NIW for AI/ML Researchers: What the AAO Data Actually Shows

There is no "AI researcher" box on the I-140

If you work in machine learning and you are reading petition guides, you have probably seen claims about an "AI researcher approval rate." There is no such number, and anyone quoting one is making it up. USCIS does not adjudicate petitions by job title. It looks at the proposed endeavor and the evidence, and the AAO classifies appeals into broad occupational groups when it publishes decisions.

AI and ML work lands in Information Technology & Computing. That is the bucket your petition shares with software engineers, data engineers, cloud architects, and IT specialists. It is the largest single category in the AAO appeal record we maintain.

What the appeal record shows for IT & Computing

Our database holds 6,362 AAO decisions on EB-2 National Interest Waiver appeals, spanning 2015 to 2026. Every one of these is an appeal — a petition USCIS already denied at least once. Across the whole set, 375 appeals were approved (5.9%), 5,292 were dismissed (83.2%), and 533 were remanded for further review (8.4%).

Within Information Technology & Computing, the AAO sustained 77 of 1,069 appeals — a 7.2% approval rate. Two narrower slices we can isolate:

• Software Engineering: 5 of 74 appeals approved (6.76%)
• Information Technology: 11 of 185 appeals approved (5.95%)

Read those numbers correctly. They describe petitions that had already failed once and were fighting an uphill appeal. First-pass USCIS approval rates are much higher and are not in this dataset. The appeal record is useful for one thing: it shows you exactly how weak petitions die, so you can avoid building one.

For comparison, Scientific Research — physical and life sciences — sits at 19.3% on appeal (93 of 481). The gap is not because computing work is less important. It is because research petitions tend to arrive with a clearer Prong 1 story. That is the part you control.

If your work is genuinely research, frame it as research

A petition that describes original AI research — published, cited, advancing a method rather than shipping a product — is read against the Scientific Research pattern even though it is filed under computing. The classification follows the evidence. If you publish at NeurIPS, ICML, or ICLR and other groups build on your methods, say so plainly and let the petition read like a research case.

Prong 1: tie your specific subfield to a real policy document

Prong 1 asks whether your proposed endeavor has substantial merit and national importance. AI as a category has been named in federal law and executive action, which gives you material to work with. The mistake is citing "AI is a national priority" in the abstract. USCIS has seen that sentence thousands of times.

Name your subfield and point to a document that names it too:

• National AI Initiative Act of 2020 — establishes AI as a coordinated federal priority
• CHIPS and Science Act of 2022 — semiconductor and AI hardware research
• Executive Order 14110 (2023) — safe, secure, and trustworthy AI
• NIST AI Risk Management Framework — directly relevant to safety and evaluation work
• NSF, DARPA, or DOE strategic plans that name your specific area

AI safety, alignment, interpretability, privacy-preserving ML, and AI for public health all map cleanly onto documents the government has actually published. Pure capability scaling is harder to anchor — not impossible, but you need to connect it to a downstream use the government has flagged.

Prong 2: showing you are well-positioned

Prong 2 is about you specifically: your record, and whether it predicts you will advance the endeavor. For ML researchers the evidence that carries weight is concrete.

Citations, read for context not just count

A raw citation total is a starting point, not an argument. USCIS wants to see that other researchers build on your work rather than cite it in passing. A short analysis that picks out the most significant citing papers and explains how they use your contribution is worth more than a number. Our citation map tool shows the geographic and institutional spread of who is citing you, which is useful for arguing influence beyond your own lab.

Conference acceptance as a quality signal

Acceptance at a top-tier venue is itself peer review. A NeurIPS or ICML paper carries a signal of merit before a single citation accrues. Say which venues and what their acceptance rates are, because the adjudicator may not know.

Open-source adoption

If you released code that other research groups use, that is direct evidence your work advances the field beyond your employer. Document downloads, dependent repositories, and named groups that adopted it. Vague GitHub-star claims are weak; "three university labs depend on this library" is strong.

Expert letters that say something

A letter calling you "a brilliant and talented engineer" does nothing. A letter that says "Dr. X's method for [specific problem] changed how my group approaches [specific task], and I know of other teams that adopted it" does the work. The best letters come from researchers who have cited you, program chairs of venues you contributed to, and government or national-lab scientists who can speak to national relevance.

Prong 3: why the job-offer process should be waived

Many ML petitioners clear Prongs 1 and 2 and then treat Prong 3 as a formality. It is not. Prong 3 asks why it benefits the United States to skip the labor certification process for you specifically. Strong arguments: your research is inherently independent and continues regardless of which employer holds your visa; the pace of the field makes delay genuinely costly; and the specialized talent shortage in your subfield means a labor-market test would not surface a comparable U.S. worker anyway.

Where weak petitions fail

Across the appeal record, dismissed petitions cluster on Prong 1 — failure to establish national importance. The recurring patterns for ML petitioners:

1. Product work dressed up as research. Building AI features at a tech company is real work, but it is not the same as advancing a field. USCIS is skeptical when a petition blurs the line. Be honest about which one you do.
2. An h-index with no context. An h-index of 12 means very different things at different career stages and in different subfields. Give the adjudicator a benchmark.
3. Only internal citations. If your work is cited mostly by your own lab or co-authors, the evidence of field-wide impact is thin. External adoption is what counts.
4. Generic Prong 1. "AI matters" is not an argument. Your specific endeavor, tied to a specific documented priority, is.

What to prepare before you file

• A full publication list with citation counts and a short context analysis of your most-cited papers
• Records of peer-review and program-committee service — invitation emails are fine
• Four to six independent expert letters, none from your employer or close collaborators
• The federal policy documents that name your specific subfield
• Open-source adoption evidence if you have it: dependents, downloads, named adopting groups

Before filing, run a case review to stress-test your three-prong argument against real AAO decisions, and check the profession lookup tool for the IT & Computing breakdown. When you are ready to draft, the petition builder works section by section through the Dhanasar framework.