Thirty-Three Dollars of Freelancer for Every Dollar of ChatGPT

Ryan Stevens, using payments data from a U.S. expense management platform, tracked what firms actually spent on freelance labour marketplaces (Upwork, Fiverr, et al.) versus AI model providers (OpenAI, Anthropic) from 2021 through Q3 2025. The finding getting all the attention: for every $1 decline in freelance platform spending, firms added roughly $0.03 in AI spending. A 33:1 cost ratio.

That's a real number from real spending data, which already puts it ahead of 90% of the "AI will replace your job" literature, most of which is vibes, surveys, and theoretical exposure scores. Credit where it's due.

But notice what's actually being measured. The study only tracks firms that were already outsourcing task work to strangers on the internet before ChatGPT launched. And it only captures payments to OpenAI and Anthropic — not Google, not Microsoft, not any of the other ways companies spend on AI — because the expense data couldn't isolate them. So the aperture is: companies that were buying piecework from freelancers, as seen through two AI vendors.

The freelance gig economy — short-term, deliverable-based, often writing or basic code — is the single most exposed category of work to generative AI substitution. Finding that firms are spending less on it post-ChatGPT is like finding that firms spent less on typing pools after the word processor. True, meaningful, and not particularly generalisable.

Here's the question the spending data can't answer: when a company drops a $500 Fiverr copywriting gig and spends $15 on ChatGPT instead, did the work vanish? Or did it move to an internal employee who now spends twenty minutes editing AI output instead of briefing a freelancer and waiting three days? The payments data shows substitution. The reality might be task reorganisation — the labour didn't disappear, it became invisible to this dataset.

To his credit, Stevens says exactly this. The paper explicitly notes that micro-level substitution doesn't tell you about aggregate labour market effects, and that the findings are consistent with AI being labour-augmenting overall. That caveat will be in approximately zero of the LinkedIn posts citing the 33:1 number this week.

The real story here isn't "AI is replacing workers." It's that firms are finding dramatically cheaper ways to accomplish tasks they were already comfortable outsourcing to strangers with no context. That's a meaningful and specific finding. It tells freelancers on task platforms something urgent. It tells the rest of the workforce almost nothing.