The first place it showed up in pay

Why the platforms first

A job is a bundle: tasks plus responsibility, relationships, the things that do not fit a ticket. A freelance gig on a platform is the task with the bundle removed. It is specified in a paragraph, priced per piece, delivered as a file, and judged on the file. That is, almost exactly, the shape of work a model does well: input-rich, output-defined, checked by the buyer, no name on it.

So if the models were going to move any market, the freelance platforms were where it would show, first and most cleanly, because the platforms record every posting and every payment. Economists noticed this and looked.

What the platform data found

A study of the largest freelance platform compared the categories of work most exposed to the chatbot (writing, editing, copy) with those least exposed, before and after the chatbot’s release in late 2022. Within months, the number of jobs posted in the exposed categories had fallen by a couple of per cent relative to the others, and the earnings of freelancers in them by around five per cent. When image generators became widely usable, the same pattern appeared in graphic design and illustration, and more steeply: the studies put the fall in image-related jobs at around seventeen per cent.

A second study, using the same platform, found that postings for writing, software and engineering gigs fell by about a fifth in the eight months after the chatbot’s release, relative to manual-intensive gigs that a model could not touch. The freelancers who remained competed for fewer postings at lower rates.

These are not large economies. They are early, precise measurements of the mechanism in the place the mechanism bites hardest.

The translators

Translation is the case with the longest history (machine translation had been eating its low-value end since 2016) and the clearest survey. In early 2024 a British authors’ organisation surveyed its members: over a third of translators said they had already lost work to generative AI, and over forty per cent said their income had fallen. Around a quarter of illustrators reported the same.

A language-learning company that had used contractors for translation cut around a tenth of them at the start of 2024, and in 2025 announced it would treat AI as the default for work a model could do. The remaining human work moved toward checking a machine’s draft, at a per-word rate below what drafting had paid: the fluency trap from the second module, monetised.

What the pattern tells you

Three things, each of which the module has already built.

The task without the job goes first. An employed writer has a manager, a house style, a deadline culture and a reputation; a freelance writer on a platform has a brief and a price. The model competes with the second far more directly than with the first, which is why in-house employment in the same occupations showed less movement in the same period.

The low-value end goes first. The postings that vanished were product descriptions, listicles, short blog posts, simple logos: the tasks whose value was already close to the cost of a competent median. The bespoke, the accountable and the reputational held better. This is the elasticity rule from the previous lesson at the level of the individual gig.

The effect arrives before the announcement. The platform data moved within months of the chatbot’s release, before most companies had a policy and before most surveys had been run. Markets that price per task price the model in immediately.

For the person on the platform

The honest advice follows from the pattern. Move up the bundle: sell the checking, the judgement and the responsibility rather than the draft, because the draft’s price is now the model’s price. Move toward work that is specific to a client’s situation rather than generic to a category, because the model’s median is generic. Use the tools yourself, since the buyer already assumes you do and prices accordingly. And know that for some categories (the product description, the stock logo) there may be no rung above the one that disappeared, and the honest move is out of the category rather than up it.

None of that is comfortable. It is what the numbers say, and the numbers arrived first for the people with the least cushion.