What AI training work actually is — and what it really pays

"AI training work" is the label a cluster of platforms use for the same arrangement: an AI lab needs examples, corrections and judgements produced by people who actually know a subject, and it pays contractors by the hour to produce them. The listings call the contractor an expert, a trainer, a data contributor or an annotator. The work behind those words is largely the same.

What the work actually is

Across the 312 live listings we carry on 10 August 2026, the tasks described fall into four recognisable shapes:

• Write the answer a model should have given. You are shown a prompt and a model's attempt, and you produce the response a competent professional in your field would have written, with the reasoning shown.

• Rank or correct what the model produced. Two answers, and you say which is better and why — or you mark exactly where a chain of reasoning went wrong.

• Set the problems. Harder projects ask domain experts to write questions a current model still fails, which is why they pay the most and reject the most.

• Label and check data. Annotation, transcription, and quality review, often in a specific language or accent. This is the lowest-paid tier and the easiest to get into.

It is contract work, hourly, remote, and typically part-time by design — a project runs while a lab needs that data and stops when it does not.

What it is not

It is not machine-learning research. You are not training models in the engineering sense: you are the domain expert whose judgement gets recorded. Most listings ask for expertise in a subject — law, medicine, a trade, a language — and no AI background at all. micro1's own guide says most roles need no prior AI experience, and the exceptions it names are ordinary seniority requirements like "3+ years of full-stack experience" for a developer role.

It is also not employment. There is no salary, no notice period, and no guarantee of hours. A project that pays $90/hour and offers four hours of work in a month paid $360 that month.

What it pays

Here is the part almost nobody publishes: the actual advertised rates, by field, across every live listing we hold. 291 of the 312 listings state an hourly band; the rest state none.

Read the table as what employers advertise, not as what people earn. The band is the median advertised minimum to the median advertised maximum in that field. Two listings quoting an annual salary in the hourly field are excluded.

The pattern is consistent and worth internalising before you spend an evening on an application: pay tracks how hard your expertise is to replace, and volume runs the opposite way. Robotics and generalist work carry more than half of all 18,182 openings at the bottom of the pay range; law, applied engineering and the research fields pay two to eight times more and have a fraction of the seats.

Where you have to live

Eligibility is the constraint that surprises people, and it is platform-wide rather than per-role. Every listing we sampled declared the same 58 countries in its structured data:

Algeria, Argentina, Austria, Bahrain, Bangladesh, Belarus, Belgium, Brazil, Canada, Chile, Colombia, Denmark, Egypt, Ethiopia, France, Germany, Ghana, Greece, Hong Kong, India, Indonesia, Iraq, Italy, Japan, Jordan, Kazakhstan, Kenya, Kuwait, Kyrgyzstan, Lebanon, Libya, Malaysia, Mexico, Morocco, Netherlands, Nigeria, Oman, Pakistan, Palestine, Peru, Philippines, Portugal, Qatar, Russia, Saudi Arabia, Singapore, South Africa, Spain, Sri Lanka, Switzerland, Taiwan, Thailand, Tunisia, United Arab Emirates, United Kingdom, United States, Uzbekistan, Vietnam

One honest caveat. The platform's own marketing pages carry a different and partly contradictory list — they name Belarus, Hong Kong, Iraq, Libya, Russia and Ukraine among regions it does not hire in, while the listings' structured data declares five of those six eligible. We publish what the listings themselves declare, because that is the machine-readable field, but if you are in a country on either list, confirm it before you invest the evening. If your country appears on neither, this whole category is closed to you today, and no amount of qualification changes that.

Who this suits

It suits someone with a real professional specialism, an evening or two a week, and tolerance for irregular work. A lawyer, a nurse, a machinist, a native speaker of a language with thin training data — those profiles are being paid at the top of the table above precisely because the supply is short.

It suits you badly if you need predictable monthly income, if you want to break into AI as a career (this is not that ladder), or if the pay only works at the rate you would need to be accepted for the hardest tier.

Where these facts come from

The rates, openings and country list are measured from the 10 August 2026 snapshot of the partner listings on this site — 312 live roles, 18,182 openings, 18 fields — not from a survey or a review site. The process and payment details are quoted from micro1's referral-program page and its expert guide.

Disclosure: JobsRadar earns a referral fee if you are hired through one of these partner listings. It does not change what we publish — the pay table above is the same table whether it flatters the platform or not — but you should know it, and the outbound links are marked as sponsored throughout the site.

See the AI-training roles currently live →

New roles land on our Telegram channel first: https://t.me/ai_training_jobs_radar