STEM AI training jobs: tasks, pay and who qualifies

What STEM AI training work involves

STEM AI training jobs pay scientists to write, solve and check the hard problems AI models are measured against. On 29 September 2026 this site listed 64 roles under Science & Math, out of 699 listings in total: 30 in chemistry, 26 in physics, 11 in biology and the life sciences and 4 in mathematics, with some listings counted in two fields.

The work comes in four main forms.

Solving problems to reference standard. micro1's physics family, such as its Physics Expert (PhD/Postdoc) listing, has you solve advanced problems in your own subfield with full derivations written up in LaTeX, checking steps in SymPy or Python.

Writing problems. micro1's Mathematics Expert writes advanced problems with complete proofs. A problem for a model has to admit exactly one defensible answer, and the proof has to be complete enough that a grader can check a model's attempt against it.

Coding benchmarks. Mercor's "Sci Code" listings, such as Chemistry PhD Coding Experts, ask you to turn a paper, dataset or repository into a scientific computing problem with grading criteria.

Red-teaming. Eight Mercor chemistry listings, including Analytical Chemist, pay you to write prompts labelled benign, dual-use or adversarial, then judge whether the model answered or refused correctly. Refusing a routine professional question counts as a failure, just as answering a harmful one does.

Bench scientists have a place too. micro1's biologist listing wants hands-on CRISPR, cloning and PCR experience, and the job is turning lab knowledge into structured material a model can learn from.

What it pays

56 of the 64 listings quote an hourly rate, from $23 to $200, with a median midpoint of $80. Physics pays best: its median midpoint is $117.50, and micro1's PhD physics listings reach $150–200. Biology sits around $85 and chemistry around $70. The eight chemistry red-team listings pay $65–75 per task rather than per hour, so what you earn depends on your pace.

Mercor also runs a bilingual STEM PhD family for AI safety work in 12 languages, where the rate follows the language market: $77–81 for Norwegian, $23–27 for Hindi. Our bilingual guide explains that pattern.

Who qualifies

A doctorate is the norm. 47 of the 64 listings mention a PhD, and 46 are rated senior. micro1's physics listings ask for recent publications with arXiv or DOI links. Mercor's coding benchmarks want Python, Git and Docker on top of the PhD.

There are exceptions. micro1's mathematics listing accepts a BSc or MSc, its biologist listing a bachelor's degree plus deep lab experience, and its Life Sciences Research Assistant takes a bachelor's or master's with two years of research, at $30–50 paid per accepted task.

Mercor posts 40 of these roles and micro1 24. 57 of the 64 are open worldwide.

How STEM assessments are judged

Most of these roles start with a subject test or a sample task, and the graders read your working as closely as your answer. A model can reach the right number through a wrong step, and catching that is the job, so the assessment looks for the same habit in you. Our guide to qualification tests covers the format; in a science or maths answer, four things carry the most weight.

State your assumptions and units. Say what you treat as constant, negligible or ideal, whether that is air resistance in a mechanics problem, temperature in a reaction or independence in a probability question, and carry units through every line.

Write out the reasoning chain. Each step should follow from the one before, so a reviewer can check it without solving the problem again.

Test the result against what you already know. In physics and chemistry, dimensional analysis and limiting cases (what happens as a mass goes to zero or a concentration grows large) catch most slips. In maths, try a special case or the boundary conditions; in biology, check the claim against the mechanism and the controls.

When you mark a model's answer wrong, say where the error starts. "Step 3 treats acceleration as constant, which stops holding once drag is included" gives the grader something to act on.

Where to start

Browse the science and maths filter by field. If your work sits closer to engineering practice or data analysis, see our guides to engineering and data science and machine learning roles.