hilighter Logohilighter Logo
EN
Get StartedSign In

ArticleFuture of work

AI Is Rebundling Freelance Work

OpenAI and Upwork, on 10 November 2026, show a shift in requested tasks and larger project budgets. That is not a count of jobs lost.

Christian Kosiek•Published •19 min read

A person in a workshop adjusts a miniature tree on a small film set beside a camera on a slider
Illustrative image supplied with this article: a person works on a miniature film set. It is not a named freelancer, an Upwork client, a study participant, or a hilighter screen. The file was not labelled as a photograph of a real person or as AI-generated. The scene does not depict a posting from the datasets below.

A client can change the unit of work it buys without sending a layoff note. The next brief may ask one person to script, generate, edit and finish a video that used to be three invoices. That is a change in the shape of the assignment. It is not, by itself, evidence that freelance jobs have disappeared.

A posting is an early signal, not a job count

On 10 November 2026 OpenAI published a page titled "Work at the Frontier: AI's effect on freelance work," with a research report, "The Rebundling of Freelance Work in the AI Age," written with Upwork. The authors, listed alphabetically, are Caroline Chin, Shagun Kala, Nadav Kunievsky, Teng Liu, Alex Martin Richmond and Carl Shan. Chin, Richmond and Shan are credited to OpenAI. Kala, Kunievsky and Liu are credited to Upwork. The PDF is dated November 2026.12

The page says freelance contracts turn over quickly, so a posting can show a change before employment statistics do. The report is explicit about the limit. ChatGPT conversations and Upwork posts are not directly matched, and neither represents the labour market as a whole. Similarities and differences are descriptive evidence about how work is organized. They do not establish that ChatGPT use caused a change in paid freelance demand. A posting can go unfilled, state an aspiration, or bundle work that is later split. A line in an advertisement is not a hire. See why an AI line in a job ad is not an AI job.2

The OpenAI page also cites two Upwork studies that sit outside this report. Upwork's 2026 Future Workforce Index says skilled freelancing rose from 28% of skilled knowledge workers in 2025 to 38% in 2026. That index is a survey of 2,400 U.S.-based workers, fielded from March to April 2026, with a margin of error of 2% at 95% confidence, plus marketplace data. The 28% in the 23 April 2025 release is "more than one in four (28%) U.S. knowledge workers," from a survey of 3,000 people between December 2024 and February 2025, margin of error ±1.8%. The same release says those workers generated a collective $1.5 trillion in earnings in 2024, extrapolated with Bureau of Labor Statistics data on management, professional and related occupations. OpenAI says those estimates extend beyond Upwork's own marketplace. They are the vendor's surveys and an earnings extrapolation. They are not a count of contracts on the platform, and the two survey waves are not the same sample. This article does not treat 28 to 38 as a census.134

Forty-three categories fit one pattern. Twenty do not tell one story.

The report matches 63 Upwork work subcategories. AI intensity is the natural log of 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 other axis is the change in that category's share of Upwork postings from the first half of 2023 to the second half of 2025. Categories with no posts in the first half of 2023, and the category "AI Apps & Integration," are left out. The figure does not measure total marketplace growth, and it does not identify a causal effect.2

Of the 63, 43 (68.3%) either had lower ChatGPT intensity and gained Upwork share, or had higher intensity and lost share. The public page says the other 20 "showed the reverse." The report splits those 20. Six categories (9.5%) combined higher ChatGPT intensity with gains in Upwork share. Fourteen (22.2%) combined lower intensity with losses. "Reverse" on the page covers both bins. It does not mean all 20 moved in one opposite direction. The report says a simple substitution story explains only part of the pattern. Client needs, production processes and specialized software can move share as well. This article follows the report's split.12

What the pages support What they do not support
43 of 63 matched categories lost Upwork share where ChatGPT use was relatively common, or gained share where it was not That ChatGPT caused those moves
6 categories gained Upwork share despite higher ChatGPT intensity; 14 lost share despite lower intensity That the other 20 are one uniform "reverse" result
Postings show requested work on one marketplace, across windows from 2023 into 2026 A November 2026 count of completed job losses, hires or individual earnings

The activity chart uses a different window. ChatGPT shares come from 36,896 conversations classified as personal or work-related, sampled from January 2025 through June 2026. Each conversation was classified with GPT 5.6 Luna from up to its first five user messages. The report says no researcher read the messages. Schoolwork, unclear purpose and conversations without an Upwork match are excluded. The design targeted 5,556 conversations a month before those exclusions. Upwork bars in that chart are each category's share of all Marketplace posts in the first half of 2026. The change plotted there is the first half of 2026 minus the first half of 2025, in percentage points, for the 15 largest gains and the 15 largest declines.2

Content writing was 14.5% of those sampled ChatGPT conversations, and professional and business writing was 10.2%, against about 1.3% and 0.4% of Upwork posts. A separate sentence says content writing 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, and that content-writing postings fell at more than twice the rate of Upwork postings overall across that two-year window. Video and animation, with lower AI intensity, grew by about a third relative to the overall marketplace. Those are the report's examples. They are not a ranking of every occupation, and the 14.5% figure is not the "three times" sentence.2

The public page adds a limit the chart does not remove. The comparison captures how people currently use ChatGPT, not every task a model can support. OpenAI calls that gap the capability overhang.1

The windows are easy to mash together. The scatter of category share runs from 2023 into late 2025. The activity-chart change runs from 2025 to 2026. The task mix inside AI-related posts compares calendar 2023 with calendar 2025. Prices and hours run through April 2026. None of this is a November 2026 wave of completed job losses. The publication is new. Much of the observed behaviour is earlier.12

A 45% wider mix is not 45% more work for one person

Between early 2023 (the first and second quarters) and late 2025 (the third and fourth), the effective number of task clusters in AI-related postings rose from 14.6 to 21.2, about 45%. Effective number is the inverse of a concentration index of task-cluster shares. A value of 21.2 means the mix is as concentrated as about 21 equally represented clusters. It describes the set of postings. It is not evidence that each freelancer performs 45% more tasks, or that every project became 45% harder.2

Inside one posting the gap is smaller. AI-related posts contain 3.35 distinct task clusters on average, against 3.19 for other posts, a difference of about 5%. AI posts are less likely to contain one or two clusters and more likely to contain three or four. The report says there is no matching rise in the five-or-more group. "About as many tasks" on the public page is this gap, not the 45% figure.12

The public example is a brief that asks for AI-generated video, editing and motion graphics together. The report uses that illustration, and also chatbot or machine-learning development combined with computer vision and API integration. A first visual from a model does not remove continuity, pacing, rights, client communication or the final check. Those checks are an editorial reading of what a finished video still requires. The study did not measure demand for each of them.12

Within AI-related posts, "build chatbots and machine-learning models" stayed the largest task cluster and fell from 21.8% of task-cluster assignments in 2023 to 17.8% in 2025. It is left off the chart because of its scale. Video production, social-media campaign management, marketing automation and AI API integration gained share inside AI-related work. AI-generated written content and several writing and data activities lost share inside that same set. Shares are of assignments in AI-related posts, clusters can overlap, and the figure describes composition, not total marketplace volume.2

Among posts with at least two tasks, historically uncommon pairs (seen together in fewer than 200 pre-AI postings, including pairs never seen) account for an average of 8.4% of pairs inside AI-related jobs and 2.4% inside other jobs. Each job is weighted equally. That comparison uses 104,066 posts. Of the uncommon pair occurrences in current AI-related jobs, 98.6% join two tasks that each appeared in at least 200 pre-period postings. The novelty is mostly familiar tasks requested together less often, not two brand-new tasks. Under both a curated list and a usage-based rule, mixed AI-and-traditional pairs are the largest slice of those uncommon occurrences: about four per 100 total pair occurrences. The appendix says this should not be read as proof that AI has broadly caused traditional tasks to be recombined with one another. The 8.4% and the table's overall rate differ because one averages within each job and the other pools pairs, so larger bundles weigh more.2

After controls for bundle size, work subcategory and month, on 104,030 posts in a 5% sample, AI-related jobs still show pairings with lower historical prevalence. Among jobs with two or three tasks, pairings had roughly 40% lower geometric mean historical prevalence. Generative AI and creative production had about 60% lower prevalence than comparable non-AI jobs. These are conditional associations. The score drops pairs seen fewer than 20 times before the generative-AI period, so entirely new pairs are outside that regression.2

A larger budget is not a higher hourly wage

Fixed-price budgets for AI-related posts are about 70% higher overall and about 52% higher after controlling for work subcategory and posting month. Posted hourly rates, using the midpoint of the posted range, are about 9% higher before that adjustment and about 1.5% higher after it. Among hired hourly contracts with positive recorded hours, hours in the first 90 days are about 41% higher for AI-related work, with the same subcategory and month controls. The top 1% of posted prices is trimmed. The comparisons run through April 2026. They do not identify a causal effect of the AI label.2

The report's introduction also says AI-related postings carry higher stated hourly rates. The exhibit is the tighter statement: after the adjustment, the hourly gap is about 1.5%. This article uses the exhibit. A larger fixed-price budget can be a wider scope. It is not take-home pay, and it is not a stable pipeline for any one person. The hours figure covers hired contracts with positive hours. It is not hours for every posting, and it is not unpaid revision time.2

Upwork's own index states a different set of marketplace figures, and they should not be averaged with the November report. The page says freelancers performing AI work earn 34% more per hour than those not incorporating AI, across every work category. It does not attach that sentence to a named quarter. Separately, it says platform comparisons are the first quarter of 2026 against the same period a year earlier. In that comparison, generative AI and creative production saw 90% growth in contract starts and a 13% decline in earnings per contract. AI-based execution tasks saw earnings fall 28%. AI-augmented professional services grew 72%, and earnings rose 22%. Freelancers doing more complex work with AI saw earnings increase 45% in the first quarter of 2026. An LLM pipeline classifies the jobs, volume and earnings are spread across tasks with relevance weights, and the page says the methods may contain measurement error. A premium per hour, a larger posted budget and a fall in earnings per contract can all be true of different cuts. They are not one number.23

A charity that once hired separate people for copy, images and distribution might now ask one freelancer for a package and expect drafts from a model. The charity might get one campaign. The contractor might inherit revisions, rights checks and coordination. Whether the fee is better depends on hours, tool cost and what the fixed price includes. That scene is an illustration. The study did not observe it.

What a smaller buyer can write down

The steps below are an editorial suggestion. OpenAI and Upwork did not test them. They are not a research result.

Write the result and the acceptance checks. Name who verifies facts, who may see confidential data, who owns the rights and who approves the file. A list of model names is not that brief.

Compare one multi-skilled contractor with specialists against the same quality bar. One person can reduce handoffs. One person can also be a single point of failure. The report does not say which structure wins.

If junior employees used to learn on first drafts, do not assume a purchased package replaces that practice. Redesign the assignment so someone still checks, explains and escalates. Buying an output can be efficient this month and thin next year's internal skill. The longer arguments are who teaches judgment after the first draft and why tasks can move inside a familiar job title. Who had paid time to practise, as distinct from a software login, is a separate problem. See why a training gap becomes a career gap.

For a European organization, a video brief is not a finding that the purchase is high-risk. This is not legal advice. Article 4, in the Commission service-desk text of the consolidated AI Act as at 27 July 2026, says providers and deployers shall take measures to support the AI literacy of their staff and of other persons dealing with the operation and use of systems on their behalf. The same paragraph says this does not require a guaranteed level for any individual. The service desk says the summary is not legally binding. A literacy measure is not a duty to hire a generalist.5

Annex III is headed "Employment, workers' management and access to self-employment." Point 4(a) lists systems intended for recruitment or selection, including targeted job advertisements, filtering applications and evaluating candidates. Point 4(b) lists systems intended to decide terms of a work relationship, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. A brief that asks for an AI video is not on that list merely because it names a model. A tool intended to score people for access to work, or to assign tasks from personal traits, is a different question. Classification depends on the intended purpose. The service desk says Annex III rules apply from 2 December 2027. This article does not decide whether any marketplace ranking tool is in scope.6

Article 50 is the transparency chapter. The same FAQ says those obligations are enforceable from 2 August 2026, and that providers of systems placed on the market before that date have until 2 December 2026 to comply with the marking duty in paragraph 2. Paragraph 4 says a deployer of a system that generates or manipulates image, audio or video constituting a deep fake shall disclose that the content was artificially generated or manipulated. Evidently artistic, creative, satirical or fictional work has a narrower disclosure. A separate rule covers text published to inform the public on matters of public interest, and it does not apply where a person holds editorial responsibility after human review. Paragraph 2 is a provider duty to mark synthetic audio, image, video or text so it is machine-readable, with an exception where the system only assists standard editing or does not substantially alter the input. This article does not decide whether a particular video is a deep fake.67

What an independent worker can price

The commercial move below is an editorial suggestion, not a demand curve from the study. Price the finished outcome, and keep a record of the whole cycle: briefing, drafts, tool cost, revisions and the check. If a model makes one step faster, a client may expect a lower price or a wider deliverable. A posted hourly rate that barely moves, beside a fixed budget that is about half again as large after adjustment, is a reason to ask what extra scope sits inside the number.

A translator who only supplies a first pass is not in the same commercial position as one who also owns terminology, a regulatory read and the final text. The study did not measure those two offers. The distinction is the purchased unit. It is not a verdict that one occupation is safe.

Permanent hiring shows a related pattern. Some entry-level ads already ask for senior judgment. See entry-level ads that ask for senior judgment. Being able to attempt a task beyond one's training is not the same as mastery. See why a wider task is not mastery.

How hilighter can help

hilighter does not staff freelance projects, set fees or decide whether a brief complies with the AI Act. It can make the skills inside a role visible when a buyer starts purchasing a wider outcome. Teams can map skills by role, compare people with those requirements in skills and talent insights, record a check or a correction in skills evidence and recognition, and set personalized growth paths. A manager's review of that evidence sits in manager assessments. The AI job impact view is indicative. It is not an Upwork forecast, and it is not a prediction that a role will disappear. The Free plan covers up to three job roles. The number of employees is not capped.

Sources

  1. OpenAI, Work at the Frontier: AI's effect on freelance work, 10 November 2026. Public summary with Upwork. 63 categories, 43 following the stated pattern, and a line that the other 20 "showed the reverse," which the full report splits into 6 and 14. Diversity from about 15 to about 21 tasks, stated on the page as 14.6 to 21.2. AI-related projects described as having larger fixed-price budgets and more hours, with similar posted hourly rates. Example brief: AI-generated video, editing and motion graphics. The page cites Upwork's 2026 index for 38% of skilled U.S. knowledge workers freelancing, and the previous annual study for $1.5 trillion in U.S. earnings in 2024, and says those estimates extend beyond Upwork's marketplace. "Capability overhang" is the page's name for the gap between current ChatGPT use and tasks AI can support. Authors named on the page.
  2. Caroline Chin, Shagun Kala, Nadav Kunievsky, Teng Liu, Alex Martin Richmond and Carl Shan, The Rebundling of Freelance Work in the AI Age, OpenAI and Upwork, November 2026. Descriptive, not causal. 43 of 63 (68.3%), 6 (9.5%) and 14 (22.2%). ChatGPT sample of 36,896 personal or work-related conversations, January 2025 to June 2026, classified with GPT 5.6 Luna; researchers did not read messages. Content writing 14.5% and professional and business writing 10.2% of that sample, against about 1.3% and 0.4% of Upwork posts. Effective task clusters 14.6 to 21.2 (about 45%), Q1-Q2 2023 to Q3-Q4 2025. Tasks per post 3.35 versus 3.19. Chatbot and machine-learning cluster 21.8% to 17.8% of AI-related task assignments, 2023 to 2025. Uncommon pairs 8.4% versus 2.4% (104,066 posts); 98.6% of uncommon occurrences join two previously common tasks. Novelty regression on 104,030 posts in a 5% sample: about 40% lower historical prevalence for two- and three-task jobs, about 60% for generative AI and creative production. Fixed-price budgets about 70% higher raw and about 52% higher after subcategory and month controls; posted hourly rates about 9% and about 1.5%; hours in the first 90 days of hired hourly contracts with positive hours about 41% higher. Through April 2026. Top 1% of posted prices trimmed. The introduction's "higher hourly rates" line is not used in place of the exhibit.
  3. Upwork Research Institute, The Future Workforce Index 2026. Survey of 2,400 U.S.-based workers, March to April 2026, margin of error 2% at 95% confidence. Skilled freelancing from 28% of skilled knowledge workers in 2025 to 38% in 2026, on that page. Marketplace: freelancers performing AI work earn 34% more per hour than those not incorporating AI, across every work category; the sentence is not tied to a named quarter. Platform year-over-year figures are Q1 2026 versus the same period a year earlier: generative AI and creative production, 90% growth in contract starts and a 13% decline in per-contract earnings; AI-based execution earnings down 28%; AI-augmented professional services up 72% in volume with earnings up 22%; more complex AI work, earnings up 45% in Q1 2026. LLM classification, relevance weights, and an explicit measurement-error caveat. Vendor platform data and a vendor survey. Not the November report's 52% or 1.5%.
  4. Upwork, press release, 23 April 2025. Survey of 3,000 U.S.-based skilled knowledge workers, December 2024 to February 2025, margin of error ±1.8% at 95% confidence. "More than one in four (28%) U.S. knowledge workers" freelancing or working independently, and a collective $1.5 trillion in earnings in 2024, extrapolated with BLS data on management, professional and related occupations (42% of the U.S. workforce, 71.5 million workers, as the release states). A separate survey of 502 executives is not used here. Conflicting skill-proficiency percentages inside the release are not used.
  5. European Commission, AI Act Service Desk, Article 4: AI literacy. Wording reflects the EUR-Lex consolidated version as at 27 July 2026, including Digital Omnibus amendments. The summary is not legally binding. Paragraph 1 does not require a guaranteed level of literacy for any individual.
  6. European Commission, AI Act Service Desk, Annex III, heading and points 4(a) and 4(b). When does enforcement start?: Annex III rules from 2 December 2027. Transparency obligations enforceable from 2 August 2026, with providers of systems placed on the market before that date given until 2 December 2026 for the Article 50(2) marking duty. Not a decision about any marketplace ranking tool.
  7. European Commission, AI Act Service Desk, Article 50. Paragraph 2: provider marking of synthetic audio, image, video or text, with an editing exception. Paragraph 4: deployer disclosure for deep fakes, a narrower rule for evidently artistic or fictional work, and a separate rule for public-interest text that does not apply where a person holds editorial responsibility after human review. Not legal advice, and not a finding about a particular video.

Sources last reviewed: 15 November 2026. The public page's "reverse" line is not used in place of the report's 6-and-14 split. The 45% diversity figure is not a per-person workload. The November budget and hourly figures are not averaged with Upwork's July index. The $1.5 trillion figure is the vendor's 2025 extrapolation, not a platform cash total. PwC's Global Workforce Hopes and Fears Survey 2026 was listed in the inbox draft and was opened; it describes employees, not these postings, so its statistics are not used. EUR-Lex was not opened separately. EU wording follows the Commission service desk and is not legal advice. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.