How we build every pay figure
A rate on this site is copied from a job listing, published by the platform, or estimated by us. Estimates are always labeled, and only rates copied from listings go into a median.
By Pietro Romeo · Method last reviewed · Live figures from
The median listed rate today is $70 an hour
- Open listings
- 4,353
- Distinct roles
- 4,078
- Listed hourly rates
- 2,848
- Middle half
- $40 to $100
Show the numbers
| Hourly rate | Rates |
|---|---|
| $0 to under $10 | 89 |
| $10 to under $20 | 248 |
| $20 to under $30 | 205 |
| $30 to under $40 | 143 |
| $40 to under $50 | 160 |
| $50 to under $60 | 220 |
| $60 to under $70 | 200 |
| $70 to under $80 | 384 |
| $80 to under $90 | 345 |
| $90 to under $100 | 139 |
| $100 to under $110 | 250 |
| $110 to under $120 | 82 |
| $120 to under $130 | 129 |
| $130 to under $140 | 41 |
| $140 to under $150 | 18 |
| $150 to under $160 | 80 |
| $160 to under $170 | 10 |
| $170 to under $180 | 28 |
| $180 to under $190 | 13 |
| $190 to under $200 | 3 |
| $200 and up | 61 |
Every pay figure has one of three sources
Listing-stated
The rate printed on the platform's own job post. We read every board daily.
Platform-stated
One standard rate a platform publishes for all its work. It shows on the platform's profile.
Our estimate
Our range for a listing that gives no rate. It always carries an estimate label.
Job pages show a listing's exact numbers. Job cards round them to the nearest $5 or $10 and turn a single number into a narrow range around it. Work paid per submission or as a one-time amount shows that amount, with no hourly conversion.
Turing posts no rate on any role, and a few other listings arrive without one. For those we estimate from typical pay in the role's field, adjusted for the countries it is open to. An estimate never enters a median and is never sent to Google as a salary. A platform that only has estimates shows "Rates not published" on comparison pages.
A role posted fifty times counts once
Platforms often repost the same role once per project or per country. We treat one platform, one title and one rate as a single role, so today's 4,353 open listings come down to 4,078 distinct roles. Medians are built from distinct rates, which keeps a heavily reposted role from pulling the figure toward its own pay.
Only stated hourly rates enter a median
- Stated hourly rate 3,119
- Hourly maximum only ("up to $X") 262
- Placeholder range, like $1 to $500 6
- Paid per task, month or year, or no rate 159
- One rate for all of a platform's work 442
- Our estimate 364
Only the green row goes into the median. Once reposts count once, its 3,119 listings give 2,848 distinct rates.
- Per-task, monthly and yearly pay stays out. Turning it into an hourly rate would mean guessing the hours.
- A range counts at its midpoint.
- "Up to $X an hour" is the best case on a role, and most people on it earn less, so those 262 rates are reported on their own.
- A range whose top is more than ten times its bottom, like $1 to $500 an hour, is a placeholder and stays out.
- A platform, field or country with fewer than 10 distinct rates shows a count, with no median.
Platform, field and country pages, the markup search engines read and our llms.txt summary all follow these rules.
General tasks and expert roles are priced apart
90% of the listed hourly rates on this board are for expert roles in law, medicine, finance, coding, science and other professions. Pooled together, the median lands near the expert rate and tells someone looking for a first annotation or recording task very little. The homepage shows both figures for that reason.
Middle half of rates Median
We sort listings by job title. A title that names work needing no professional background, such as annotation, labeling, video or voice recording, transcription, rating AI answers or bilingual generalist work, and that names no profession, field of study or seniority, is a general task. Everything else is an expert role. The title decides it because platforms often ask for a degree on plain recording and transcription work.
The general-task figure rests on fewer rates (281 today), and some of them are low country rates: video capture work open to Asia and Latin America can pay under $10 an hour. Those rates are posted on this board, so they stay in. Platforms that publish one rate for all their work, or no rate per listing, are in neither figure.
Some pay never shows up on a listing
- Your rate is set in the offer, and it can move with your experience, your country and the project.
- Bonuses, rate changes and invite-only projects happen inside the platform after you join.
- We read 20 platforms. Work posted anywhere else is missing from every figure.
- A listing doesn't say how many people it will hire or how much work follows onboarding.
Check the rate on the platform before you spend time on an assessment. If our figure doesn't match the platform's own page, tell us and we'll correct it.
Paid placements never touch the pay figures
- A pin or a star on a job card marks a paid placement.
- Cards in the sidebar are ads and partner offers.
- Featured, the default order, puts paid placements first. The rest are ranked by your location, pay, how recent the listing is and our rating of each platform, which reflects how reliable it has been for applicants and whether we partner with it.
- Newest sorts by date, with some platforms grouped lower.
- Highest pay sorts by rate alone.
Medians, counts and datasets are computed from every listing with the rules above, whoever paid for placement. Our commercial relationships with platforms and tools are listed in the affiliate disclosure.
Every month's data is archived with a DOI
Live figures change daily, so a monthly snapshot is archived on Zenodo under DOI 10.5281/zenodo.23042213, with each past version kept unchanged, and mirrored on Hugging Face. Anyone can check a number we published against the month it came from. The data is licensed CC BY 4.0.
Two related datasets sit beside it: platforms with their ownership and verified hiring terms (DOI 10.5281/zenodo.23042850), and our study of Trustpilot reviews of AI training platforms (DOI 10.5281/zenodo.23018069). Field definitions and downloads are on the open data page.