The state of the AI training job market

Cumulative hires by time since referral

Share of hires with a known hire date, rounded to 5 points

Per-platform figures are not published. Referral volume and follow-through vary by platform, so a per-platform number reflects our referral process as much as the platform.

How 60 companies that pay people to train AI hire, pay and accept contributors. Every fact links to the company page we checked it on.

One JSON file holds every figure on this page, plus monthly pay, contract length and weekly hours. You can download or link it freely, with no sign-up or API key.

For citations, use the dated monthly snapshot. Snapshots are frozen once their month closes: https://aitrainer.work/data/market-trends/market-trends-2026-11.json

The open listings behind the job board, as JSON. No sign-up or API key. Responses refresh about once an hour and fall under the same CC BY 4.0 license as the dataset.

Listings are newest first. A job's pay_source of estimate means the platform published no rate and the figure is ours. The whole board is also one file at /data/jobs.json (gzipped at /data/jobs.json.gz).

Machine-readable spec (OpenAPI 3.1): /openapi.json.

If you show these listings, link each one to its url. To pick roles by platform, field, country or pay and get an RSS or JSON link, use the job feed builder.

Add our MCP server to Claude, ChatGPT, Cursor or any assistant that supports MCP, and it can search open AI training roles and pay for you. No sign-up or key.

• match_jobs_to_profile ranks open roles for one person from their profession, skills, languages, country and minimum pay. Nothing about the person is stored.
• search_jobs finds open roles by keyword, platform, field, country or minimum hourly pay, most relevant first.
• get_job returns one role in full: requirements, hours, interview and the countries it accepts.
• get_interview_prep explains how a platform hires and gives practice questions for a role, for a mock interview.
• list_platforms compares platforms by open roles, median pay and payout speed.
• get_pay_stats gives market pay and the hiring numbers from the monthly Pay Index.

In Claude, open Settings, then Connectors, choose "Add custom connector" and paste the URL. In Claude Code, run claude mcp add --transport http aitrainer-work https://aitrainer.work/mcp. In Mistral's Le Chat, open Connectors, click "Add Connector", switch to the "Custom MCP Connector" tab, name it aitrainer_work and paste the URL.

Free to use for commercial and non-commercial research, reporting, or applications. Attribution with a link to this page is required.

The license covers every file and endpoint on this page: /data/market-trends.json, /data/companies.json, /data/companies.csv, the jobs feed at /data/jobs.json and /data/jobs.json.gz, and the /api/jobs.json API.

Full license: creativecommons.org/licenses/by/4.0

Corpus

All listings observed since tracking began, including listings that have since closed. The dataset has no active/expired split, because that split depends on the build date.

First-seen dates

The date a listing was first observed by this project, which can be later than the platform's publication date. Listings already open when tracking began on a platform are dated to their first observation. Each platform's first month of coverage includes a backlog of listings already open at onboarding. These months are flagged includes_backfill: true and should be excluded from growth comparisons.

Pay figures

Hourly-priced listings only, in USD. Where a listing advertises a range, the midpoint is used. Each distinct rate counts once: a role relisted on the same platform with the same title and rate is one data point, and sample_size counts distinct rates, not listings. Excluded: rates below $3/hr or above $400/hr (parsing errors), max-only rates where the platform publishes only a ceiling, open-ended placeholder ranges whose top is more than 10 times the bottom, per-task, monthly and annual pay, platform-wide disclosure bands, and any rate that is our estimate rather than one the listing states, including rows saved before estimates were tagged. platform_balanced_median_hourly_usd is the median of per-platform medians (platforms with at least 10 distinct rates), because one platform supplies about half of all distinct rates.

Fields

Field (domain) is assigned from listing title, description, and platform tags. Approximately one third of listings have no confident field assignment and are excluded from field breakdowns. Field names used by older releases of the classifier (for example Law, Biology, Voice) are folded into the current ones (Legal, STEM, Voice Acting).

Listing lifespan

Days between first and last observation of a listing. Listings still open at generation time are excluded, which biases the median slightly downward.

Sample suppression

Breakdowns with fewer than 10 listings are suppressed.

Known limitations

Coverage is limited to tracked platforms and does not represent the full market. Pay disclosure varies by platform, so per-platform sample sizes vary widely. Read each median alongside its sample_size.

Referral outcomes: scope

Outcomes for candidates aitrainer.work referred to AI training platforms since March 2026, pooled across platforms and rounded. Counts are floored (referrals to the nearest 1,000, placements to the nearest 50), the rate is a whole percent, and time figures are rounded to 5 days and 5 percentage points.

Referral outcomes: time to hire

Days from a candidate's first recorded referral to the platform's confirmed hire date, pooled across platforms that report a hire date for each candidate. Reported as a median to limit the influence of outliers. Most dated hires come from one platform, so the figure mostly reflects that platform's timing.

Referral outcomes: placement rate

Placements divided by candidates referred, pooled across all tracked partners. A referral is an introduction, and many referred candidates never complete the partner's application, so this rate describes the referral funnel and should be read as a lower bound. It does not represent any individual applicant's probability of being hired.

Referral outcomes: known limitations

Platform-reported and refreshed periodically, so figures rise as outcomes are recorded. Hires not attributed back to aitrainer.work are not captured, which biases the rate downward. Per-platform figures are not published.