U.S. visas for AI engineers

Which parts of an ML career actually map onto the criteria

The criteria at 8 C.F.R. § 214.2(o)(3)(iii)(B) for O-1A and § 204.5(h)(3) for EB-1A were written long before modern machine learning, which means the mapping is not obvious and a lot of strong candidates undersell themselves.

Authorship of scholarly articles is the most straightforward. Papers at NeurIPS, ICML, ICLR, ACL, CVPR, and their peers count, and in this field conference proceedings carry more weight than in most disciplines — a point worth making explicitly in the petition, because an adjudicator without domain knowledge may not know that a top conference outranks many journals here. Preprints on arXiv can support the criterion when accompanied by evidence of standing: citations, adoption, or venue acceptance.

Judging the work of others is satisfied by conference and workshop reviewing, and most working researchers have done it without thinking of it as a credential. Area chair or program committee membership is stronger still. Keep the confirmation emails and the reviewer-assignment records — reconstructing them years later is tedious and sometimes impossible.

Original contributions of major significance is where AI engineers have unusual advantages and usually fail to use them. A model card on Hugging Face with download counts, a repository with meaningful stars and dependent projects, a technique adopted into a widely used framework, a benchmark others now evaluate against, an architecture cited and built on — these are direct, quantifiable evidence of independent adoption. That is exactly what this criterion asks for, and it is more persuasive than a citation count divorced from context.

Critical or essential capacity applies to a role at a recognised lab or a well-funded company, but remember it needs two showings: your role, and the organisation's distinguished reputation, each evidenced separately. Naming a well-known employer is not evidence of its standing — funding, market position, independent coverage, and third-party rankings are.

High salary is often satisfiable in this field and rarely documented properly. Compensation for senior ML roles is well above the general software wage, but the criterion requires a comparison, not a number. Pair your offer letter or payroll evidence with DOL OES data or a credible industry compensation survey for the same role, level, and metro. Equity is difficult to use here without a defensible valuation.

• Top-tier conference papers — and explain to the adjudicator why the venue is top-tier in this field.
• Peer review, program committee, and area chair work, with confirmation records kept.
• Model and repository adoption metrics: downloads, dependents, forks, downstream products.
• Framework or benchmark adoption where the field now builds on your work.
• Salary benchmarked against wage-survey data for the same role and location.

What USCIS discounts, and why

Several things that feel impressive inside the industry do very little in a petition, and knowing that in advance saves you from building a case around them.

Employer prestige alone. Working at a famous lab is not a criterion. The role and the organisation's reputation each need independent evidence, and a recognisable name substitutes for neither.

Raw citation counts. A number without analysis says nothing an adjudicator can use. What works is context: who is citing the work, in what venues, and what they did with it. A hundred citations that show a technique being adopted into production systems is a stronger showing than a thousand that show routine background referencing.

GitHub stars in isolation. Stars are a popularity signal that adjudicators have learned to discount. Dependent repositories, downstream products, package download counts, and named organisations using the work are the evidence that carries.

Internal impact. A system that transformed your own employer's operations, evidenced only by your employer, satisfies almost nothing. The criterion asks for significance to the field. Find the independent adopters.

Letters from your coauthors and managers. The most common weakness in AI petitions is a stack of glowing letters from people who worked directly with the beneficiary. Independence is what makes a letter persuasive — a smaller number of detailed letters from experts with no coauthorship or employment relationship, who can describe the specific work and its effect, is worth considerably more.

O-1A, EB-1A, or EB-2 NIW

Most AI engineers at this stage are choosing between three categories, and they test genuinely different things.

O-1A is the temporary work visa and usually the first step. Three of eight criteria, no annual cap, no lottery — but it requires a U.S. employer or agent as petitioner and cannot be self-petitioned. Founders can often petition through their own company, though only where someone other than the beneficiary can hire, supervise, and terminate.

EB-1A is the self-petitioned green card. The criteria list overlaps heavily with O-1A, but it is adjudicated against a materially higher standard of sustained national or international acclaim, and the final merits determination is where these cases are decided. A record that clears O-1A comfortably is frequently not yet enough for EB-1A, and there is no shame in holding O-1A status for a few years while the record matures.

EB-2 NIW asks a different question entirely: not how acclaimed you are, but how nationally important your proposed endeavor is and whether you are positioned to advance it. For AI engineers working on something with clear public relevance — safety and evaluation, healthcare applications, critical infrastructure, defence-adjacent work — NIW is often the better fit even where the acclaim record is thinner. The failure mode is describing the endeavor as an occupation. "Research in artificial intelligence" fails prong one; a specific problem with a specific beneficiary population and a plausible scale does not.

One practical note on sequencing: EB-2 sits in a more oversubscribed visa category than EB-1, so for applicants born in India or China the NIW route can mean a substantially longer wait for a visa number even though the petition standard is lower. That trade-off is worth working through with an attorney before choosing.

If you are on an H-1B or OPT right now

A lot of AI engineers are reading this while on F-1 OPT or STEM OPT with an H-1B lottery outcome pending, and the timing questions are more urgent than the category questions.

The H-1B is cap-subject for most employers, allocated by registration lottery in the spring, with employment starting no earlier than October 1. Miss the window or lose the lottery and you wait a year, regardless of your record. That is exactly why O-1A is worth evaluating in parallel: it has no cap, no lottery, and no seasonal window, so it can be filed whenever the evidence is ready.

Cap-exempt employment is the other route around the lottery — universities, affiliated nonprofits, nonprofit research organisations, and governmental research organisations can file year-round. For researchers this is a genuine option rather than a technicality.

Both H-1B and O-1A tolerate dual intent, so you can pursue an EB-1A or NIW while holding either. Start the green card process earlier than feels necessary: the evidence assembly alone — soliciting independent letters, running adoption analysis — routinely takes months before anything is filed.

The categories in full

Each guide covers who qualifies, what evidence satisfies the standard, how the process runs, what the government charges, and the RFE patterns to avoid — with the regulations cited.

Common questions