this is an argument against using AI in petition preparation. Used appropriately, AI tools can assist attorneys with certain organizational, drafting, and review tasks. But attorneys considering the use of these tools should independently evaluate whether and how they may be used consistent with their professional obligations, including duties relating to client confidentiality, data security, competence, supervision, and the verification of AI-generated work. Those considerations may vary depending on the particular tool, its terms and data-handling practices, the information being provided to it, and the applicable rules of professional conduct.
The failure mode is treating the draft as the product. The draft is a hypothesis about what the case could be. The evidence determines what the case is. The work that matters, and the work that a serious attorney review should consist of, is closing the distance between the two, one claim at a time.
If the founder in the opening scenario filed his twenty-minute draft, he would submit a petition that describes someone slightly more accomplished than he is, supported by evidence that proves the credentials of someone slightly less qualified. After the audit, the petition describes exactly who he is, and every sentence of it points to a page in the file. That is a much less exciting document. It is also the version that presents the record more credibly.
