Math & STEM

The highest-paying corner of the AI-training economy, where a clean proof is worth more than almost any other skill you can sell online.

If you can write a rigorous proof and, harder still, spot exactly where someone else's proof breaks, you own a skill that the largest AI labs in the world are currently paying $30 to $250 an hour to rent. That is not a headline plucked from a get-rich-quick pitch. A single Lean 4 formal-proof role posted by Alignerr in 2026 openly advertises $170 to $200 per hour for mathematicians who can formalize advanced proofs - Alignerr's Lean 4 Proof Engineer listing. Meanwhile the national average across all AI-training work sits near $31 an hour - Mercor's salary breakdown. The gap between those two numbers is the whole story of this guide.

The reason for that gap is structural, not temporary. Large language models have gotten frighteningly good at mathematics - a frontier system reached gold-medal standard at the 2025 International Mathematical Olympiad - Epoch AI's FrontierMath work documents just how fast the ceiling is rising. For a closer look at which AI math benchmarks are actually solved and which are still unproven, see this explainer on O-mega.ai, a site run by the same team as AITraining.jobs. And yet the labs are spending more, not less, on human mathematicians. The benchmark FrontierMath was built with over 60 mathematicians, including IMO question writers and Fields medalists, because the models still cannot be trusted to generate their own hardest problems and grade their own proofs. That contradiction, models that are brilliant and still unreliable, is exactly why this work exists and why it pays what it pays. This guide explains what the work actually is, who pays what, how to get your first paid task, and how to avoid the growing pile of scams built to prey on people chasing these numbers.

Contents

1. What STEM, math and proofs AI-training work actually is
2. Why this human work exists at all (first principles)
3. What you do day to day: concrete task walk-throughs
4. Who it is for, and the honest barrier to entry
5. The skills that get you paid more, and how to build them
6. The platforms that hire for this work, compared
7. Realistic pay: ranges and what moves your number
8. How to get started and land your first paid task
9. Scams and pitfalls to avoid
10. How AI agents and 2026 model progress are reshaping this work
11. A decision framework: is this right for you?

1. What STEM, math and proofs AI-training work actually is

STEM, math and proofs is the category of AI-training work where labs pay people who genuinely understand mathematics to produce, evaluate, and structure the data that teaches models to reason. It sits at the top of the pay ladder for a simple reason: the pool of people who can write a correct proof of a hard theorem is tiny, and the pool who can reliably grade one is smaller still. Where general data labeling asks "is this a stop sign," this work asks "is step four of this derivation valid, and if not, precisely why." That is a judgment task, not a perception task, and judgment about rigorous mathematics is one of the scarcest inputs the AI industry buys. The result is a market where a math PhD in almost any country can command a premium global rate that a generalist labeler never will.

The work is not one job but a family of related tasks, and most people who do it end up doing a mix. You might author novel problems that stress-test a model, write reasoning traces (step-by-step derivations that must be logically airtight), grade and rank the model's own attempts to find where it went wrong, or, at the most technical and best-paid end, build the verifiers and reinforcement-learning environments that let a model be trained with automatic correctness checks. The exact blend depends on the platform and the project, and it shifts over time as labs move from one training bottleneck to the next. What stays constant is the underlying demand: a model is only as good as the mathematical data it learns from, and that data has to come from humans who can tell right from almost-right. If you want to see how this category sits alongside its neighbors, you can compare pay across every kind of AI-training work or read the related guide on RLHF and model evaluation, which is the broader discipline this work feeds into.

2. Why this human work exists at all (first principles)

Start from the machine, not the marketing. A large language model is a next-token predictor trained on a vast corpus, and prediction is not the same thing as proof. A model can produce an argument that looks like a proof, complete with the right vocabulary and the right shape, while quietly asserting a lemma that does not hold or skipping the one step that actually needed justification. This failure mode is not a bug that a bigger model erases: it is baked into what "generate the most plausible continuation" means. A plausible-looking wrong proof is, by construction, the thing the model is most likely to produce when it is out of its depth. To make a model reliable at mathematics, someone has to supply the signal that separates "plausible" from "correct," and that signal is not in the raw internet. It has to be manufactured by people who can tell the difference.

There are only two ways to manufacture that signal, and both need humans who understand the math. The first is human judgment as the reward: a mathematician reads the model's output and scores it, and those scores train a reward model that then guides the base model, the loop known as reinforcement learning from human feedback. The second is a machine-checkable ground truth: you write a problem whose answer can be verified programmatically, or you formalize the proof in a system like Lean that mechanically checks every step, so the reward comes from a deterministic verifier instead of a person - this RLVR explainer lays out the mechanism. Both paths run straight through human expertise. The first needs a human to be the judge. The second needs a human to build the verifiable problem or formal proof in the first place, because a model cannot reliably generate its own ground truth without contaminating it. The economic punchline follows directly: the scarcer and more rigorous the mathematical judgment required, the higher the price, because the labs cannot substitute cheap labor or, yet, the model itself. That is the entire basis for the rates in this guide.

• Models are fluent, not sound. Fluency is cheap and abundant; soundness is scarce and expensive.
• The training signal is manufactured, not found. Correct hard-math data barely exists on the open web at the frontier, so it has to be created.
• Contamination forces human authorship. A model cannot grade its own novel proof without the risk of grading its own mistake as correct.
• Verification is a bottleneck money is thrown at. Grading an Olympiad-level answer can take a domain expert hours, so labs pay to shorten that bottleneck.

Read those four points together and the market makes sense. This is not busywork dressed up as a career: it is the labs buying the one input they cannot synthesize, mathematical correctness certified by a mind that understands it. The moment a task becomes machine-checkable and routine, its price falls and it gets automated. The moment a task requires genuine mathematical creativity or the ability to certify a hard proof, its price holds or rises. Everything else in this guide, the platforms, the rates, the outlook, is downstream of that single dynamic.

3. What you do day to day: concrete task walk-throughs

The abstract description ("author, evaluate, rank, verify") becomes real when you see what a shift actually looks like. Most projects hand you a queue or a batch of items, a multi-page rubric, and a deadline, and you work through items one at a time inside a browser-based annotation tool or a shared document. You are almost always working in LaTeX, because your solutions and traces have to be typeset cleanly enough to become training data, and you frequently have free access to a frontier model whose outputs you are grading (a perk that Outlier alone values at over $200 a month). The rhythm is closer to careful grading than to fast clicking: a single hard item can take you thirty minutes to an hour to do properly, and doing it properly is the entire point, because your accuracy on early tasks determines whether you get routed to the higher-trust, higher-paying work later.

To make it concrete, here are four representative tasks you might see in a single week, each drawn from how the platforms describe the work. Notice that they escalate in both difficulty and pay, and that the highest-paid one barely resembles "labeling" at all. This escalation is the ladder you climb, and understanding it is how you avoid getting stuck at the bottom rung earning generalist rates when your credential could earn double.

• Author a hard problem with a worked solution. You invent a novel Olympiad-adjacent problem, then write the full, correct, step-by-step solution. Outlier's math role describes this as synthesizing "high-complexity problem sets" that challenge advanced models.
• Grade a reasoning trace. You read the model's step-by-step derivation and mark exactly where it goes wrong, or confirm it is airtight. This is the highest-signal task because catching a subtle error requires the same skill as proving the result.
• Rank two model answers. Given two solutions to the same problem, you decide which is better and write why. These preference judgments are the raw material of reward-model training.
• Formalize a proof in Lean. You translate an informal proof into machine-checkable Lean 4 code, so a deterministic checker can verify every step. This is the top of the pay range and the frontier of the work.

The interpreting point to hold onto is that these four tasks are not equally valued, and the difference is not about how long they take but about how few people can do them. Ranking two answers is something a strong undergraduate can learn to do consistently; formalizing a nontrivial proof in Lean is something a small number of specialists on earth do well. Your daily reality on any given project is some blend of these, weighted by your credential and by which bottleneck the lab is currently paying to relieve. The people who earn the most are not the fastest clickers: they are the ones who reliably do the grading and authoring tasks that the model itself still fails, because that is precisely the data the lab cannot get any other way. If the evaluation side interests you, the RLHF and evaluation pay hub covers the adjacent judgment-focused roles in depth.

4. Who it is for, and the honest barrier to entry

This work is for people with real mathematical training, and it is honest to say so plainly, because the barrier to entry is the single most misunderstood thing about it. The high rates you see advertised are gated behind credentials and, more importantly, behind demonstrated ability. A bachelor's degree in mathematics or a hard STEM field (or current enrollment in one) is the realistic floor for entering the accessible platforms. A master's widens your access and lifts your rate. A PhD in pure or applied math, statistics, or theoretical physics is what unlocks the six-figure-hourly-equivalent tier on the credential-gated platforms, and some of them are explicit about wanting a recent PhD or advanced candidacy from a strong program. The work rewards depth, and it filters hard for it, which is exactly why it pays what it does.

That said, the barrier is more nuanced than "no PhD, no entry." The credential opens the door, but what keeps you inside and moving up the pay ladder is your accuracy and your discipline with instructions, not your diploma. The most common way strong mathematicians fail these gigs is not a lack of math ability: it is rushing through a multi-page rubric and missing a rule, which tanks their quality score and gets them deprioritized. So the honest picture has two layers. First, a real quantitative background is close to non-negotiable, and if you are trying to enter with no math credential and no demonstrable ability, this is the wrong category for you (data labeling or research studies would be more realistic entry points). Second, among people who do have the background, the ones who succeed are the careful ones, and the ones who wash out are usually the careless ones. If you have a degree and you are meticulous, you are a strong candidate. If you have a PhD and you can formalize proofs, you are close to the top of the market.

There is a geographic asterisk on all of this that matters enormously. The best self-serve platforms, DataAnnotation and most of Outlier's expert projects, only accept a handful of countries (the US, Canada, UK, Ireland, Australia and New Zealand). If you live elsewhere, the door is not closed, but it is narrower: your realistic path is the credential-gated global platforms, where a verified math PhD can still command a high international rate. A math PhD is one of the most portable credentials on this market. A generalist labeler in a low-cost region competes on price; a verified specialist competes on scarcity, and scarcity travels.

5. The skills that get you paid more, and how to build them

The credential gets you in; the skills get you paid at the top of your tier. There is a clear hierarchy of what the labs actually value, and it is worth being precise because the market rewards some abilities far out of proportion to how "advanced" they feel. The single most valuable skill is not solving hard problems: it is grading them correctly, which means reading a proof and stating exactly why a given step is invalid. Anyone with a degree can compute; being able to say precisely why line four does not follow is the paid skill, because it is the same skill the model lacks and the same skill the reward signal needs. If you can do only one thing supremely well, make it this.

Around that core sit a set of enabling skills, and the gap between a mid-tier and a top-tier contributor is usually one of these rather than raw brilliance. Below are the ones that most reliably lift your rate, roughly in order of how often they are the deciding factor. Treat this as a checklist for where to invest before you apply, because arriving with these already sharp is what separates a fast acceptance from a 30-day retry wait.

• LaTeX fluency. Nearly every serious math role requires it. If you are rusty, fix that first: it is table stakes, not a differentiator, and fumbling it wastes your paid hours.
• Precise error diagnosis. The ability to localize the exact flaw in a wrong proof. Practice by grading solutions from competition archives and forcing yourself to name the first invalid step.
• Following long rubrics exactly. Read the multi-page instructions twice before you start. The most common failure is a competent mathematician who skimmed a rule.
• Clear technical English. You are producing training data that must be unambiguous. Sloppy prose gets correct math rejected.
• Python and formal methods (Lean, Coq). Not required everywhere, but they open the highest-paid verifier and RL-environment work, where the ceiling lives.

The interpreting point is that this list is deliberately weighted away from what mathematicians usually think of as their strength. Your ability to prove a deep theorem is assumed; it got you the credential. What separates pay tiers is the unglamorous, professional stuff: typesetting cleanly, reading instructions like a lawyer, writing unambiguously, and, if you want the very top, learning the tooling of formal verification. The good news is that all of these are learnable in weeks, not years, if the mathematical foundation is already there. The Lean skill in particular is a high-leverage investment right now: the demand for autoformalization is expanding fast, with active 2026 research on translating natural-language proofs into machine-checkable form - see this ProofBridge paper on auto-formalization in Lean - and the roles that need it pay near the top of the entire market. If you also write code, the coding and software evaluation guide and its pay and jobs hub cover the adjacent, similarly well-paid work of evaluating and building software.

6. The platforms that hire for this work, compared

The platforms that pay for STEM, math and proofs work are not interchangeable, and choosing the right ones for your credential is the highest-leverage decision you will make. They differ on three axes that matter: who they accept (open self-serve versus PhD-gated), how work reaches you (a standing task queue versus project waves versus matched assignments), and where the ceiling sits (a $50/hr cap versus a $250/hr specialist track). Getting this wrong is the most common way people underearn: a math PhD grinding a $40/hr self-serve queue is leaving well over a hundred dollars an hour on the table by not being on a credential-gated platform, while a bachelor's-only applicant who applies only to the PhD-gated platforms may never get matched at all.

Here is the landscape, with rates drawn from each platform's own pages plus third-party 2026 reporting. Treat every figure as a range that depends on your credential and the specific project, not a promise of a steady wage.

A few things jump out of that table, and they should shape your application strategy directly. Outlier and DataAnnotation are the accessible on-ramps: a math bachelor's can get in, but their ceilings (roughly $40 to $100 an hour) are lower than the gated platforms because the credential bar is lower. Glassdoor's large self-reported sample puts DataAnnotation's annotator role near $52/hr on average, though that number skews high because it excludes unpaid screening time - Glassdoor's DataAnnotation data. Mercor, AfterQuery and Handshake AI are where the $150-plus work lives, gated hard on a real PhD or equivalent, with Mercor's "apply once, get matched" model reusing a single 20-minute AI interview across many opportunities. Alignerr is the outlier for formal proofs specifically, openly posting $170-200/hr for Lean 4 work, which is the clearest signal in the market that formal-verification skill is the premium play right now.

The strategic reading is to stack two or three platforms matched to your credential rather than betting on one. Task availability on any single platform is spiky (waves and dry spells are the norm, not the exception), so diversification is not greed, it is stability. If you hold a PhD, lead with Mercor, AfterQuery and Handshake AI, and add Alignerr if you can formalize proofs. If you hold a bachelor's or master's, start with Outlier and DataAnnotation and use strong early accuracy to climb onto specialist tracks. One resource worth knowing about is AITraining.jobs, which aggregates live STEM and math openings and posted pay across these platforms in one place, so you can see what is actually hiring this week without checking six sites by hand. You can also browse the live STEM and math roles directly to gauge current demand before you commit to an application.

7. Realistic pay: ranges and what moves your number

The honest headline is that $30 to $250 per hour is the true range, and the number you personally hit is the product of three multipliers stacked on top of each other: your credential, your platform, and your country. Domain expertise is the single biggest lever, and the distribution is heavily skewed. Across all AI-training work the national average is about $31/hr and the median sits near $25 to $33, with only the top decile clearing $100 - Mercor's rate breakdown. STEM and math pull toward the top of that curve because the credential bar is higher, but being on a platform does not put you at $150: being a math PhD who passed the specialist assessment does. As of April 2026, ZipRecruiter pegged the average AI mathematics trainer at $24.74/hr - ZipRecruiter's math trainer data - which tells you the average task is a generalist task; the premium is what you build toward, not what you start at.

Here is the pay ladder by credential, which is the table that matters most because it is the one you can actually move yourself up. Read it as a map of where you are and where the next rung is, not as a promise that arriving at a tier guarantees the top of its range.

The country multiplier is the biggest asterisk and the one most guides gloss over. The same task pays wildly different amounts by region, and the best-paying self-serve platforms restrict who can even join. If you are in a top-tier country you have full access and can chase the credential-gated ceilings. If you are elsewhere, your leverage is your credential, not your location: a verified PhD travels far better than a generalist role does. The table below sketches the rough picture, and the interpreting takeaway is simple: lead with your credential, never with your local rate, because the whole point of specializing is that scarcity is global while cheap labeling is local.

Two forces will move your number up over time, and both are within your control. The first is accuracy compounding into trust: platforms route their best-paying, highest-trust projects to contributors with proven quality scores, so over-delivering on your first tasks is the highest-return thing you can do. The second is skill escalation: moving from grading into authoring, and from authoring into verifier and formal-proof work, walks you up the tiers in the table above. The force that moves your number down is the one to watch: your effective hourly (paid hours divided by total hours, including unpaid screening and task-hunting) can run 20 to 40 percent below your headline rate. Track it honestly, because the most common disappointment in platform reviews is not low rates, it is inconsistent availability dragging the effective number down.

8. How to get started and land your first paid task

Getting started is a sequence, not a leap, and doing it in the right order is what separates a smooth entry from months of frustration. The mistake most people make is applying broadly and carelessly, treating the assessments as formalities. They are not formalities: they are the entire filter, and on the harder qualifications the acceptance rate runs in the 10 to 20 percent range, with most failures coming from carelessness rather than lack of math skill. Fail a DataAnnotation qualification and you may wait 30 days to retry - see this DataAnnotation review - so treating the first attempt like a real exam is not optional. The steps below are ordered to maximize your odds on that first, high-stakes attempt.

Work through these in sequence. The first three happen before you ever touch a paid task, and they are where the outcome is actually decided.

1. Sharpen your LaTeX and warm up on a proof. Before applying, confirm you can typeset a clean solution quickly. This is the single most common gating skill and the easiest to fix in advance.
2. Pick two or three platforms matched to your credential. PhDs lead with Mercor, AfterQuery and Handshake AI; bachelor's and master's holders start with Outlier and DataAnnotation. Diversify, because availability is spiky.
3. Treat the unpaid assessment like a final exam. Read every instruction twice, do not skim the rubric, and never take the test tired. Quality beats speed, and the labs test whether you can follow scoring criteria consistently, not whether you have opinions - Operations Army's assessment guide.
4. Do the AI interview seriously. On Mercor and similar platforms, the ~20-minute interview is reused across many opportunities, so a strong showing pays off repeatedly. Speak to concrete examples of rigorous work.
5. Over-deliver on your first paid tasks. Early accuracy determines whether you get routed to the higher-paying specialist tracks. This is where the real money is decided: not the base rate, but earning the trust that unlocks the premium queue.

The interpreting point behind this sequence is that the leverage is front-loaded. Almost everything that determines your eventual earnings, which platforms you are on, whether you pass the gate, and whether you climb onto the high-trust track, is decided in your first two weeks. After that the system largely runs on the reputation you built early. So resist the urge to rush the assessments to "start earning sooner." An extra hour of care on a qualification test, or on your first ten paid items, compounds into months of better-paying work. And keep a simple log from day one: paid hours versus total hours, and effective rate per platform. That log is how you discover, honestly, which platform is actually worth your time, which almost never matches the headline rate. If you want a running view of what is hiring while you build up, AITraining.jobs tracks live STEM and math openings and pay across the platforms in one feed.

9. Scams and pitfalls to avoid

The high pay in this category has attracted a matching wave of scams, and the scale is not trivial: reports to the Federal Trade Commission about job scams tripled between 2020 and 2024, with reported losses jumping from $90 million to over $513 million in that window - Digital Biz Talk's 2026 scam breakdown. The good news is that the fraud follows a small number of predictable patterns, and one bright-line rule catches almost all of it: a legitimate platform never charges you to work. Money flows one direction only, from the company to you. Every platform named in this guide follows that rule. The moment anyone reverses the flow, asking you to pay a fee, buy a kit, or deposit and forward a check, you are talking to a scammer, not an employer.

Here are the patterns worth memorizing, because they cover the overwhelming majority of what you will encounter. Read them once, and any offer that trips one of them should end the conversation immediately.

• Any upfront fee. No "registration," "training," "background-check," "certification," or "starter kit" fee. This is the single most common 2026 scam pattern, and real employers never charge it.
• Payment by gift card, crypto, wire, or a check you forward. The last is the classic fake-check scam: the check bounces, and you owe the bank the money you already forwarded.
• Unrealistic promises with no screening. "$150/hr, no experience required, start today" is bait. Real high rates require real credentials and always involve a qualification task.
• Unsolicited recruiter DMs. Scammers clone real job posts and impersonate recruiters on LinkedIn, WhatsApp and Telegram. Apply only through a platform's official site.
• ID or banking details before any work. A request for a banking password or a photo of your ID before you have done anything is a data-harvesting scam.

Beyond outright fraud, there are honest pitfalls that cost real people real money. The biggest is misjudging your effective hourly: the headline rate ignores unpaid assessments and task-hunting, and a $50/hr platform with three hours of dry-queue hunting per paid hour is not a $50/hr platform. A second is income lumpiness: work arrives in waves with quiet stretches between, so treating any single platform as a steady salary sets you up for disappointment. A third, more subtle, is being silently deprioritized for a rushed early task or, ironically, auto-flagged as AI-assisted when your writing is simply clean and formulaic (some platforms run detectors that occasionally catch honest workers). The defense against all three is the same discipline: diversify across platforms, log your real effective rate, and protect your early accuracy like the asset it is. If you want to sanity-check any platform against the broader market, comparing pay across all AI-training categories is a fast way to spot an offer that is too good to be true.

10. How AI agents and 2026 model progress are reshaping this work

This is the question that decides whether the work is worth building toward, and the answer is more encouraging than the doomer framing suggests, but it demands that you position yourself correctly. Start with the market signal: the AI data-labeling market is estimated at about $2.32 billion in 2026 and forecast to reach $6.53 billion by 2031, a roughly 23 percent compound annual growth rate - Mordor Intelligence's forecast. The industry is spending more on human-supplied training data, not less. But that aggregate hides the crucial split: within it, spending is bifurcating. Commodity labeling is being automated and pushed to the lowest-cost regions, while specialized human experts, mathematicians most of all, command higher rates because their judgment cannot be commoditized. You want to be on the specialist side of that split, and this category is squarely on it.

Now the nuance specific to math. Frontier models are genuinely closing the gap on mathematical reasoning: a leading system reached gold-medal standard at the 2025 IMO, and open research systems have posted near-perfect scores on competitions like the Putnam - Epoch AI's benchmark work. If you stop reading there, it looks like the humans are being trained out of a job. But the actual dynamic is the opposite. As models improve, the only data that still moves them is data at or beyond the current frontier, which is precisely the hardest, most expert-dependent data to produce. FrontierMath had to recruit 60-plus mathematicians including Fields medalists to build problems the models could not already solve, and it did so under strict anti-contamination protocols. The better the models get, the scarcer and more valuable the humans who can still generate problems the model fails become. Progress does not remove the human: it pushes the human further up the difficulty curve, where fewer people can operate and the pay is highest.

The one real structural shift is the rise of reinforcement learning with verifiable rewards (RLVR), where a deterministic checker, an exact-answer comparison, a unit test, a Lean proof-check, replaces a human grader for routine correctness. This does automate some "grade this answer" work. But it does not eliminate humans; it moves them up the value chain. Instead of grading answers one by one, experts are now paid, and paid more, to build the verifiers and RL environments themselves, which is harder, scarcer work. This is why the market for formal-proof and verifier skill is the premium play right now, why Lean 4 roles openly post $170-200/hr, and why platforms like Mercor have been acquiring RL-environment companies to meet demand. Net-net for 2026 and beyond: the floor for commodity labeling is under pressure, but the ceiling for genuine mathematical expertise, proof writing, verifier design, and research-grade reasoning, is holding and arguably rising. If you can write and grade a rigorous proof, or build the machine that checks one, you are on the right side of this trend. To see how the same forces play out in adjacent expert categories, the law, medicine and finance guide and its pay hub cover the other credential-gated specialties, while red-teaming and AI safety covers the judgment work of stress-testing models.

11. A decision framework: is this right for you?

Reduce everything above to a single decision. STEM, math and proofs AI-training work is worth your time if, and only if, three things are true: you have real mathematical ability backed by a credential (bachelor's minimum, PhD for the top tier), you are meticulous with instructions and willing to protect your early accuracy, and you can tolerate lumpy income by stacking platforms rather than expecting one steady paycheck. If all three are true, this is one of the very few online markets where deep expertise is rewarded at genuine expert rates, and the trend is running in your favor, not against it. If any one is false, adjust: no meticulousness means you will wash out regardless of talent; no credential means you should start with a more accessible category and build toward this one; no tolerance for lumpiness means you should treat it as a supplement to other income, not a replacement.

Here is how to act on that judgment, matched to where you actually stand today. Pick the row that describes you and follow it, then diversify from there.

• You have a PhD in math/physics/stats. Lead with Mercor, AfterQuery and Handshake AI; add Alignerr if you can formalize proofs in Lean. Aim for the $100-250/hr specialist tracks and treat the AI interview as a repeatable asset.
• You have a bachelor's or master's in a hard STEM field. Start with Outlier and DataAnnotation, pass the math qualification with exam-level care, and use early accuracy to climb onto specialist tracks toward $50-90/hr.
• You have quantitative skill but no formal credential. This category will filter you out. Start with data labeling or research studies, build a track record, and study toward the credential this work requires.
• You are outside the top-tier countries. Skip the US-only self-serve queues and target the credential-gated global platforms, where a verified PhD earns a high international rate. Lead with your credential, never your local market rate.

The through-line of every row is the same first-principles truth this guide opened with: you are being paid for the one input the labs cannot manufacture without you, mathematical correctness certified by a mind that understands it. Build the skill that certifies hard proofs, protect the reputation that routes you to the best work, and position yourself on the specialist side of a market that is spending more, not less, on exactly what you can do. To keep a live pulse on what is hiring and what it pays, AITraining.jobs aggregates current STEM and math roles and posted rates across these platforms, and you can always browse the live openings to see where the demand is this week. This guide's author, Yuma Heymans (@yumahey), founder and CEO of AITraining.jobs and co-founder of the AI recruitment platform HeroHunt.ai, spends his days watching how labs source exactly this kind of scarce mathematical talent from San Francisco, and the pattern is consistent: the harder the proof, the higher the price.

This guide reflects platform pay ranges, market data, and hiring practices as of July 2026. AI-training rates, platform requirements, country eligibility, and tax thresholds change frequently, so verify current details on each platform's official site before applying, and treat all figures as ranges rather than guarantees.