Frontier Labs Want Your PhD: Research-Grade AI Training Work
2026-07-28 · Expert Match AI team
If you have a PhD - or you're partway through one - the AI training market has a specific job for you, and it isn't labeling data. Frontier labs need people who can work *at the edge of what their models can do*: writing problems the model fails, grading proofs the model attempts, and explaining exactly where the reasoning breaks. We currently track 31 live science roles averaging $77/hr across priced listings, with the top academic tier reaching $300/hr.
Why labs pay for research-grade judgment
Models have long since saturated textbook problems. What a frontier lab can't generate internally is a supply of *novel, verifiable, genuinely hard* questions - and expert judgment on the model's attempts at them. That's a research skill, not an annotation skill: designing a problem with a clean statement, a hidden subtlety, and a checkable answer is exactly what you do when you write a qualifying exam or referee a paper.
The work splits into a few recognizable lanes. Problem authoring: writing physics, math, chemistry, or biology problems hard enough that the model fails informatively - the scarcest skill on this list and the reason the top tiers pay what they do. Solution grading: evaluating model-generated derivations and proofs step by step, flagging where an argument goes wrong and writing the rationale the model trains on. Red-teaming and safety review: domain experts stress-testing what models will say in sensitive technical areas. If you've TA'd, refereed, or written exams, you've already done the core motions.
What the tracks pay right now
The physics track is the deepest. micro1 currently lists a full academic ladder: PhD/postdoc at $80-$140/hr, postdoc/junior professor at $80-$150/hr, and professor/principal-investigator and high-energy/mathematical-physics tiers at $80-$160/hr - the rate literally steps up with your CV. Above that sits Surge's Professors & PhD Researchers program at $150-$300/hr, the highest-paying academic listing on our board.
Chemistry and biology run a tier lower but are hiring: Chemistry Professor/Researcher (PhD) at $70-$90/hr, plus specialist listings in toxicology and related areas around $50-$90/hr. On the research-scientist end, Mercor lists an LLM Research Scientist (pre-training & post-training) at $100-$120/hr. And for grad students who want volume rather than prestige tiers, DataAnnotation's Math & STEM projects ($40-$60/hr) and Outlier's Math AI Trainer ($30-$50/hr) are the accessible on-ramps.
Full context is in the live salary report: science averages $77/hr across 23 priced listings, mid-pack among fields - but that average hides the spread. The floor is generalist eval work; the ceiling is credentialed problem authoring, and the gap between them is roughly 4x.
How screening works for academics
Platforms screen academics on credentials first, then a work sample. Your CV line does real work here: listings distinguish "PhD," "postdoc," and "professor/PI" explicitly, so state your degree, institution, field, and publication area in exactly those words - automated screeners match strings, not prestige. Expect a paid or timed qualification task: typically authoring one problem or grading a few model solutions against a rubric. The bar isn't brilliance; it's whether your problem is *well-posed* and your grading is *consistent*. Our physics and math skills pages show what the market currently asks for by name.
Two honest caveats. First, queues fluctuate: a physics track that's hiring this month may pause next month, which is why researchers - like everyone else in this market - do better holding 2-3 platforms than betting on one. Second, this is contract work with no benefits, and universities have consulting-disclosure policies - check yours before you start, especially if you're on a visa or grant-funded.
Where to start
If you're research-active, start with our PhD researchers page and apply directly to the tier that matches your CV - the professor tiers exist precisely because labs will pay more for your judgment than a generalist queue will. If you're mid-PhD, anchor on an accessible STEM queue, bank the assessment pass, and move up-tier as your credential does. All 31 live science roles are on the board now, sorted by pay - or upload your resume and we'll rank them against your background.