AI for O-1, EB-1, and NIW Applicants (Part 1)

Almost every O-1, EB-1, and NIW inquiry we receive now involves AI. Some applicants want to know how our firm uses AI. Others want to do some of the work with AI and want to know how this affects the timing and the cost of the petition preparation. And some come to us with already prepared AI-assisted drafts of personal statements, recommender letters, and entire petitions that they want us to review.  

AI is, without any doubt, incredibly useful in petition preparation. It can save a lot of time and improve accuracy – if used right. It can also create problems that cost the applicant time, money, and credibility with the officer reading the file. Every AI draft requires heavy human editing before it goes into the file.

This two-part series comes out of a webinar I presented on the subject. Part 1 covers how to use AI well when you are preparing your own petition materials: which tools to use and how to draft recommender letters and a support letter that do not read as machine-written. Part 2 covers the use of AI when you work with counsel, including how to vet an attorney's AI fluency, how to utilize AI efficiently, and how to avoid collaboration mistakes that raise, not reduce, your legal bill.

Part 1: Preparing Your Petition With AI

Start with an honest picture of what AI can and cannot do

Use AI as a research and drafting assistant. Do not use it as a legal advisor.

AI cannot assess your eligibility for a specific visa type or recommend the best visa pathway. It cannot predict how an adjudicator will behave. It cannot replace judgment about strategy, which is where petitions are won or lost. AI gets information from online sources, some are reliable and some are not. Most importantly, this information is limited. Most valuable information in is the heads of experienced attorneys who have been in business for decades, it never makes it to the internet. If you intend to prepare your petition by yourself, do yourself a favor and run your strategy by at least one experienced lawyer before starting to prepare the petition.

What AI can do, and do well, is accelerate drafting, organization, and translation, all under your direct review. That last clause is the key: you must read, verify, and rewrite what the model gives you. AI is also excellent in reviewing large amounts of information and identifying potential issues or internal inconsistencies.

Choose the right tools for drafting

Whatever tool you choose, use a paid tier that does not train on your data (if there is an option for using your data for training, turn it off). Use due diligence to understand where your data is stored and how it is used by each tool, and make sure you’re comfortable with your choice. For drafting, I recommend using the best available model and extended thinking turned on. Free tiers give you weaker models and, more importantly, less control over your data.

For long-form legal writing, which includes personal statements, cover letters, recommender letters, and RFE responses, Claude is my personal preference as of the date of this post. New models keep coming out, and it is important to continue testing new ones. For now, I find Claude’s prose more natural and showing less of the generic “AI voice.” It follows complex, multi-part instructions more faithfully. And it handles long context well, which matters enormously here: with a tool like Cowork you can put your full CV, the criteria framework, all of your evidence, and your prior drafts into a single session and have the model work across all of it at once.

For research, use only tools that link to their sources

Require clickable citations, and treat those links, not the model's conclusions, as your source of information.

Perplexity is good for this, though Claude, ChatGPT, and other systems have comparable modes. Again, use the strongest model with extended thinking. Avoid any AI legal research that produces answers without verifiable links, and open every source to confirm it actually says what the AI claims it says.

Used this way, research is one of the highest-value AI applications. AI is very good at helping you build out the record on the impact of your projects, the standing of the organizations you have worked for, the reach of the outlets that published your articles, the selectivity of the hackathons you placed in, the admission standards of your memberships, or the existence of federal and state government programs supporting your NIW endeavor.

Decide on confidentiality before you paste anything

Before anything goes into a chat window, decide what level of protection the material needs.

Never paste passports, A-numbers, addresses, or financial and medical records into consumer AI without first checking the data settings. Use enterprise or paid tiers with no-training options for sensitive content. Strip or redact identifiers, including metadata in documents you upload.

Also keep in mind that a model can assemble your identity from fragments across a long session even when no single message identifies you. Read each tool's retention and training policy before your first use. They differ meaningfully, and they change.

Let AI brainstorm the mapping to criteria

AI is genuinely useful for suggesting which regulatory criteria each item in your record could support. Treat that output as a brainstorming starting point rather than a final classification. Ask the model to flag weak fits alongside strong ones, and cross-check by running the same prompt in a fresh session or in a different system for a second pass.

Put your technical work into plain language

This is also one of the best uses of AI in the petition preparation process. Officers are not specialists in your field, and to convince them that your work is important you need first to explain it in clear terms.

Ask for an explanation of your scientific paper to a non-specialist in 200 words. Then ask a second question: why does this research matter to the field? Correct the result using your own technical knowledge and common sense, and then use those explanations in both your support letter and your recommender letters.

Recommender letters

Target length is three to five pages per letter.

The key to success is to use AI in a way that it is not obvious, which means that every AI draft needs to be reviewed and substantially revised to achieve accuracy and convincing human voice. Officers have now read an enormous volume of AI-generated recommender letters, and they recognize the patterns instantly. They also have their own AI systems. A letter that reads as machine-produced is a credibility risk, and credibility is the currency of these petitions.

The AI tell signs show up in vocabulary, sentence structure, and rhythm. Here is what to avoid:

• Heavy use of em dashes.
• The opener “In my [X] years of experience…”
• “Not only… but also…” structures
• Climbing tricolons: “dedicated, brilliant, and truly exceptional.”
• The closer “I give my highest recommendation without reservation.”
• Restating the same accomplishment in different words across the letter. Every paragraph must introduce a new fact.
• Assumptions of any kind. The model should stay strictly inside the facts in your evidence and your input.
• The “it was not just X, it was Y” construction and its many cousins.

That last one is worth an example, because it is everywhere:

AI draft: The product was not a demonstration or a limited pilot. It was designed and built for full production deployment under live client contracts.

My rewrite: Mr. X designed and built this product from scratch and led it to full deployment under live client contracts.

Below, are my suggestions for what to use:

• Active voice. “Dr. X redesigned the protocol,” not “the protocol was redesigned by Dr. X.”
• The reading level of a bright 16-year-old. Short sentences, concrete nouns, no abstraction without an immediate example.
• A clear structure. First, who the recommender is and how they know you. Then one specific accomplishment per paragraph, each followed by a paragraph explaining why it matters, either to the field (for original contribution or general recommendation) or to the company where you worked (for critical role). Then a direct statement of recommendation tied to your visa category.
• Specificity. Include exact numbers, dates, locations, project names, specific quantifiable outcomes. Generic letters are not persuasive.
• Variation across letters. No verbatim repetition, and no shared sentence structures, opening lines, or characterizing adjectives between letters in the same case. A useful check: run all drafts through one AI session and ask it to identify similarities and suggest changes before they get singed.

Then edit. The AI draft is a scaffold, not a letter. The recommender must review the draft with extreme care and rewrite by hand to make the letter sound human, and also add their own phrasing, an anecdote, or an observation only they could make. That is what makes a letter convincing.

The petition support letter: organize the evidence first

Target length is 15 to 30 pages.

Before you write a word of the support letter, build and number your exhibits, map them to criteria, and produce a table of exhibits.

Then prompt the AI to draft the letter. I recommend drafting section by section rather than all at once:

• Introduction (one to two pages): the beneficiary, the petitioner, and a brief statement of why the petition meets the standard.
• The beneficiary's field (about half a page): a plain-English explanation. Skip this if your field is common and easily understood.
• Criterion-by-criterion sections.
• Final merits and totality argument (two to three pages): for EB-1A, this is where the sustained acclaim argument lives.
• Conclusion (about half a page): a direct request for approval.

Avoid padding with repetitive summary paragraphs, and apply the same anti-AI-tell rules you used for the recommender letters. Treat the output as a first rough draft – review it with extreme care, make sure that all evidence is addressed accurately and thoroughly, and then rewrite and polish by hand.

One prompt worth running at the end, on the assembled package: identify any inconsistencies across this document and the exhibits, including but not limited to dates, names, exhibit numbers, or factual claims. This is a task machines do better than humans, and inconsistencies are a common trigger for an RFE.

Everything above assumes you are working on your own materials. Part 2 of this series tells how AI changes the process of preparing the petition in collaboration with an AI-savvy lawyer.