Methodology
How we compile pay
AITraining.jobs publishes what AI-training work pays across 65 platforms and 12 kinds of work. Those numbers are only worth trusting if you can see where each one comes from. This is exactly how we source, label, and keep them current.
Every figure is labeled
The confidence label on every pay number
This industry runs on inflated averages. We attach a confidence label to every figure so you can tell a live wage from a hopeful estimate at a glance. These are the four tiers, strongest first.
Live from job feed
Read straight from a platform's own open-role feed and dated to when we last pulled it. This is the strongest figure we can show: a rate a platform is advertising right now.
Platform-stated
Published by the platform itself on its own pages. Real, but often an average weighted toward the top tier, so we treat it as a claim rather than an observed wage.
Third-party reports
Drawn from worker reports and aggregators (Glassdoor, Reddit, Trustpilot, primary news reporting). Directional and useful for sentiment, but not exact.
Estimate
A reasoned figure where no live or stated source exists. Always shown as a range, clearly labeled, and never presented as a precise or guaranteed wage.
Live pay
How we compute the live medians
Where enough open roles state an hourly rate, we show today's real market rate instead of a curated estimate. This is the site's edge, so the calculation is deliberately conservative.
- We read hourly rates straight from each platform's own job feed, pulled fresh every day, and take the median and the typical pay across the open roles that state a rate. Typical pay is the 25th percentile of what roles pay (the middle of each advertised range) to the 90th-percentile advertised ceiling, so it includes the well-paid expert, coding and agentic tier, and it is measured the same way for every list of roles: the whole market, a platform, a kind of work, a platform’s roles of one kind, or a search. Every pay figure opens the roles it is measured from, and that list states the same figure.
- We count distinct advertised rates after removing duplicate relistings of the same job, so one role posted many times cannot skew the median. A rate count is never reported as an open-role count: they measure different things.
- We exclude advertised ceilings ("up to $X/hr"), which are marketing maximums, not observed wages. A platform whose feed states only ceilings shows no live median, and we say so rather than dress a ceiling up as a going rate.
- We only call a figure a market rate when it spans several platforms with a large enough sample. A rate from a single platform is shown as that platform's rate card, clearly labeled, and never published as the going rate for the whole category.
Coverage
What the live feed can and cannot see
The live medians are computed only from platforms that publish an open job board or worker API. That boundary shapes what the numbers mean, so we state it plainly.
Of the 65 platforms we track, 29 are hiring now, their open roles read from a live feed we pull every day. That set is not just the entry tier: it already includes premium expert networks like Mercor, AfterQuery, Outlier (Scale AI), and Handshake AI, so the live band reaches genuine top-tier hourly work: the highest rate advertised on it today is $450/hr, and every open role is listed by pay.
Two things sit outside it, and we would rather name them than imply the feed sees everything. First, networks that onboard invite-only and publish no public board (Surge AI is the clearest example: one of the largest expert-data firms, listed in our directory, but its contributor work is matched privately and its public API is empty). Second, the rarest and highest-ticket work: the roughly $500 to $1,000/hr elite expert engagements (medical fellows, management consultants, VC partners), and the $20,000 to $300,000 per-environment reinforcement-learning builds, which are commissioned privately rather than posted by the hour.
So the two numbers on each page mean different things, on purpose. The live median is the honest going rate for open, hourly, publicly-posted work. The top of each category's range comes from the curated benchmark, labeled as such, precisely because that frontier tier is matched off-board and never reaches a feed we could measure.
We also draw a line at who does the paying. We list places that pay an individual to do the work (by the hour, the task, or the accepted bounty), not companies hiring salaried staff to build training data in-house (Mechanize, for one, employs engineers on six-figure salaries to build RL environments) or business-to-business evaluation tools sold to the labs themselves (HUD, for one, is an agent-evaluation SDK). Both are real parts of the AI-training economy, but neither is a way an individual gets paid to train AI, so neither belongs in this directory.
Curated benchmarks
When there is no live rate
Not every kind of work has enough openly-posted rates to measure. For those, we fall back to a curated benchmark, compiled by hand from public sources and labeled as such.
Curated benchmarks come from each platform's own pages, live job feeds, and worker reports (Trustpilot, Glassdoor, Reddit, and primary news reporting). Each one carries the same confidence label as any other figure, so it is never mistaken for a live wage.
The two dates on the site mean different things, and we keep them separate on purpose. Live data reflects open roles as of 18 Oct 2026 (it refreshes every morning). Our curated benchmarks were last hand-verified 25 Jul 2026. We would rather show a range and its source than a confident figure we invented.
How we stay free
On referral links
Some links on AITraining.jobs are referral links. If you sign up through one, AITraining.jobs may earn a commission at no extra cost to you. It never changes what a platform pays you.
Here is exactly how that referral relationship does, and does not, affect what you see. It never touches the facts: a pay figure and its confidence label are compiled the same way whether or not a platform pays us, and the two sorts you can choose, highest pay and newest first, rank purely by the number and the date, with no referral input at all.
The Top Networks list on the home page is a featured list: it shows the networks we have a referral relationship with, in order of that relationship, and a line right under it says that we may earn a referral fee when you join through those links. The home page's platform filter and its “Platforms hiring right now” list also put those same networks first, then every other platform by how many roles it has open, and the referral disclosure on that page covers them.
Everywhere else we rank a SET of platforms (the full platform and role directories, and your suggested platforms when signed in), the order is neutral: partner status is only one signal in a quality score, alongside live hiring activity, pay, trust, and how many kinds of work a platform hires for. Its weight is deliberately bounded there, so a weak platform never ranks well just for paying us. Those directories are meant to be trusted, and a ranking you cannot trust is worthless to you and to us.
The default balanced roles feed leads with our referral partners: the newest couple of roles from each partner network come first (in order of our referral relationship, the same basis as the Top Networks list, not by what the role pays you), and after that the feed continues one role per platform in turn, newest first, with a partner's role still coming before the rest within each turn. So partner roles are more prominent near the top, but the feed is not all partners: non-partner roles fill the rest and appear throughout. Where a partner role genuinely pays above the market, we mark it with that fact ("Top-tier pay", "Above-average pay"), measured from the same live rates the rest of the site uses, so the emphasis rests on a real feature of the role, not a label. The moment you switch the feed to Highest pay or Newest, the partner ordering drops away entirely and roles rank purely by the number or the date.
When you describe the job you want, or match your LinkedIn profile, an AI model reads what you are looking for and every open role is ranked by how closely its title matches it in meaning: the roles that match come first, then every other role, closest first. In that ranking a partner's role gets a small lift that can put it ahead of an equally close match, never ahead of a clearly closer one, and it never decides which roles count as a match. With Newest or Top pay, the matching roles rank purely by the date or the number.
“Similar roles hiring now” uses that same balanced order, so its first roles are usually our partners'. It appears under the roles of a platform we have no referral relationship with, and as a small card after you apply to one. The card lists up to three open roles of the same kind of work on other platforms, open where you are, at most once every three days, and never in a visit where you already applied through a referral link.
Clicking a role opens a short summary of it: its pay, where it is open, the start of its description and what the platform is. With a free profile, an AI model also checks the role against your background and tells you whether you are a strong, possible or unlikely match, and why. For a role at one of our referral partners the summary carries its apply link straight away. For a role at a platform we have no referral relationship with, the apply link is part of your free profile: create one (your LinkedIn link or a sentence about your work, an email and a password) and the link appears. Beside such a role the summary lists up to three comparable roles at our referral partners that you can apply to right away: the same kind of work (or, where our partners have too few of those, other work that platform also hires for), open where you are, with their pay and how each platform pays.
Either way, this referral disclosure appears on every surface where partner status affects the order, and clicking any link takes you to that platform under its own terms.
Written by Yuma Heymans, Founder & CEO of AITraining.jobs.
Founder and CEO of AITraining.jobs and co-founder of the AI recruitment platform HeroHunt.ai, based in San Francisco, with years spent mapping how AI-training and data work is sourced, staffed, and paid.