要点
- On 10 November 2026, OpenAI published a blog, "Work at the Frontier: AI's effect on freelance work," with a research report co-authored with Upwork, "The Rebundling of Freelance Work in the AI Age." The authors, listed in alphabetical order, are Caroline Chin, Shagun Kala, Nadav Kunievsky, Teng Liu, Alex Martin Richmond, and Carl Shan.
- Content writing and professional and business writing accounted for 14.5% and 10.2% of more than 36,000 sampled ChatGPT conversations, compared with about 1.3% and 0.4% of Upwork Marketplace job posts in the first half of 2026. Of 63 matched work subcategories, 43, or 68.3%, either had lower ChatGPT intensity and gained Upwork share, or higher intensity and lost share, from the first half of 2023 to the second half of 2025.
- Between early 2023 and late 2025, the effective number of task clusters in AI-related postings rose from 14.6 to 21.2, about 45%. Historically uncommon task pairs, those appearing in fewer than 200 pre-AI postings, made up 8.4% of pairs within AI-related jobs, against 2.4% within other jobs. Fixed-price budgets were about 70% higher raw for AI-related jobs and about 52% higher after controlling for subcategory and month. Hours worked in the first 90 days after hire on hourly contracts with positive recorded hours were about 41% higher after the same adjustment. Price and hours data run through April 2026, and the top 1% of posted prices is trimmed.
- ChatGPT classifications were conducted using GPT 5.6 Luna. The report says no researcher read user messages or looked at individual user data. The analysis does not measure aggregate Marketplace growth or identify causal substitution of jobs. It is OpenAI and Upwork research, not an independent academic replication.
- Some writing categories really did lose share, and that is part of the evidence. Clients may be asking for AI in the brief while still buying implementation, production, and delivery that ChatGPT use does not replace.
On 10 November 2026, OpenAI and Upwork released a research report, "The Rebundling of Freelance Work in the AI Age," and OpenAI posted a shorter account in its Work at the Frontier series. The authors, listed alphabetically, are Caroline Chin, Shagun Kala, Nadav Kunievsky, Teng Liu, Alex Martin Richmond, and Carl Shan, from OpenAI and Upwork. Where the blog and the report disagree, we follow the report. The checkable claim is not that AI is simply substituting for freelancers. It is that on Upwork, many categories that are common in ChatGPT use lost marketplace share, while AI-related postings became about 45% more diverse and carried larger fixed-price budgets, about 70% higher raw and about 52% higher after controlling for work subcategory and posting month, plus more hours on hired hourly contracts.
[1][2]What the report actually measured
The report compares two views of work that are already digital and remote. ChatGPT conversations show what people do directly with a general-purpose model. Upwork job posts show what clients ask freelancers to do, how those activities are packed into one project, and what budgets or hourly rates they attach. This is not a picture of the whole labor market. OpenAI's blog cites Upwork estimates that extend well beyond this marketplace: the 2026 Future Workforce Index, which found that 38% of skilled U.S. knowledge workers freelance, and a previous annual study that put 2024 U.S. earnings from freelance knowledge work at $1.5 trillion. The new report is narrower. It describes composition and bundling on this marketplace, not the size of U.S. freelance work.
ChatGPT shares cover January 2025 through June 2026. Methods say monthly samples from individual consumer accounts, targeting 5,556 conversations a month, produced 36,896 personal or work-related conversations after dropping schoolwork, unclear-purpose chats, and those with no Upwork subcategory match. GPT 5.6 Luna did the labeling. The report says no researcher read user messages or looked at individual user data, so the ChatGPT activity mix is model-labeled by OpenAI.
Upwork figures use two time windows. One exhibit tracks category share from the first half of 2025 to the first half of 2026, in percentage points, as a share of all Marketplace posts in the first half of 2026. The 63-category scatter runs from the first half of 2023 to the second half of 2025. AI intensity is a category's share of ChatGPT use in the second half of 2025 divided by its share of Upwork posts in the first half of 2023. The scatter drops AI Apps & Integration and any category with no Upwork posts in that 2023 baseline. AI-related postings, under Upwork's internal pipeline, are those that request AI skills, require AI tools, or involve improving AI-generated output. Price and hours data run through April 2026.
[1][2]Where ChatGPT use and Upwork demand diverged
The two mixes are not the same. Writing is a much larger share of ChatGPT use. Video and animation are a larger share of Upwork postings. Content writing and professional and business writing accounted for 14.5% and 10.2% of the sampled ChatGPT conversations, compared with about 1.3% and 0.4% of Upwork job posts in the first half of 2026.
Subcategories with higher ChatGPT intensity tended to lose Upwork marketplace share. Those with lower intensity more often gained share. Content writing is the clearest example. It accounted for roughly three times the share of ChatGPT activity in the first half of 2025 as of Upwork postings in the first half of 2023. Over that two-year period, content-writing postings fell at more than twice the rate of Upwork postings overall. Video and animation show the other side of that pattern, with lower ChatGPT intensity and postings that grew by about a third relative to the overall marketplace.
Of the 63 subcategories, 43, or 68.3%, followed the simple pattern of lower ChatGPT intensity and a gain in Upwork share, or higher intensity and a loss of share. The other 20 did not. Six, or 9.5%, had higher ChatGPT intensity and still gained Upwork share. Fourteen, or 22.2%, had lower intensity and still lost share. A substitution story explains most of the scatter, but not all of it.
[1][2]Larger budgets, more hours, mixed tasks
The authors call this a rebundling of work: clients packing different tasks into the same job. Inside AI-related postings, the mix of work broadened. Between early 2023 and late 2025, comparing the first two quarters of 2023 with the last two of 2025, the effective number of task clusters rose from 14.6 to 21.2, about 45%. That measure converts concentration into an equivalent count of equally common task types, so a higher number means the work is less bunched in a few activities. The average AI-related post still listed only slightly more tasks than other posts, 3.35 compared with 3.19, a difference of about 5%.
The bigger change is which tasks are packed together. Historically uncommon task pairs, those appearing in fewer than 200 pre-AI postings, accounted for an average of 8.4% of pairs within AI-related jobs, compared with 2.4% within other jobs. The blog's example is a single brief that asks for AI-generated video, editing, and motion graphics, with the freelancer hired to deliver a finished video.
Posted prices and hours point toward larger projects rather than a higher price for each hour. Fixed-price budgets, the amount a client posts for a project or deliverable, were about 70% higher raw for AI-related jobs, and about 52% higher after controlling for work subcategory and posting month. Posted hourly rates were only about 9% higher raw, and about 1.5% after the same adjustment. Among hired hourly contracts with positive recorded hours, hours worked in the first 90 days after hire were about 41% higher for AI-related work, after the same subcategory and month controls. The top 1% of posted prices is trimmed. The report says the comparisons are descriptive, not causal.
[1][2]What the evidence cannot show
The report is explicit about what it does not measure. The 63-category figure does not measure aggregate Marketplace growth, and it does not identify causal substitution of jobs. Share changes are about mix, not about whether the marketplace grew or shrank. Hours are observed only on hired hourly contracts that recorded some hours, and only in the first 90 days. The ChatGPT side is a consumer sample classified by GPT 5.6 Luna, and captures current use, not the full range of tasks a model can support. The blog calls that gap a "capability overhang."
The work is also not an independent academic replication. The authors are from OpenAI and Upwork, studying those two platforms' own data. A reader should treat the findings as a carefully documented internal study, not as a result re-run by a third party.
[1][2]The fair case
The fair reading starts from a split the data actually show. Clients may be asking for AI in the brief while still buying the implementation, production, and delivery that using ChatGPT does not replace. Heavy ChatGPT use for writing can sit alongside rising demand for video, integration, and bundled delivery. A freelancer who can finish the job, rather than only generate a draft, is closer to what many postings are asking for.
The other half of the evidence should not be moved to a footnote. Some writing categories really did lose share, and content writing lost share at more than twice the overall rate after being heavily represented in ChatGPT use. If a freelancer's work looks like the text people already take to ChatGPT, this marketplace has already shifted against that work, at least as a share of postings. Rebundling and substitution can happen at the same time, in different categories.
[1][2]Who should care, and what to watch
Writers who sell draft-like work on Upwork should care, as should freelancers who already combine production, integration, or delivery with AI tools, and clients rewriting briefs. Readers of headlines that treat AI as the end of freelance demand should care too. The report does not show a uniform collapse, and it does not show that demand is safe.
In the next six months, the tests are whether content-writing's share of postings keeps falling, whether the gaps in fixed-price budgets and hired hours persist as more tools ship, and whether later data from Upwork or another marketplace still look like rebundling, or start to show a demand drop this window does not capture.
[1][2]Close
On this marketplace, in this window, the main pattern is a rebundling of work: a wider mix of AI-related tasks, less familiar pairings, larger project budgets, and more hours, alongside real share losses in some writing categories. That is not a finding about the whole labor market, and it is not a forecast that demand will hold. It is a claim that can be checked against later Upwork vintages and other platforms over the next six months.
[1][2]