OpenAI and Upwork data shows AI reshuffling freelance work, squeezing single-task gigs

Freelance markets are an early-warning system for labor. Contracts are short and clients rewrite requirements with every posting, so shifts appear there before official employment statistics. OpenAI’s economic research team, in its Work at the Frontier series, paired Upwork posting data with ChatGPT usage patterns from January 2025 to June 2026. Upwork’s own research says 38 percent of skilled U.S. knowledge workers freelance, and its earlier study estimated freelance knowledge work generated $1.5 trillion in U.S. earnings in 2024. The report also notes what OpenAI calls a capability overhang, the gap between how people use ChatGPT today and what AI can support. OpenAI recently published country-level work-usage data showing usage growing more varied (our coverage).

Across 63 work categories, the pattern is clear. In 43, activities relatively common in ChatGPT lost share on Upwork while less common ones gained. In the other 20, the reverse held. That’s a real signal of substitution in some kinds of work, and OpenAI deserves credit for saying plainly that ChatGPT use alone doesn’t explain the changes.

AI-related postings combined tasks that rarely appeared together before, such as AI-generated video, editing and motion graphics in one brief. Their task mix became 45 percent more diverse between early 2023 and late 2025, roughly like going from about 15 equally common tasks to about 21. Those projects carried larger fixed-price budgets and more hours, while posted hourly rates stayed similar.

That is good news for freelancers who can integrate and a warning for those whose work was a single, well-defined deliverable. The generalist who can stitch AI output into a finished product looks more valuable. The specialist doing one task a chatbot now handles reasonably well is in a shrinking category. This isn’t the end of freelancing; it’s a story about who gets squeezed inside it. Inside companies the same shift is happening as agents like OpenAI’s Dots take on routine work (how Dots works) and cheaper models such as GPT-6.1 Terra target office tasks (our analysis), although Terra isn’t in this study’s window.

This is research by an AI company and a freelance marketplace, both of which benefit from a narrative that AI grows the pie. The methods are described and the hedges are there, but the underlying data should be opened to independent economists. Larger budgets that partly reflect broader scope aren’t the same as higher pay per hour of skilled work, and the report itself notes hourly rates were similar.

If the pattern scales, AI fluency becomes table stakes in freelance work and value moves to scoping, judgment and assembly. That can be healthy. It can also leave a long tail of workers in shrinking categories with no clear path into growing ones. More in our future of work coverage and OpenAI coverage.