Why AI projects pay experts, and where it’s heading

The work, in one paragraph

AI developers train and test their models partly with human judgement. Expert work is the part that needs a professional: writing hard questions and model answers in your field, grading and comparing an AI’s answers, writing marking criteria, and explaining step by step how you reached a judgement. You may see it called “AI training work”, “AI expert jobs” or “human data”. It is usually remote contract work, project by project. Our guide explains the work and the vetting in detail.

Why developers pay for it

Good public text may run out soon. Researchers at Epoch AI estimate that AI models will have used up the stock of good-quality public text written by people somewhere between 2026 and 2032^{source 1·, grade B}. Stanford’s AI Index 2026 repeats that estimate and finds no clear evidence that AI-generated text can fully replace it in early training^{source 2·, grade B}. Our view: this is one reason developers pay for new human data, written on purpose by people who know the field.

Models are now trained to do tasks, not just to answer. Developers increasingly train models on tasks with a way to check the result, in set-ups often called “environments”. People who build them told Epoch AI that knowing the field matters more than machine-learning skills when writing those tasks^{source 3·, grade B}.

Someone has to check the answers. Testing a model in law, medicine or finance takes people who can tell a good answer from a confident wrong one; the guide gives public examples, such as physicians writing the marking criteria for a health benchmark. In the EU, providers of the general-purpose AI models the AI Act classes as having systemic risk must evaluate them, including adversarial testing, to identify and reduce those risks^{source 4·, grade A}. These rules apply from 2 August 2025; models already on the market before then have until 2 August 2027^{source 4·, grade A}.

Who pays. AI developers, usually through specialist data companies that find, vet and pay the experts. We don’t name them.

Where it’s heading

Most forecasts point up, but they disagree on how far, and they come from firms that sell market reports. Read them as direction, not as precise sizes. The numbers are in the block below. Three signals stand out:

• Both forecasts we checked expect the market to grow several times over by 2030 to 2034^{source 5·, grade C, source 6·, grade C}.
• Companies that supply training data report fast growth. One said its gross revenue run rate doubled in four months; another reported eighteen-fold growth in a year^{source 8·, grade B, source 9·, grade B}. These are the companies’ own figures.
• People are moving into the work. On LinkedIn, “data annotator” was the fourth fastest-growing job title in the US^{source 7·, grade A}.

What could change it

We’d rather you hear the limits from us:

• AI-written data is already useful. The AI Index finds it already helps in later training stages, such as fine-tuning and reinforcement learning^{source 2·, grade B}. Our view: some tasks may need fewer people over time.
• Developers build in-house teams too, not only through outside companies^{source 3·, grade B}.
• Growth for the market is not work for you. Roles vary by field and language. Projects can end or change at short notice, as the guide explains.
• Forecasts are estimates. The two firms below use different definitions and different years, and their figures don’t line up.

What this means for you

If you have real professional experience, these sources suggest demand for expertise like yours is growing. That is all the numbers say. They don’t say you’ll be accepted, how often there will be work, or what you’ll earn. We can’t promise work, interviews or pay.

What we can do: show the roles open in your field each week, dated, with the pay as the platform posts it, and explain how the vetting works.

The numbers

Each source is linked in “Sources” at the end of the page. Forecasts are estimates by the firms named; we have not checked their methods, and some are sold reports we could read only in summary.