State of AI Training Pay: Q3 2026

The median advertised hourly rate across the AI-training listings we tracked on October 1, 2026 was $52. The middle half of distinct rates ran from $23 to $95 an hour. Those figures rest on 740 distinct advertised rates from 26 platforms, drawn from 7,043 open listings.

A listing is an offer to apply, not a timesheet. The number says nothing about how many hours a worker gets, whether the work is there every week, or what they are eventually paid. Every figure in this report is what postings said, read on one date and frozen there, so it will say the same thing next month that it says today. The live version is the Pay Index, which is recomputed every night.

What the work pays

“AI training” covers a legal specialist reviewing model answers, a model-evaluation project and a short annotation task. They are different jobs with different advertised rates, and one platform often repeats the same pay sentence across hundreds of its listings. We count that sentence once, so a repeated advertisement cannot stand in for hundreds of independent price decisions.

The spread by field is wide. Law listings had a median of $120/hr across 40 distinct rates, finance $100/hr across 62 and healthcare $89/hr across 82. Linguistics sat at $36/hr and voice and speech work at $30/hr. Task type cuts the same listings another way: research and analysis had a median of $76/hr, AI model evaluation $65/hr across 273 rates (the largest sample of any task), and data annotation $30/hr. One listing can carry both a specialist field and an evaluation task, so these are separate views of the same listings, not categories to add up.

Platforms differ just as much. AfterQuery, Mercor and Handshake AI advertised medians of $83 to $93/hr; Prolific and Meridial $30; Welo Data $14 across 100 distinct rates. A platform’s median describes what it advertised in this sample. It is not a promise to every applicant, and it does not compare workers with the same qualifications.

Four snapshots, and which ones compare

On the first of every month, an archive job saves what the Pay Index published that day. There are four captures so far:

Read the last column before the second. July and August took the median over every listing. On August 1 the index switched to counting each distinct advertised rate once, and the first archive taken under the new rule is September’s. That is a different statistic, not just a different sample size, so a move from $41 in August to $50 in September does not show anyone being paid more. Even the first row is not what its archive key implies: it was captured on July 29, not July 1.

The only like-for-like comparison is September 1 to October 1: $50 to $52.

The September rise came from coverage, not raises

A market median can rise because platforms pay more, or because different platforms are in the sample. Breaking the two snapshots out by platform separates the two:

Of the 16 platforms with at least five distinct rates on both dates, 7 advertised a lower median on October 1, 5 were unchanged and 4 were higher. Prolific’s median fell from $40 to $30, Mercor’s from $90 to $85 and Welo Data’s from $17 to $14. The market median rose anyway, and the clearest reason is micro1: its listings were missing from our September sample (its job board blocks the servers our nightly collection runs from) and back by October 1, adding 106 distinct rates with a median of $80, well above the market’s.

The same effect runs the other way. AfterQuery’s sample shrank from 78 to 43 distinct rates over the month. A week after the snapshot, on October 7, the live index read $60, after Alignerr’s listings were collected again with their pay (67 distinct rates). A market-wide median in this space is a statement about which platforms were visible on the day. That is why this report freezes its date, and why the platform rows are a better guide to direction than the headline number.

Openings move quickly

On October 1, 7,043 listings were open and 4,751 had closed since tracking began. In the last complete week before the snapshot, September 21 to 27, we recorded 330 newly seen listings and 369 closures. These are public listings, not hires. A listing can close because a project filled, paused or stopped being published, and the data cannot tell those apart. Some closures and reappearances also reflect our own coverage, as in the section above, rather than anything the platform did.

Among listings that have closed, the median time from first sighting to last confirmation was seven days. That is a statistic about closed listings only, and it understates the lifespan of listings still open. The practical lesson is modest but real: a rate on an old screenshot is weak evidence that the opportunity still exists. Check the source listing before you apply.

Who says what they pay

6,016 of the 7,043 open listings (85%) stated some pay. That includes per-task and project amounts, which are left out of the hourly medians because converting them would mean guessing how long the work takes. 5,716 listings carried a comparable hourly figure.

The site-wide share hides a wide spread between platforms, so the comparison worth making is platform by platform:

Hedging is concentrated. 2,916 of the 5,716 hourly listings (51%) qualified their figure, with wording such as “tend to pay”, “up to” or “approximately”, and 2,717 of those were Prolific’s, where 94% of hourly listings are worded that way. Mindrift hedged every hourly figure it posted. A site-wide hedging rate would mostly describe one platform, which is why the table breaks it out.

Neither column measures generosity or honesty. A platform near the bottom has not been shown to pay less, and one at the top has not been shown to pay what it advertises. The leaderboard measures one thing: whether a reader can see a number before applying.

How we calculated the rates

The HumanSourcer Pay Index collects publicly visible listings from AI-training platforms. It reads structured pay fields where a platform provides them and otherwise extracts the figure from the listing text, keeping the source wording. A missing rate stays missing. A per-task amount is never divided by a guessed number of hours to make an hourly figure.

• Ranges use their upper figure. A median here is a median of advertised figures under that rule, not a midpoint and not a typical realized wage.
• Each distinct (platform, pay wording) pair counts once. Where a listing has no pay wording, its number is used instead. This stops one repeated sentence from dominating the distribution. It can also merge two genuinely separate offers that happen to use identical wording, which understates the sample rather than inflating it.
• Only US-dollar rates are included. Listings paying in pounds or euros are excluded from every dollar figure rather than converted at a chosen exchange rate.
• Dollar figures are rounded to whole dollars. The middle half’s lower bound is $23.38 before rounding.

Every figure above comes from three records of the same day, which agree to the listing: the October 1 archive of the Pay Index, the full listing dataset the site published that day, and the lifecycle data generated the same day. Researchers can request the underlying data at /for-researchers/.

To cite this report: HumanSourcer, “State of AI Training Pay: Q3 2026”. Median advertised AI training rate $52/hr on October 1, 2026 (n=740 distinct advertised rates across 26 platforms).

Use the report when choosing work

For entry rules and how payment works, start with the sourced Outlier and DataAnnotation guides, or the current Turing AI Advancement Work listings. DataAnnotation’s advertised starting rates are recruiting copy, rather than the discrete listing rates measured here.

The free lesson on how pay and tiers work explains providers’ public pay rules. To learn about an evaluation task, read judging two responses and try a free sample. These resources teach general skills; HumanSourcer is independent of the platforms, and practice does not predict selection or earnings.