OpenAI–Upwork study: AI-related freelance projects bundle more work and carry larger fixed budgets, but the data don't show job replacement

OpenAI and Upwork published the study on November 10, 2026, as part of OpenAI's Work at the Frontier series 1. The 24‑page report, 'The Rebundling of Freelance Work in the AI Age', is written by OpenAI and Upwork authors 2. It does not report a count of jobs lost. It reports four things:

• Some activities that are common in ChatGPT use lost share of Upwork postings.
• AI‑related postings spread across more kinds of work.
• Those postings combine tasks that clients rarely requested together before generative AI.
• AI‑related postings carry larger posted fixed‑price budgets and, for hired hourly contracts with positive recorded hours, more hours.

The authors say the findings 'do not point to a simple substitution story' 1. They also describe their comparisons as descriptive, and say the data cannot establish a causal effect of AI 2. The study comes from a company that sells ChatGPT and a marketplace that sells freelance work, and we have not independently reanalyzed the data.

The headline figures are easy to mix up because they use different postings, windows and contract types. Each item below is from the report 2.

• ChatGPT vs Upwork activity shares. Population: 36,896 sampled personal or work‑related conversations from consumer ChatGPT accounts, January 2025 to June 2026, against all Upwork Marketplace posts in H1 2026. Result: content writing was 14.5% of ChatGPT conversations vs about 1.3% of Upwork posts; professional and business writing was 10.2% vs about 0.4%.
• ChatGPT 'intensity' vs category share. Population: 63 matched Upwork subcategories. Intensity compares ChatGPT share in H2 2025 with Upwork share in H1 2023; share change runs H1 2023 to H2 2025. Result: 43 categories (68.3%) either had higher intensity and lost share or lower intensity and gained share. Six had higher intensity and gained share; 14 had lower intensity and lost share.
• Task‑mix concentration. Population: task‑cluster assignments in AI‑related Upwork posts, averaging Q1–Q2 2023 against Q3–Q4 2025. Result: the effective number of task clusters rose from 14.6 to 21.2, about 45%.
• Distinct tasks per post. Population: AI‑related vs other Upwork posts, through 2025. Result: 3.35 vs 3.19, about 5% more.
• Historically uncommon task pairs. Population: posts with at least two tasks (N = 104,066), through 2025. 'Uncommon' means fewer than 200 pre‑AI postings contained the pair. Result: the average within‑post share of such pairs was 8.4% in AI‑related posts vs 2.4% in others.
• Posted fixed‑price budget. Population: positive posted budgets, through April 2026, top 1% trimmed. Result: about 70% higher raw, and about 52% higher after controlling for work subcategory and posting month.
• Posted hourly rate. Population: midpoint of the posted hourly range, same window. Result: about 9% higher raw, and about 1.5% higher after the same controls.
• Hours worked. Population: hired hourly contracts with positive recorded hours, first 90 days after hire, through April 2026. Result: about 41% higher after controlling for work subcategory and calendar month.

Two definitions matter. 'AI‑related' is a classification by Upwork's internal LLM‑based pipeline, applied to each post as a whole. And 'effective number of task clusters' is 1 divided by a concentration index (HHI). A value of 21.2 means the mix is as spread out as about 21 equally common clusters. It does not mean 21 tasks per job; the per‑post average is 3.35.

The static comparison shows different activity mixes. ChatGPT conversations skew toward writing and advice, while Upwork postings lean toward video and animation, design, web and mobile development, QA testing and digital marketing [2](https://cdn.openai.com/pdf/openai‑upwork‑full‑report.pdf). The ChatGPT side is a sample of consumer‑account conversations, classified by an OpenAI model into Upwork's taxonomy. It is not a measure of what freelancers' clients do. The longer‑run comparison asks whether categories that were heavily used in ChatGPT lost share of Upwork postings. For 43 of 63 subcategories the pattern held in one direction or the other: higher ChatGPT intensity with lost share, or lower intensity with gained share. The authors single out content writing as a particularly prominent example. Its postings fell at more than twice the rate of Upwork postings overall between H1 2023 and H2 2025. Video and animation, with lower ChatGPT intensity, grew about a third relative to the overall marketplace. That is a pattern of share, not of jobs. The report says the figure 'does not measure aggregate Marketplace growth or identify a causal effect of ChatGPT use on Upwork demand', and it excludes the AI Apps and Integration category and categories with no H1 2023 postings. A category can lose share while its absolute postings hold steady or grow; because the exhibit is a relative‑share measure, it does not tell us which happened. Twenty of the 63 categories ran against the pattern. The report also cites earlier research that found employment and earnings declines for freelancers in more AI‑exposed categories, so its own framing leaves room for displacement in specific segments.

Within AI‑related postings, the authors find demand shifting toward integration, multimodal production and applied business workflows. The largest cluster, building chatbots and machine‑learning models, fell from 21.8% to 17.8% of task‑cluster assignments between 2023 and 2025. Tasks such as video production, social‑media campaign management, marketing automation and AI API integration gained share [2](https://cdn.openai.com/pdf/openai‑upwork‑full‑report.pdf). AI‑related posts contain only slightly more tasks on average (3.35 vs 3.19). The distribution shows fewer one- and two‑task posts and more three- and four‑task posts, with no increase at five or more. The more distinctive result is the mix. Among AI‑related posts, 8.4% of task pairs were historically uncommon, against 2.4% elsewhere. Among pairs found in at least 200 current postings, AI‑related posts account for 54% of appearances of historically uncommon pairs, versus 13% for established pairs. A separate regression scores novelty only for task pairs observed in at least 20 pre‑AI posts; entirely new pairings and rarer pairs are excluded from that score. After controlling for bundle size, subcategory and posting month, AI‑related posts with two or three tasks had pairings with roughly 40% lower geometric‑mean historical prevalence. The gap was about 60% in the generative AI and creative production group. The appendix says where the novelty comes from. Of the uncommon pair occurrences in AI‑related jobs, 98.6% join two tasks that were each already common in pre‑period postings; they were just rarely requested together. Mixed pairs, one AI‑associated task and one traditional task, are the largest component: about 4 uncommon pairs per 100 total pair occurrences, against roughly 1 to 2 for each other pair type. The report cautions that this should not be read as showing AI has broadly caused traditional tasks to be recombined with one another. An example from the announcement is a single posting that asks for AI‑generated video, editing and motion graphics, so one freelancer delivers a finished video [1](https://openai.com/index/ai‑effect‑on‑freelance‑work/).

The 52% figure is the one most likely to be misread. It is the adjusted difference in posted fixed‑price project budgets between AI‑related and other postings in the same subcategory and posting month [2](https://cdn.openai.com/pdf/openai‑upwork‑full‑report.pdf). It is not realized pay, earnings or a wage premium, and it says nothing about whether total demand rose. A fixed‑price budget bundles value, scope and the labor needed into one number, so a larger budget could reflect a larger project rather than a higher price for the same work. The hourly data offer a partial check, not a decomposition. Posted hourly rates, which price worker time, differ by about 1.5% after adjustment, so the posted price of an hour is close across the two groups. Hours worked in the first 90 days are about 41% higher for AI‑related work, but that comparison covers only hired hourly contracts with positive recorded hours. The authors conclude that AI‑related work is associated with more labor and larger projects rather than a higher price per hour, and the announcement says the higher budgets 'may partly reflect' broader scope [1](https://openai.com/index/ai‑effect‑on‑freelance‑work/). Using hourly‑contract evidence to interpret the fixed‑price gap is an inference: the two contract types are different postings, and hours are observed only for hired contracts with recorded hours. The report offers possible mechanisms, such as clients experimenting more or workers owning larger bundles, but does not test them. The 1.5% rate gap and the 41% hours gap come from different eligible populations (posted hourly ranges versus hired contracts with positive recorded hours), so they should be read separately. They should not be multiplied into a cost or pay estimate, and the report does not offer one. Nor does it claim the hourly and fixed‑price gaps should match. All three are associations from OLS regressions with subcategory and month controls. The report states they do not identify a causal effect of the AI‑related label, and that label itself comes from an LLM classifier.

For someone pricing or scoping AI‑related freelance work, a reasonable reading is to compare scope before comparing budgets. A fixed‑price budget well above a non‑AI comparable may partly reflect extra deliverables, such as AI‑generated video combined with editing and motion graphics, but the study does not show how much. Because the adjusted hourly‑rate gap is small and hours are higher on hired hourly contracts with recorded hours, the authors infer that scope matters more than the price of an hour; that is an inference for fixed‑price work, not a direct measurement. For anyone hiring, a high fixed‑price posting is not evidence of a higher hourly price in this data. For anyone reading headlines, 'larger budgets' and 'losing share' are measured on different populations and cannot be added together.

Limits the report or announcement states or implies:

• Upwork and ChatGPT data are not directly matched and do not represent the labor market as a whole 2.
• Work inside firms can change without a new job posting, so postings capture only part of the change 1.
• Time windows differ across exhibits: H1 2023 to H2 2025 for category shares, calendar 2023 versus 2025 for task mix, H1 2026 for the static comparison, and through April 2026 for prices and hours.
• The task‑novelty analysis stops at the end of 2025, and the authors describe the results as an early baseline.

The study supports a narrower claim than job replacement or job creation: on this marketplace, AI‑related projects look broader and differently assembled, and they are associated with larger posted fixed‑price budgets and, separately, with more hours on hired hourly contracts with recorded hours, at similar posted hourly rates. The banner image is a June 15, 2025 photograph of the building at 1515 Third Street in San Francisco, credited to Coolcaesar under CC BY 4.0 and used for archival context on OpenAI. It does not show study participants, freelancers or results.