AI Training Work for Scientists: The Honest Guide
AI labs pay PhD-level experts to create and grade training data in their field. The pay is real, the work is legitimate, and the variability is the part nobody advertises. What to expect and where to sign up.
What this work actually is
Frontier AI labs need experts who can tell a subtly wrong chemistry mechanism from a right one, grade a model's reasoning about experimental design, and write the kind of hard domain questions that expose model weaknesses. That labor market runs through platforms that recruit, vet, and pay scientists as independent contractors. The work is task-based, remote, NDA'd, and paid hourly or per task. It is critical reading and expert judgment, not data entry, which is why advertised rates for PhD-level contributors typically sit in the $30–100+/hour range depending on field and track.
The platforms
Mercor
Apply via referral →A marketplace matching experts to AI lab projects, with per-project applications and interviews. The link lands on the life-science expert track, where roles like Biology Research Scientist currently advertise $60–100/hour, among the highest rates anywhere in this market. More selective per project; worth a profile even if nothing fits immediately.
Handshake MOVE (Handshake AI)
Apply via referral →Handshake's AI training program recruits graduate-level and PhD experts across the sciences for model evaluation and data creation, structured as a fellowship with production work paid hourly. Onboarding includes vetting and an assessment in your domain. Of the two platforms this one leans hardest toward academic credentials, which is an advantage for this site's readers.
Other platforms exist in this market (Outlier, DataAnnotation, Alignerr among them); we list the ones we can vouch for from direct experience.
Prefer talking about your own field to grading model output? Expert networks pay scientists by the hour for phone calls with consultants and investors, at higher rates but far fewer hours. See how GLG works for scientists.
The honest downsides
Task supply is the whole game, and it fluctuates without notice or explanation: a specialty queue that paid for twenty hours last week can be empty this week, so never budget against this income for anything with a due date. Onboarding and assessments are often partly unpaid. You are a 1099 contractor with no benefits, projects end abruptly, and NDAs mean the work produces nothing citable for your CV. Accounts can be deactivated by automated quality systems with limited appeal. And the opportunity cost is real: for an academic scientist, hours spent here compete directly with the papers and applications that move your actual career. Used as a bounded side channel, 5 to 15 hours a week, it is among the best-paying legitimate side work available to scientists right now. Used as a lifeline, it will disappoint.
How this fits your job search
If you are between positions, this pairs naturally with an active search: browse live structural biology jobs, check what roles in your field actually advertise in salary, and join the talent network to get matched while the side work covers the gap.
Frequently Asked Questions
How much do AI training platforms actually pay PhDs?
Advertised rates for PhD-level domain experts typically run $30 to $100+ per hour depending on field, task type, and platform, with specialized science and reasoning tracks at the upper end. Treat any single number skeptically: effective pay depends on task availability, unpaid onboarding time, and how fast you work. Plan on the first weeks paying less per real hour than the sticker rate.
How many hours a week is realistic alongside a postdoc?
Most scientists doing this treat it as 5 to 15 hours a week, evenings and weekends. Task supply fluctuates: some weeks have abundant work in your specialty, some have none. It is dependable extra income on average and undependable in any given week, which is exactly the wrong shape for rent money and the right shape for savings.
Is this allowed alongside my postdoc or university appointment?
Usually yes as independent contracting done on your own time and equipment, but check your appointment's outside-activity policy and any funder rules (some fellowships restrict outside work hours). Never use lab data, unpublished results, or university resources in training tasks, and remember platform NDAs cut both ways.
What does the work actually involve?
Writing and reviewing expert-level questions and answers, grading model responses for factual accuracy, ranking outputs, and red-teaming models in your domain. For scientists that means chemistry, biology, and data-analysis reasoning tasks. It is closer to grant-review-style critical reading than to data entry, which is why they want PhDs and pay accordingly.
Do these gigs help or hurt a scientific career?
Neutral to mildly positive used sparingly: the money is real, and seeing how frontier labs evaluate model reasoning is genuinely educational if you are considering AI-adjacent careers. The risk is opportunity cost. Ten hours a week that should have gone to your own papers is expensive at any hourly rate.