The work that trains the models

An industry that does not appear in the industry’s story

Every model you have used was shaped by people who were paid by the task. They wrote reference answers, ranked two responses against each other, flagged unsafe outputs, drew boxes around objects, transcribed speech, checked whether a citation was real, and argued in a comment thread about whether a guideline covered an edge case.

The names to search for are Scale AI and its worker platform Outlier, Surge AI, Appen, Toloka, Sama, TELUS Digital, Invisible Technologies, Mercor and Turing. Between them they engage hundreds of thousands of people. Almost none of that appears in how the field describes itself, and it is one of the widest doors into AI work that exists for someone with no job history.

What it actually pays, and how

Pay is per task or per hour, and it splits sharply by what the task requires:

• General labelling and transcription: the low end. In low-cost markets this is frequently a few dollars an hour, sometimes less once unpaid time is counted.
• Language work in an under-served language: better, because supply is short. Fluent Yoruba, Marathi, Vietnamese or Amharic is a genuine scarcity.
• Domain expertise: the high end. Qualified doctors, lawyers, accountants, translators and competitive programmers writing evaluation data are paid tens of dollars an hour, occasionally more.

The structural things nobody puts in the advert:

• Qualification tests are often unpaid, and can take hours.
• Projects end without notice. You may work forty hours one week and none the next.
• Accounts get deactivated for quality-score reasons with limited appeal. Read the platform’s own community forums before committing.
• Content moderation and safety work carries real psychological cost. Reviewing violent or sexual material for months damages people. It is documented, it is not weakness, and if a role involves it you should know that going in and treat support as a requirement rather than a perk.

Be clear-eyed. Some of this work is well paid and interesting. Some of it is precarious piecework dressed in the language of a technology career.

Why it is still worth considering

Two reasons, and neither is the money.

First, it shows you what models get wrong. Nothing else gives you as many hours staring at bad model output next to a definition of what good would have been. That is the raw material of evaluation, which the evidence module treats as the scarcest hireable skill in the field. Somebody who has ranked four thousand responses has real intuitions about failure modes that a course cannot give.

Second, you learn the vocabulary of data quality from the inside. Annotation guideline, gold set, inter-annotator agreement, adjudication, label drift, edge-case memo. These are the exact words used by the teams that build training and evaluation data, and using them correctly in an interview marks you as someone who has done the work rather than read about it.

Turning the work into evidence

The failure mode is treating these platforms as a career ladder. They are largely flat by design: there is no promotion track from labeller to research engineer, and waiting for one wastes years.

The move that works is converting the experience into an artefact you own. All of these are things you can do without breaching a confidentiality agreement, because they describe your own process rather than the client’s data:

1. Write a guideline for a task of your own. Take a public dataset, define a labelling task in one page, and state how to handle five edge cases.
2. Label a sample twice, a week apart, and measure your agreement with yourself. Report the percentage. Self-agreement below about 85% means the guideline is ambiguous, not that you are careless.
3. Get one other person to label the same 100 items and compute agreement between you. Then write up the disagreements and how you resolved them.
4. Publish the guideline, the numbers and the adjudication notes.

That package is a portfolio piece for any data-quality, evaluation or trust-and-safety role, and very few applicants have one.

What to check before signing up

• Is the pay stated per hour or per task, and what does the community say the effective hourly rate is?
• Is the qualification test paid?
• How does the platform pay out, in what currency, and what are the fees?
• Does the work involve harmful content, and is that disclosed up front?
• What are the terms about who owns work you produce and whether you may describe the project publicly? Usually you may not name the client. You can almost always describe your own methods.