Why AI work fits the criteria
The O-1A visa and the EB-1A green card both ask whether a person has risen to the top of their field, measured through a set of criteria in the regulations. Machine learning produces exactly the kinds of evidence those criteria were written around: peer-reviewed papers, peer review itself, citations, adoption by others, and high pay. USCIS’s own guidance goes further and mentions AI by name.
- The Policy Manual chapter on O-1 beneficiaries gives fellowship in the Association for the Advancement of Artificial Intelligence as an example of a qualifying membership, because fellows are chosen by a panel of current fellows for significant, sustained contributions.
- The EB-1A chapter counts published presentations at nationally or internationally recognized conferences as scholarly articles, which matters in a field whose leading venues are conferences.
- The NIW chapter names certain critical areas of artificial intelligence as an example of a critical and emerging technology.
The criteria, mapped to machine-learning work
The O-1A has eight criteria under 8 CFR 214.2(o)(3)(iii) and the EB-1A has ten under 8 CFR 204.5(h)(3); both require evidence under at least three. The two extra EB-1A criteria cover artistic exhibitions and performing-arts sales, which rarely apply to AI work, so the eight below are the working set for both.
| Criterion | What counts in ML | What falls short |
|---|---|---|
| Awards | Best-paper or outstanding-paper awards at recognized conferences; selective national fellowships; major competition wins with documented selectivity | Travel grants, internal company awards, student scholarships without a selective field-wide pool |
| Membership | Fellowships elected by existing fellows for outstanding achievement, such as AAAI or IEEE Fellow | Paid or general membership in a professional society |
| Published material about you | Coverage of your work in major media or trade press that names you, with circulation or readership figures | Your employer’s blog or press release; coverage that never mentions you |
| Judging | Completed reviews for major conferences or journals; area chair or program committee service; grant panel review | Reviewer invitations with no proof the reviews were done |
| Original contributions | A method, model, dataset, or library that others adopted, shown through citations in context, downstream use, and letters naming the impact | Repository stars or download counts with no evidence of who relies on the work |
| Scholarly articles | Peer-reviewed papers at recognized conferences and journals | A record built only on preprints |
| Critical role | Lead or key role on a research team at a distinguished lab, university, or company; founding a recognized company | A title without evidence of what you led or decided |
| High salary | Total compensation, including equity grants, compared with wage data for the same occupation and location | Comparing against a broad national average for all workers |
Peer review is often the easiest criterion to document
USCIS lists reviewing papers for scholarly conferences and serving as a peer reviewer for publications as examples of judging the work of others. Many ML researchers review every year without thinking of it as evidence. The petition has to show the reviews were completed, so the useful records are the invitation, the assignment, and the confirmation of submission from the conference system.
Original contributions need proof of adoption
This criterion asks for contributions of major significance to the field, and it is where ML petitions most often overreach. A popular repository is a start, but the petition needs to show who uses the work and what it changed: papers that build on the method, companies or projects that depend on the library, benchmarks that adopted the dataset. Letters from independent experts should name the specific contribution and its effect, as the guide to original contributions explains.
Salary is usually strong, if compared correctly
Machine-learning compensation is high, but the comparison has to be with the right group. Wage data for the specific occupation and area, such as the Bureau of Labor Statistics figures for computer and information research scientists, or a reputable industry survey for the role and level, is more persuasive than a national average for all workers.
Records that are hard to recreate later
- Reviewer invitations and submission confirmations for every conference and journal
- Citation exports with notes on how later papers used your work, not only counts
- Evidence of downstream use: dependent projects, production deployments, adoption announcements
- Press coverage that names you, with the outlet’s readership or circulation
- Offer letters, equity grants, and pay records
When you work in industry
Researchers at companies often publish less than academics, and USCIS allows for that. When a criterion does not readily apply to an occupation, the petition can offer comparable evidence. USCIS’s own example is that presenting work at a major trade show can be comparable to scholarly articles for someone in industry, and that highly valued equity can stand in for salary for a startup founder. The petition has to explain why the listed criterion does not fit and why the substitute is of equal weight. Industry ML engineers commonly build their case on original contributions, a critical role at a distinguished company, and high pay, with technical talks and patents in support.
The final merits question
Meeting three criteria is the threshold. USCIS then weighs the whole record to decide whether it shows sustained acclaim at the top of the field, and its guidance lists factors that suit researchers: a citation rate or h-index that is high relative to the field, papers in particularly highly ranked venues, work at leading institutions, unsolicited invitations to speak at major conferences, and a named role on a competitive government grant. Comparative data matters at this step, since a citation count means little until it is set against typical researchers at the same career stage in the same subfield.
Gaps that draw requests for evidence
- Team papers with no explanation of the applicant’s own role in the work
- Adoption claims backed only by stars, forks, or downloads
- Reviewing claimed without proof the reviews were completed
- Citation counts with no comparison to the field
- Support letters from close collaborators only, with none from independent experts
Sequencing the filings
Approval odds differ sharply between the two. USCIS approved 91.0% of O-category petitions in the third quarter of FY2026, against 32.2% of EB-1A petitions, according to its published data. Many AI researchers secure an O-1A first, which needs no lottery and can be filed through an employer or agent, and file for the EB-1A once the record has grown, a path covered in the O-1A to EB-1A bridge. The EB-2 national interest waiver is a parallel option with a different test. Its focus on a specific proposed endeavor, and USCIS’s treatment of certain areas of AI as critical and emerging technology, often make it a strong second filing, as the guide to proving national interest explains.