Math & STEM

If you have real mathematical training (a degree in progress, a master's, a PhD, or years of quantitative work) and you want a remote math job that pays you to use it from home, STEM, math and proofs is the single highest-paying category in the AI training economy. Qualified people earn roughly $30 to $200 per hour for it, and the very top of the market (frontier-lab proof verification, research-grade reasoning) reaches $200 to $250 per hour on platforms like AfterQuery and Handshake AI. That is not a get-rich-quick pitch. It is a real, if lumpy, freelance market where credentials and the ability to write a clean proof are worth more than in almost any other online gig. This guide explains exactly what the work is, who pays what, how to get in, and where the traps are, so you can decide whether it is worth your time.

What "STEM, math & proofs" AI training actually is

Large language models are famously unreliable at rigorous mathematics. They can produce a plausible-looking argument that is subtly (or catastrophically) wrong, skip steps, or "hallucinate" a lemma that does not exist. To fix this, AI labs pay humans who genuinely understand mathematics to do four broad things, and your day is usually some mix of them:

• Author hard problems. You write novel, high-difficulty math problems (and their fully worked, correct solutions) that stress-test a model. Outlier's math role describes this as "synthesize high-complexity problem sets" that challenge advanced models.
• Write and grade reasoning traces / proofs. You produce step-by-step derivations (a "reasoning trace") that are logically airtight, or you read a model's proof and mark exactly where it goes wrong. This is the "proofs" part, and it is the most valuable because grading a proof correctly requires the same skill as writing one.
• Rank and compare model outputs. Given two or more AI answers to the same problem, you decide which is better and why. This is the human-preference data that sits at the heart of RLHF (reinforcement learning from human feedback): the policy model generates completions, a reward model scores them, and your judgments are what teach the reward model what "good" looks like.
• Build verifiers and RL environments. A newer, higher-paid category: you help build the automatic checks (a verifier, a unit test, a rubric) that let a model be trained with Reinforcement Learning with Verifiable Rewards (RLVR), where a programmatic check confirms correctness instead of a human doing it every time. Mercor builds these "realistic data-rich worlds" with "rigorous tasks and verifiers" and pays experts to design them.

Why is this so well paid compared to image tagging or chatbot rating? Because the supply is tiny. As one research summary put it, at Olympiad level "annotating a single question requires domain experts (like IMO participants) to spend hours and cost hundreds of dollars." The labs cannot cheaply substitute this. A 2026 arXiv paper on math proof reward models exists precisely because proof verification is a bottleneck that money is being thrown at. That bottleneck is your leverage.

The kinds of math in demand

The category is broad. Real postings and platform pages ask for people across:

• Competition and Olympiad math (AMC/AIME/IMO-style problem design and grading).
• Undergraduate and graduate pure math (real and complex analysis, abstract algebra, topology, number theory) where clean proofs matter most.
• Applied math, statistics and probability (the NoDesk PhD Math & Statistics tutor listing via Mercor is typical).
• Theoretical physics and quantitative fields treated as adjacent (many STEM roles bundle physics, chemistry and CS reasoning).
• Formal / machine-verified proofs (Lean, Coq) for the most advanced RL-environment work, which pays at the top of the range.

How much you actually get paid

Here is the honest picture. "Getting paid $30 to $200/hr" is true, but the number you personally hit depends on three things stacked on top of each other: your credentials, your platform, and your country. Domain expertise is, in the words of one platform's own breakdown, "the single biggest lever."

Pay by seniority / credential

This is the table that matters most. It reflects the aggregate ranges reported across Mercor, Handshake AI, and general 2026 market reporting.

The distribution is skewed. Across all AI-trainer work the national average is about $31/hr, the median sits around $25-33/hr, the upper quartile around $53-58/hr, and only the top decile clears $100/hr. STEM and math specifically pull toward the top of that curve because the credential bar is higher, but you still have to qualify into it. Simply being on a platform does not put you at $150/hr; being a math PhD who passed the specialist assessment does.

Pay by platform

The eight platforms you asked about are not interchangeable. They differ in who they accept, how they pay, and what the ceiling is. Rates below are drawn from each platform's own pages plus Glassdoor and third-party reporting; treat them as ranges, not promises.

A few things jump out of that table. Outlier and DataAnnotation are the most accessible and the most beginner-friendly for someone with a math degree, but their ceilings ($40-100/hr) are lower than the credential-gated platforms. Mercor, AfterQuery and Handshake AI are where the $150-250/hr math work lives, but they gate hard on an actual PhD or equivalent. Terac is a "paid feedback" add-on, excellent per-task pay but not a source of steady hours. Sources: Breaking Even's platform pay roundup, RemoWork's 2026 platform list, and each platform's own page.

Pay by country (this is the big asterisk)

The same task pays wildly different amounts depending on where you live, because platforms use region-tiered pay and because the best platforms only accept a handful of countries at all. The best-paying platforms (DataAnnotation and most Outlier expert projects) only accept the US, Canada, UK, Ireland, Australia and New Zealand.

If you are outside the top-tier countries, the realistic move is to target the credential-gated global platforms (Mercor, micro1, AfterQuery, Handshake AI) where a verified PhD can still command a high global rate, rather than the US-only self-serve queues. A math PhD in any country is far more portable than a generalist labeler in any country.

The platforms in detail

DataAnnotation

The most common starting point for people with STEM degrees who are not (yet) PhDs. You take an unpaid starter assessment (about an hour, testing writing, reading comprehension and logical reasoning), then unpaid domain qualification tests (Math, Coding, Chemistry, Physics, etc., each roughly 1-2 hours). Pass the math qualification and you unlock STEM projects that start at $40+/hr and can reach $100/hr for coding-adjacent work. Glassdoor's 1,213-salary sample puts the AI Data Annotator role at roughly $52/hr average, with a 25th-75th band of about $43-63/hr, though that self-reported sample skews high because it excludes unpaid screening time. Payment is via PayPal every 3-7 days. The real friction: qualification tests are unpaid, acceptance is selective, and some applicants report never hearing back.

Outlier

Outlier's math expert page is the clearest job description in the industry. Tasks: "develop and answer Math-related questions," "evaluate and author complex reasoning traces," and "assess and rank AI responses" for scientific rigor. Requirements: a bachelor's or higher in math (current enrollment accepted), LaTeX proficiency, strong English writing. Base is "up to $50/hr" with surge-pricing bonuses averaging +7.5% (top performers +11%). The catch, per Glassdoor's 1,528-review dataset and third-party reviews, is that math/coding sit at the top of Outlier's range while text tasks pay $10-15/hr, task availability "fluctuates heavily," and rates in 2026 are lower than a year prior. It is legit and pays weekly (PayPal/ACH), but do not count on 40 steady hours.

Mercor

Mercor is a "apply once, get matched" platform: you do a ~20-minute AI interview once, and it is reused across opportunities. Its whole model is connecting frontier labs with 30,000+ credentialed experts for post-training, evaluation, rubric design and RL-environment construction. For PhD-level math, reported rates run $110-130/hr, with the platform-wide mode for accepted contributors around $90-120/hr and a ceiling past $200/hr for the rarest specialties. Mercor is aggressively expanding its RL-environment capability (it acquired Sepal AI in early 2026 and announced the Deeptune acquisition in July 2026), which means more, not less, demand for people who can build verifiable math tasks. The honest caveat: pay only accrues when a project is live, and there are "no tasks" waiting periods.

AfterQuery

A Y Combinator-backed research lab that builds expert-generated training data and RL environments. Work arrives in project waves, typically 2-3 week assignments at 10-20 hours/week, followed by quiet queues, with weekly (Friday) payouts. Third-party reports put generalist reviewers at $35-75/hr and credentialed specialists (PhD/MD/JD) at $100-250/hr; live postings have shown things like $150-200/hr for a niche modeling expert. For a math PhD who can commit in bursts, it is one of the highest ceilings available.

Handshake AI

Handshake's AI fellowship program specifically recruits PhDs in Physics, Biology, Chemistry or Math (in progress or completed within the last 10 years) to evaluate AI responses and design hard prompts. Pay tiers run generalist $22-30/hr, Master's/PhD $30-150/hr, and medicine/law/finance specialists up to $175-300+/hr, paid weekly, with rates published upfront and fixed for the project. It is well-suited to current grad students because it grew out of the university-careers platform.

Alignerr, micro1 and Terac

Alignerr (powered by Labelbox) spans roughly $25-150/hr; about 20% of trainers clear $90/hr, but some projects sit at $18/hr and the effective ceiling for non-specialists is lower, so it rewards getting onto a specialist track. micro1 uses an AI screener ("Zara") and pays annotator $20-40/hr, evaluator $20-65/hr, and engineering $50-150/hr, making it a good fit for math people who also code. Terac is different in kind: an AI-native expert panel where you do studies (interviews, feedback sessions, consultations) that can pay $20-300 each, verified by a short voice conversation. It is a strong supplement, not a full-time income.

Requirements, tools and skills

Credentials

There is a real ladder. At the bottom, a bachelor's in math or a hard STEM field (or current enrollment) is enough for Outlier and to attempt DataAnnotation's math qualification. In the middle, a master's widens access and lifts rates. At the top, a PhD in pure or applied math, statistics, or theoretical physics is what unlocks the $100-250/hr work on Mercor, AfterQuery and Handshake AI. Some listings will accept a bachelor's plus 10+ years of professional quantitative experience in lieu of an advanced degree.

Skills that actually get you paid more

• Writing clean, rigorous proofs and being able to grade them. This is the differentiator. Anyone can compute; being able to say precisely why step 4 is invalid is the paid skill.
• LaTeX. Nearly every serious math role requires it for typesetting solutions and traces. If you are rusty, fix that first; it is table stakes.
• Clear technical English writing. You are producing training data that must be unambiguous. Sloppy prose gets your work rejected regardless of whether the math is right.
• Following long, fussy instructions exactly. The most common reason good mathematicians fail these gigs is not the math, it is not reading the multi-page rubric carefully.
• Some coding (Python) and, at the frontier, formal methods (Lean/Coq). Not required everywhere, but they open the highest-paid RL-environment and verifier work.

Tools you will use

The models themselves (many platforms give free access to frontier models, which Outlier values at $200+/month), a LaTeX editor, sometimes an in-house annotation interface, occasionally a formal-proof assistant, and standard freelance plumbing (PayPal, a time tracker such as Insightful on some platforms).

How to get started, step by step

1. Fix your LaTeX and pick a proof or two to warm up. Before you apply, be sure you can typeset a clean solution quickly. This is the single most common gating skill.
2. Apply to two or three platforms at once, matched to your credential. If you have a PhD, lead with Mercor, AfterQuery and Handshake AI. If you have a bachelor's or master's, start with Outlier and DataAnnotation. Diversify, because task availability on any one platform is spiky.
3. Treat the (unpaid) assessment like a real exam. DataAnnotation's own community consistently says the failure mode is rushing. Quality beats speed: read every instruction, do not skim, and do not take the test exhausted. Acceptance rates on the harder qualifications are reported in the 10-20% range, and most of the people who fail failed on carelessness, not competence.
4. Do the AI interview seriously (Mercor, micro1). It is reused across opportunities, so a strong 20 minutes pays off repeatedly. Speak to your specific math background and concrete examples of rigorous work.
5. On your first paid tasks, over-deliver on rigor and formatting. Your early accuracy determines whether you get routed to the higher-paying, higher-trust projects. This is where the real money is: not the base rate, but getting onto the specialist track.
6. Stack platforms and log your effective hourly. Track paid hours vs. total hours (including task-hunting and unpaid screening). Your headline rate and your effective rate can differ by 20-40%.

Realistic earnings

Set expectations honestly. The $150-200/hr headline is real but it applies to credentialed specialists on live, high-trust projects, and it is rarely 40 hours a week. A grounded picture:

• A math bachelor's on Outlier/DataAnnotation, once qualified, realistically nets $25-45/hr on the tasks they get, with availability that ebbs and flows. Many people treat it as a $500-1,500/month side income rather than a salary.
• A math master's or advanced grad student who gets onto specialist tracks can push toward $50-90/hr on good projects.
• A math PhD on Mercor/AfterQuery/Handshake AI can command $100-200/hr while a project is live, but income is bursty (2-3 week waves, then quiet). Someone stitching multiple platforms together can build this into a serious income, but it takes hustle and there are dry spells.

The Glassdoor and ZipRecruiter salary pages that show $50k-130k+ annualized figures are real for consistent, high-tier contributors, but they assume you stay busy. The most common disappointment in reviews is not low rates, it is inconsistent task availability and unpaid screening dragging down the effective hourly.

Payment and taxes

Almost every platform hires you as a 1099 independent contractor (US) or the local equivalent, not an employee. That means:

• No tax is withheld. You are responsible for your own. In the US, budget for ~15.3% self-employment tax plus your income-tax bracket, and set aside 25-30% of what you earn.
• You will get a 1099 if you cross the threshold. Note the 1099-NEC reporting threshold rose to $2,000 for 2026, but you must report all income regardless of whether a form is issued. If you are paid via PayPal you may also see a 1099-K; do not double-count the same income across a 1099-NEC and a 1099-K.
• Payment methods and cadence vary. DataAnnotation pays via PayPal every few days; Outlier weekly via PayPal/AirTM/ACH; Mercor and AfterQuery weekly; Terac via wallet-to-bank, PayPal or gift card. Confirm the method before you invest hours.
• Track your expenses. As a contractor you can typically deduct legitimate business costs (a share of internet, a laptop, software). Keep records.

Pros and cons

Pros: genuinely high pay for the right credential, fully remote and flexible (no minimum hours on most platforms), intellectually real work (you are writing and grading actual mathematics, not tagging cat photos), and a direct on-ramp for grad students and academics tired of adjunct pay and grant writing. Free access to frontier models is a nice perk.

Cons: income is lumpy and unpredictable (waves and dry spells), unpaid assessments and task-hunting eat into your effective hourly, rates in 2026 have softened at the generalist end, best platforms are geo-restricted, and you carry all the self-employment tax and admin burden. There is also the risk of being silently deprioritized or flagged (some platforms auto-flag suspected AI-assisted submissions, which is ironic and occasionally catches honest workers).

Red flags and scams to avoid

The legitimate platforms in this guide never charge you to work. Money flows one direction only: from the company to you. Use these rules:

• Never pay an upfront fee. No "registration fee," "training fee," "background-check fee," "certification," or "starter kit." This is the single most common 2026 scam pattern. Real employers do not charge you.
• Distrust unrealistic promises. "$150/hr for general AI training, no experience required" is bait. Real generalist work is $12-45/hr; the high rates require real credentials.
• Never accept payment by gift card, crypto, wire transfer, or a "deposit this check and forward part of it" arrangement. The last is the classic fake-check scam and ends with you owing the bank.
• Verify the domain and recruiter. Scammers clone real job posts and impersonate recruiters on LinkedIn. Apply through the platform's official site, not a link in an unsolicited DM.
• Be wary of a hire that skips all screening. Real STEM/math work always involves a qualification task or interview. An instant "you're hired, just pay to onboard" is a scam.

The 2026 outlook

Demand for exactly this skill is rising, not falling, and there is a nuance worth understanding. The overall AI data labeling market is about $2.3-2.6 billion in 2026 and forecast to more than double by 2031 (roughly 22-28% CAGR by various estimates). Within that, spending is bifurcating: commodity labeling is being automated and pushed to low-cost regions, while specialized human experts (mathematicians, doctors, lawyers) are commanding higher rates because they cannot be commoditized. Math PhDs "consistently earn the highest rates in AI training" precisely because models still struggle with mathematical reasoning.

The one real headwind for math specifically is RLVR. As labs get better at verifiable rewards, some routine "grade this answer" work gets automated by programmatic checkers. But that shift does not remove humans, it moves them up the value chain: instead of grading, experts are now paid to build the verifiers and RL environments, which is harder and better paid, and it is exactly why Mercor is acquiring RL-environment companies. Net-net for 2026 and into 2027: the floor for generic labeling is under pressure, but the ceiling for genuine mathematical expertise (proof writing, verifier design, research-grade reasoning) is holding up and arguably rising. If you can write and grade a rigorous proof, you are on the right side of this trend.

Bottom line

STEM, math and proofs is the best-paid, most durable corner of the AI training economy. If you have a math or hard-STEM degree, start with Outlier and DataAnnotation, take the assessments seriously, and build up your accuracy to reach specialist tracks. If you have a PhD, go straight for Mercor, AfterQuery and Handshake AI, where $100-200/hr is real. Diversify across two or three platforms, track your effective hourly honestly, set aside a third for taxes, and never pay anyone a fee to work. Done right, it is one of the few online gigs where deep expertise is actually rewarded at expert rates.