Can AI hallucinate facts in a USCIS petition?

Yes. General-purpose AI models (ChatGPT, Claude, Gemini) will generate plausible-sounding but factually incorrect content, including specific citations, statistics, and regulatory references. In a USCIS petition, a hallucinated fact is not just a mistake — it is a potential misrepresentation that can result in denial and bar the petitioner from future filings.

What are the bar ethics obligations for attorneys using AI?

The ABA and most state bars require that attorneys using AI for legal work: (1) supervise the output and take responsibility for its accuracy, (2) maintain competency in the tools they use, (3) protect client confidentiality when sending data to third-party AI systems, and (4) disclose AI use to clients if it materially affects the work. The key obligation is that the attorney remains responsible — AI output cannot be filed without review.

Is it safe to upload client documents to AI tools?

It depends on the tool. Consumer AI products (ChatGPT free, Copilot in personal mode) typically use inputs for model training. Enterprise agreements with OpenAI, Anthropic, and similar providers typically exclude training use. Attorneys must verify the data processing agreement before uploading client documents. Purpose-built legal AI platforms with explicit data protection commitments are safer than general-purpose tools.

What is RAG and why does it matter for petition drafting?

RAG (Retrieval-Augmented Generation) is an AI architecture where the model generates text only from documents you provide, not from its training data. For petition drafting, this means every factual claim in the AI output is traceable to a specific client exhibit. RAG-based systems are structurally different from general-purpose AI — hallucination is prevented by design, not just reviewed-for afterward.

Do I need to disclose AI use to USCIS?

USCIS has not issued formal guidance requiring disclosure of AI use in petition preparation. However, some courts and federal agencies are developing AI disclosure requirements. The ABA's guidance is that attorneys must supervise AI output and take full responsibility for it. Whether or not disclosure is required, the attorney-of-record signs the petition and is accountable for its accuracy.

What state bars have issued guidance on AI use in legal practice?

Several state bars have issued formal ethics opinions on AI use, including California (State Bar Formal Opinion 2023-204), Florida, New York, and others. Most follow the ABA framework: competency in the tool, supervision of output, client confidentiality protection, and candor obligations. The specifics vary by jurisdiction — attorneys should consult their state bar's ethics opinions before deploying AI in client matters.

AI in Legal Practice

Is AI Safe for Your Immigration Practice? A Frank Assessment

The real risks of using AI for EB1A and O-1 petition drafting — hallucination, bar ethics, data confidentiality — and a framework for evaluating AI tool safety.

April 29, 2026·Updated May 6, 2026·11 min read

Every week, more immigration attorneys are experimenting with AI for petition drafting. Most start with ChatGPT or Claude, ask it to draft a criterion section, and are impressed by how plausible the output sounds.

"Plausible" is exactly the problem.

This article is a frank assessment of the real risks of AI in EB1A and O-1 petition preparation — hallucination, bar ethics, and data confidentiality — and a framework for deciding which tools are actually safe to use in a professional practice.

3 tiers

AI tool risk categories

High-risk (consumer AI), moderate-risk (enterprise AI), and appropriate-for-legal (purpose-built with fact grounding)

ABA Rule 1.1

The competency obligation

Attorneys must understand the tools they use — including how AI fails, not just how it succeeds

RAG

The correct architecture

Retrieval-Augmented Generation grounds every AI claim in actual uploaded evidence — structurally prevents hallucination

The Core Risk: Hallucination

AI language models generate text by predicting what words should follow what came before. They are trained to produce fluent, coherent, authoritative-sounding output. They are not trained to be accurate.

When a general-purpose AI writes petition content, it produces language that sounds like a well-drafted USCIS filing. It will cite the correct regulatory structure (8 CFR 204.5(h)), use the correct legal vocabulary ("sustained national or international acclaim," "final merits determination"), and produce text that an experienced attorney would recognize as competent.

What it will also do, reliably: invent specific facts.

What hallucination looks like in petition drafting

Ask a general-purpose AI to draft a Criterion 5 section for a hypothetical client. In a typical output, you might find:

"Dr. Smith's 2022 paper 'Novel Approaches to Distributed Systems' has been cited 847 times across 23 countries, making it one of the most-cited works in its field over the past five years."

If Dr. Smith has not written this paper, or if the citation count is different, or if this claim appears in a filed petition — the petition now contains a materially false statement. Under 8 CFR 103.2(a)(2), this can result in denial and a finding that the attorney assisted in a fraudulent filing.

This is not a theoretical risk. It is a predictable consequence of using general-purpose AI without a verification layer. The AI does not know it is wrong. It is not lying — it is pattern-matching from training data. The result reads as authoritative and will pass casual review.

Why legal AI is uniquely dangerous for hallucination

Most domains where AI hallucination causes problems are recoverable. A hallucinated restaurant recommendation is a nuisance. A hallucinated fact in a USCIS petition is a misrepresentation — with downstream consequences for the client's immigration history and the attorney's bar standing.

The other factor is specificity. Petition drafts require specific facts: citation counts, award selection rates, salary percentile comparisons, employer revenue figures. These are exactly the types of specific numerical claims that AI models generate confidently and incorrectly. A model that has seen thousands of petition drafts in its training data knows what numbers go in which positions — but not what the actual numbers are for your client.

The confidence-accuracy gap is worst for specific numbers

AI models are most convincing when they are most likely to be wrong. A hallucinated narrative paragraph about a client's "impact on the field" is easy for an attorney to identify as too generic. A hallucinated specific — "cited 847 times," "one of 12 finalists selected from 3,400 applicants" — is much harder to catch because it sounds precise. These are the hallucinations that reach USCIS.

Cracked legal document representing the hallucination risk in AI-generated EB1A petition content

The Three Categories of Risk

1. Factual hallucination

The AI invents or embellishes specific facts: citation counts, award details, salary figures, company metrics. The attorney does not catch it because the language sounds authoritative and the attorney is reviewing many pages of AI-generated text.

Mitigation: Every factual claim in AI-generated petition text must be verified against an actual exhibit before filing. This is not optional — it is the baseline minimum.

2. Legal hallucination

The AI cites non-existent cases, inverts legal standards, or misapplies the Kazarian v. USCIS, 596 F.3d 1115 two-step. Immigration law has enough nuance that errors are easy to miss for attorneys who are not deeply familiar with current adjudication trends.

Mitigation: Attorney review of all legal analysis. AI-generated legal arguments should be treated as a starting point, not a final product. For the correct Kazarian framework, see the Kazarian standard complete reference.

3. Data confidentiality exposure

When you paste client documents into a public AI interface, you are sending your client's personal information, immigration status, employment history, and salary data to a third-party server. Whether that data is used for model training depends on the product and the agreement.

ABA Model Rule 1.6 requires attorneys to take "reasonable measures" to prevent the disclosure of client information. Using a free consumer AI tool to process client documents does not meet this standard.

Mitigation: Use only tools with explicit data processing agreements that prohibit training on client data. Verify the agreement — do not assume.

Bar Ethics: What the Rules Actually Require

Competency — ABA Rule 1.1

ABA Model Rule 1.1 requires attorneys to maintain competency in the tools they use. Comment 8 to Rule 1.1 was updated to include "the benefits and risks associated with relevant technology." Using AI competently means understanding what it can and cannot do — including understanding that AI output must be reviewed, not trusted.

Supervision — ABA Rule 5.3

ABA Model Rule 5.3 addresses responsibilities regarding non-lawyer assistance. AI tools function as non-lawyer assistants under this framework. The supervising attorney is responsible for the output. "The AI drafted it" is not a defense for an inaccurate petition.

Confidentiality — ABA Rule 1.6

Before uploading client documents to any AI system, attorneys must:

Verify the data processing agreement
Confirm client data will not be used for training
In some jurisdictions, obtain client consent for third-party processing

Candor to tribunal — ABA Rule 3.3

ABA Model Rule 3.3 is the most direct rule in the USCIS context. Filing a petition with materially false statements — even unknowingly — is a candor violation. The obligation to verify AI-generated content flows directly from this rule.

State bar guidance is jurisdiction-specific — check yours

Several state bars have issued formal ethics opinions on AI use, including California (Formal Opinion 2023-204), Florida, and New York. Most follow the ABA framework but vary in specifics — particularly on disclosure obligations to clients and whether consent is required before processing documents in third-party AI systems. Check your state bar's ethics hotline or published opinions before deploying AI in client matters.

Scales of justice representing the bar ethics balance between AI utility and attorney responsibility in immigration practice

Evaluating AI Tools: A Three-Tier Framework

Not all AI tools carry the same risk for legal work.

AI tool risk tiers for immigration petition drafting — evaluation framework
Criterion	Regulatory Name	2024–2025 Pattern	Risk Level
T1	High-risk — not suitable for petition content	Consumer AI (ChatGPT free, Copilot personal): no enterprise data agreement, generates from training data only, cannot trace claims to client exhibits, may use inputs for model training.	High risk
T2	Moderate-risk — use with caution	Enterprise AI (ChatGPT Enterprise, Claude Pro) with data protection agreements. Still generates from training data — hallucination risk remains. Requires manual attorney verification of every factual claim.	Moderate
T3	Appropriate for legal work	Purpose-built legal AI: explicit data protection agreements prohibiting training use, fact-verification against uploaded evidence (every claim traced to a source exhibit), attorney review workflow, audit trail.	Strong

The key distinction between Tier 2 and Tier 3 is automatic fact grounding. A Tier 3 tool does not merely generate content — it generates content that is demonstrably traceable to the client's actual documents. Hallucination is structurally prevented, not just reviewed-for.

What "Hallucination-Free" Actually Means

When a legal AI tool claims to be "hallucination-free," interrogate what that means in practice.

Marketing claim: "Our AI doesn't hallucinate."
What to ask: "How is that verified? Can you show me which exhibit each claim in the petition letter is cited to?"

True hallucination prevention in petition drafting requires Retrieval-Augmented Generation (RAG): the AI generates content only from the documents you upload, not from its training data. Each claim is explicitly tied to a specific passage in a specific exhibit. For the technical architecture behind this, see how RAG powers EB1A petition drafting.

This is mechanically different from a language model that "tries to be accurate." A general-purpose model has no mechanism to know that the citation count it generated is incorrect — it is making its best prediction based on patterns in its training data. A RAG-based system cannot generate a citation count that does not appear in an uploaded exhibit, because it only generates from what it can see.

The question to ask of any AI tool claiming to be safe for petition work: "Where, specifically, is each factual claim in this output coming from?" If the answer is anything other than a specific uploaded document, the hallucination risk is not controlled.

RAG is an architecture, not a feature — verify what the tool actually does

Some AI tools describe their approach as "grounded" or "evidence-based" without implementing true RAG. Grounding in this context means the model was shown the documents and asked to stay consistent with them — but the model can still generate from its training data if the question is underspecified or the document is silent on a detail. True RAG systems can show you, for each sentence in the output, which passage in which document it was drawn from. If a tool cannot do this, hallucination is managed, not prevented.

A Practical Decision Framework

Before using any AI tool for petition work, answer these questions:

1. Where is the client data going?

Is there a data processing agreement?
Is training use explicitly excluded?
Is the data encrypted in transit and at rest?

2. How are claims verified?

Can I trace every factual claim in the AI output to a specific client exhibit?
Does the tool automatically cite sources, or do I have to manually check?

3. Who is responsible for errors?

Does the tool have an attorney review workflow?
Is there an audit trail showing what the AI generated vs. what I approved?

4. What does "AI-generated" mean in my jurisdiction?

Some courts and agencies are beginning to require disclosure of AI use. USCIS has not yet issued formal guidance, but this is an evolving area. Monitor your state bar's published ethics opinions.

Magnifying glass over a document stack representing attorney due diligence in evaluating AI tools for immigration practice

The Bottom Line

AI can safely accelerate EB1A petition preparation — but only with the right architecture. The risk is not that AI is inherently dangerous for legal work. The risk is that most readily available AI tools are not designed for legal work and create hallucination and confidentiality risks that attorneys may not anticipate.

The standard an attorney should hold any AI tool to: would I be comfortable explaining exactly how I used this tool if the petition were audited?

If the answer is yes — because every AI-generated claim is traceable to a client exhibit, the data never left a secure environment, and the attorney reviewed and approved every word — then AI is a legitimate and powerful tool for EB1A practice.

If the answer is no, the tool is not ready for use in filings. For the full picture of how AI-assisted preparation works end-to-end, see the EB1A petition drafting efficiency guide.

Immigration Copilot is built to the Tier 3 standard: every claim is traced to an exhibit, client documents are never used for training, and the attorney approves before filing. Get started →

EB1A Practice Tips

Get bimonthly guides for immigration attorneys

Criterion deep-dives, workflow tips, and USCIS updates. No spam. Unsubscribe any time.