Can AI Get Immigration Right?
The first independent evaluation of how leading AI models perform on U.S. immigration law
The Problem
Many people are turning to AI for immigration guidance. Immigrants, employers, attorneys, and international students are already using AI tools to navigate one of the most complex legal systems in the United States — asking about visa eligibility, filing deadlines, work authorization, and what to do when their status is at risk.
The consequences of a wrong answer are serious. A missed deadline can mean losing work authorization. Bad advice on status can trigger removal proceedings. Misinformation about a visa category can cost an employer a key hire — or a researcher their opportunity to work in the United States.
Until now, no one had tested whether AI gets these answers right.
What We Did
The Talent Mobility Fund partnered with Vals AI to build the first expert-validated benchmark for AI performance on U.S. immigration law. Working with a team of practicing immigration experts, we developed 519 questions covering the full range of situations real people face — from international students navigating Optional Practical Training (OPT) to Deferred Action for Childhood Arrivals (DACA) recipients trying to understand their options.
We tested 25 leading AI models against 262 of those questions, scoring each against a detailed expert rubric that specifies everything a complete and accurate answer must cover. The remaining 257 questions are available for AI labs to license and use.
What We Found
The knowledge AI needs is fragmented across agencies, courts, and legal practice. Better tools require collaboration between AI developers and the immigration legal community.
The best-performing model, Claude Fable 5.1, answered correctly only 75% of the time — leaving out roughly 1 in 4 pieces of information an immigration expert would consider essential.
For people with DACA or Temporary Protected Status, the least expensive and most accessible models scored as low as 40%.
Models satisfied only 7.3% of essential answers involving government rules from 2025 and 2026, compared to 64.1% for established law.
Every model was better at stating general rules than working out what those rules mean for the specific person asking.
Even the best-performing model with search enabled met only 43.8% of recent-rule items.
Read the Full Report
The full benchmark report includes detailed results by model and persona, annotated example questions, and methodology.
About This Project
This benchmark was developed by the Talent Mobility Fund, a program of Renaissance Philanthropy, in partnership with Vals AI. The evaluation was conducted in October 2026.
For questions about the benchmark, to license the validation set, or to discuss future iterations, contact TMF at [email protected].
Fund
Talent Mobility Fund © 2026
The Talent Mobility Fund is a program of Renaissance Philanthropy.