The Rebundling of Freelance Work in the AI Age
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "The Rebundling of Freelance Work in the AI Age".
Warren: Here's the gist of the paper. This report studies how freelance work is changing in the AI era by comparing direct AI use with client requests on Upwork and examining how clients combine tasks into projects.
Ruth: The findings suggest that AI is changing not only which activities people undertake but also how clients assemble work into projects.
Warren: Activity Mix Divergence AI use and freelance demand have different activity mixes, with ChatGPT activities skewing toward writing and advice, while Upwork postings lean toward technical implementation and visual production.
Ruth: This pattern is consistent with businesses continuing to hire freelancers for work that requires implementation, production, and delivery.
Title and authors: Ruth: We’re talking about "The Rebundling of Freelance Work in the AI Age" and who wrote it. The authors are Chin, Kala, Kunievsky, Liu, Martin, and Shan.
Warren: They come from different spots—economics, quantitative trading—which is interesting because this paper tries to bridge that gap between market demand and actual task organization.
Iris: And what they do here is compare the direct use of AI with client requests on Upwork to see how the mix changes.
Ruth: It’s not just about listing jobs; it’s about seeing how clients bundle those tasks into projects, which is a different kind of view than just looking at job postings on their own.
Warren: It shows that the labor market is adjusting in ways we haven't fully mapped out yet because of this combination of direct use and client demand.
The paper's summary: Ruth: So, what’s the actual finding? The paper shows that there are different activity mixes between using AI yourself and hiring freelancers.
Warren: ChatGPT use is heavily focused on writing and giving advice, while Upwork postings lean much more toward things like technical implementation and visual production.
Iris: And this pattern suggests that businesses are still hiring people for the work that actually requires putting things together, producing deliverables, and getting them done.
Ruth: It points to a continuing need for workers who handle those kinds of implementation and final production steps, even with AI tools available.
The paper's improvements: Warren: Now, the authors suggest some ways to improve the research itself. They point out that AI systems can be designed to do more than just focusing on a few dominant applications like chatbot development or just text generation.
Iris: And they suggest optimizing systems to focus on delivering finished products and those applied business workflows, like integrating AI into existing software or making videos with AI.
Ruth: The paper even suggests engineering systems that specifically look for task bundles that have less historical precedent before the generative AI era, even if a project has the same total number of tasks as another one.
Warren: And they also suggest managing budgets and hours based on the scope of the project rather than just charging one fixed hourly rate.
Conclusion: Ruth: So to wrap up, this paper on "The Rebundling of Freelance Work in the AI Age" shows that AI isn't just changing what tasks we do, but fundamentally how clients assemble those tasks into a final piece of work.
Warren: The implication is that value might increasingly be found in defining the problem, connecting different tools and activities together, applying your own judgment to that combination, and delivering a finished outcome.
Iris: That means project boundaries can redraw themselves as clients put different capabilities together to meet their goals.
Ruth: It’s a shift from just being an implementer of one task to being an orchestrator of multiple tasks through AI assistance.
Warren: Makes you wonder what the next big thing in this area is, and where we should look next for more evidence on this kind of work organization.
Caroline Chin, Shagun Kala, Nadav Kunievsky, Teng Liu, Alex Martin Richmond, Carl Shan
OpenAI · Upwork
econ.GN, cs.AI, cs.CY, q-fin.EC
Submitted: 2026-11-10
Updated: 2026-11-10
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: This report studies how freelance work is changing in the AI era by comparing direct AI use with client requests on Upwork and examining how clients combine tasks into projects.
Key concepts
- Activity Mix Divergence
- This refers to the difference between what people do directly with AI (like writing) and what clients hire freelancers for (like technical implementation). AI use is concentrated in certain activities, while freelance demand leans toward production and delivery.
- Project Budget Drivers
- AI-related job postings show significantly larger fixed-price budgets compared to non-AI work. This suggests that the value of AI projects is tied to delivering broader, more complex outcomes rather than just charging a higher rate for individual hours spent on tasks.
- Novel Task Bundles
- AI jobs combine tasks in ways that are less common than before. Clients are increasingly seeking freelancers who can use AI alongside traditional skills to create new combinations of activities, such as combining AI video generation with editing.
Terminology
Summary
This report studies how freelance work is changing in the AI era by comparing direct AI use with client requests on Upwork and examining how clients combine tasks into projects. p. 0 THE REBUNDLING OF FREELANCE WORK
IN THE AI AGE
Caroline Chin1 Shagun Kala2 Nadav Kunievsky2
Teng Liu2 Alex Martin Richmond1 Carl Shan1
1OpenAI 2Upwork
November 2026
Abstract
This report studies how freelance work is changing in the AI era. We compare the activities
people undertake in ChatGPT with the work clients request on Upwork and examine
how clients combine tasks into projects. AI-related postings also carry higher stated project budgets and hourly rates, even within
narrowly defined work categories. Together, the results provide marketplace evidence that AI is changing
not only which activities people undertake, but also how clients assemble work into projects.Where the paper says it
Where the paper says it
Activity Mix Divergence
AI use and freelance demand have different activity mixes, with ChatGPT activities skewing toward writing and advice, while Upwork postings lean toward technical implementation and visual production. This pattern is consistent with businesses continuing to hire freelancers for work that requires implementation, production, and delivery. p. 0 ChatGPT activities skew toward
writing and advice, while Upwork postings lean toward technical implementation and visual
production. This pattern is consistent with businesses continuing to hire freelancers for work that
requires implementation, production, and delivery. ChatGPT activities skew toward
writing and advice, while Upwork postings lean toward technical implementation and visual
production. This pattern is consistent with businesses continuing to hire freelancers for work that
requires implementation, production, and delivery. There is not a
simple substitution pattern between Upwork tasks and those common on ChatGPT. This contrast
suggests a continuing role for workers who provide technical implementation, creative production,
and finished deliverables. There is not a
simple substitution pattern between Upwork tasks and those common on ChatGPT. This contrast
suggests a continuing role for workers who provide technical implementation, creative production,
and finished deliverables.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Project Budget and Scope
AI-related postings carry larger project budgets, with fixed-price budgets being approximately 52 percent higher in AI-related postings than for non-AI work after accounting for work subcategory and posting month. This suggests that the value of AI-related freelance work may lie partly in delivering broader projects that combine many capabilities. p. 0 Posted fixed-price budgets are approximately 52
percent higher in AI-related postings than for non-AI work after accounting for work subcategory
and posting month. The value of AI-related freelance work may lie partly in delivering broader
projects that combine many capabilities. Hourly contracts allow us to distinguish between the posted price of worker
time and the quantity of labor purchased. While adjusted differences in posted hourly rates are small,
AI-related work involves substantially more hours per hired contract.Where the paper says it
Where the paper says it
Task Reorganization
AI-related work is becoming more diverse, as the average share of historically uncommon task pairs in AI-related postings is higher among projects with at least two tasks. Work within AI-related postings is becoming less concentrated in a few activities, and among postings with at least two tasks, the average share of historically uncommon task pairs is higher for AI-related work. p. 0 Work within AI-related postings is becoming less
concentrated in a few activities. Among postings with at least two tasks, the average share of
historically uncommon task pairs is higher for AI-related work. Most of the tasks that gained share within AI-related postings describe a finished product, such as
a video, or a business function, such as managing a marketing campaign. This suggests that clients
increasingly seek freelancers who can use AI as one part of delivering a broader project.Where the paper says it
Where the paper says it
Novel Task Bundles
AI-related jobs combine tasks in less historically familiar ways, with historically uncommon task pairings accounting for an average of 8.4 percent of pairs within AI-related jobs, compared with 2. p. 2 AI-related jobs contain slightly more activities on average, but they combine those activities in novel
ways—even when compared with non-AI jobs of similar breadth in the same work subcategory and
month. AI-related postings also carry higher stated project budgets and hourly rates, even within
narrowly defined work categories. AI-related jobs contain slightly more activities on average, but they combine those activities in novel
ways—even when compared with non-AI jobs of similar breadth in the same work subcategory and
month. AI-related postings also carry higher stated project budgets and hourly rates, even within
narrowly defined work categories. Work within AI-related postings is becoming less
concentrated in a few activities. Among postings with at least two tasks, the average share of
historically uncommon task pairs is higher for AI-related work. in production. AI can shift this tradeoff by changing the relative costs of performing particular tasks,
coordinating across them, and transferring context from one task or worker to another.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion on Work Organization
Upwork and ChatGPT do not represent the labor market as a whole, but together they offer an early view of a broader shift: AI is changing not just the tasks people perform but the way those tasks are divided, combined, and brought to market. Value may increasingly lie in defining the problem, connecting tools and activities, applying domain judgment, and delivering a finished outcome. p. 3 This contrast suggests a division between activities people frequently
undertake through a general-purpose AI interface and those for which clients seek specialized remote
workers. Direct ChatGPT use is more concentrated in text-based and advisory activities, while Upwork
postings are more concentrated in technical implementation, visual production, and work that often
requires specialized tools, domain expertise, or delivery of a finished product. Job
postings show which activities clients seek, how those activities are bundled into projects, and the
budgets or rates clients attach to them. Freelance markets may reveal these changes relatively quickly
because clients can readily post emerging work, purchase one component of a larger role, or seek
combinations of tasks that previously belonged to different specialists.Where the paper says it
Where the paper says it
References
Acemoglu, D., & Restrepo, P. p. 2 2 Related research and contribution
A longstanding literature argues that the labor-market effects of technological change are best understood
by examining tasks rather than occupations. One common idea in this literature is that technology can
substitute for workers in some activities while increasing the value of complementary human inputs,
and the creation of new tasks can offset some of the resulting displacement (Autor, 2015; Acemoglu and
Restrepo, 2019; Autor et al. 2 Related research and contribution
A longstanding literature argues that the labor-market effects of technological change are best understood
by examining tasks rather than occupations. One common idea in this literature is that technology can
substitute for workers in some activities while increasing the value of complementary human inputs,
and the creation of new tasks can offset some of the resulting displacement (Autor, 2015; Acemoglu and
Restrepo, 2019; Autor et al.Where the paper says it
Where the paper says it
Autor, D. H. (2015). Why are there still so many jobs? p. 3 We treat Upwork Marketplace
job postings in H1 2023 as a baseline for how demand for remote work was distributed across activities. ChatGPT provides a way to use AI to perform many activities, but some may still prefer to hire human
workers for many reasons, including if AI capabilities are lacking on their specific tasks. (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025).Where the paper says it
Where the paper says it
Autor, D., Chin, C., Salomons, A., & Seegmiller, B. (2024). New frontiers: The origins and content of new work. p. 1 Yet,
technological change can reshape tasks and workflows well before occupations disappear or new jobs
enter official statistics. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today.Where the paper says it
Becker, G. S., & Murphy, K. p. 2 (2026) find that AI improves
performance on tasks within its capabilities but can reduce accuracy on tasks outside this “jagged
technological frontier. ”
Recent theory emphasizes that jobs are endogenous bundles of tasks rather than fixed occupational
containers, with boundaries set by the tradeoff between specialization and coordination costs (Rosen,
1978; Becker and Murphy, 1992; Ocampo, 2022). Which
tasks are bundled into a job reflects, in part, a classic tradeoff in the organization of work. Rosen (1978)
emphasizes how substitution possibilities across different kinds of labor shape the division of labor,
while Becker and Murphy (1992) show that greater specialization must be balanced against the costs of
coordinating complementary activities across workers.Where the paper says it
Where the paper says it
Brynjolfsson, E. p. 2 (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025). (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025). First, we compare the activities people undertake directly with ChatGPT with the
work clients purchase from freelancers, and examine how the mix and stated value of Upwork demand
are changing. Second, we study how activities are organized within job postings—whether AI-related
projects contain broader task bundles and whether they combine tasks in ways that had less precedent
before the generative-AI era.Where the paper says it
Where the paper says it
Where the paper says it
Chandar, B. p. 8 This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009). This distinction between price and quantity margins is consistent with recent evidence that labormarket adjustment associated with AI can operate strongly through employment and hiring while wage
responses are smaller or more heterogeneous (Chandar and Klein Teeselink, 2026; Rinz, 2026; de Souza,
2025). This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009). This distinction between price and quantity margins is consistent with recent evidence that labormarket adjustment associated with AI can operate strongly through employment and hiring while wage
responses are smaller or more heterogeneous (Chandar and Klein Teeselink, 2026; Rinz, 2026; de Souza,
2025). Yet,
technological change can reshape tasks and workflows well before occupations disappear or new jobs
enter official statistics. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today. The first report in OpenAI Economic Research’s “Work at the Frontier” series provides evidence
of one way users are starting to broaden their work activity.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Dell’Acqua, F. p. 2 Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025). At the same time, the benefits depend on the task: Dell’Acqua et al. (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025).Where the paper says it
Where the paper says it
Demirer, M. p. 2 ”
Recent theory emphasizes that jobs are endogenous bundles of tasks rather than fixed occupational
containers, with boundaries set by the tradeoff between specialization and coordination costs (Rosen,
1978; Becker and Murphy, 1992; Ocampo, 2022). Because AI can change the costs of performing tasks,
coordinating across them, and transferring context between them, it can split and rebundle jobs even
when no task is fully automated (Demirer et al. (2026) find that AI improves
performance on tasks within its capabilities but can reduce accuracy on tasks outside this “jagged
technological frontier. ”
Recent theory emphasizes that jobs are endogenous bundles of tasks rather than fixed occupational
containers, with boundaries set by the tradeoff between specialization and coordination costs (Rosen,
1978; Becker and Murphy, 1992; Ocampo, 2022).Where the paper says it
Where the paper says it
de Souza, G. p. 8 This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009). This distinction between price and quantity margins is consistent with recent evidence that labormarket adjustment associated with AI can operate strongly through employment and hiring while wage
responses are smaller or more heterogeneous (Chandar and Klein Teeselink, 2026; Rinz, 2026; de Souza,
2025). 2 percent) combined lower ChatGPT intensity with losses in Upwork
marketplace share. The presence of activities in these bins suggests that AI may complement freelance
services in some activities, while changes in client needs, production processes, and specialized software
also shape which activities gain or lose marketplace share. Categories not shown are included in the totals used to calculate both sets of percentages. Change in share is the H1 2026 Upwork share minus the H1 2025 share, in percentage points.Where the paper says it
Where the paper says it
Where the paper says it
Eloundou, T. p. 2 Research on generative AI documents both broad technical exposure and uneven effects in practice. Eloundou et al. Yet,
technological change can reshape tasks and workflows well before occupations disappear or new jobs
enter official statistics. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today.Where the paper says it
Where the paper says it
Gans, J. S. (2026). Endogenous Task Bundling, Skills and Automation (Working Paper No. p. 2 ”
Recent theory emphasizes that jobs are endogenous bundles of tasks rather than fixed occupational
containers, with boundaries set by the tradeoff between specialization and coordination costs (Rosen,
1978; Becker and Murphy, 1992; Ocampo, 2022). Because AI can change the costs of performing tasks,
coordinating across them, and transferring context between them, it can split and rebundle jobs even
when no task is fully automated (Demirer et al.Where the paper says it
Garicano, L. p. 2 Because AI can change the costs of performing tasks,
coordinating across them, and transferring context between them, it can split and rebundle jobs even
when no task is fully automated (Demirer et al. , 2026; Gans, 2026), with effects that depend on how
easily tasks can be separated from the jobs they sit in (Garicano, Li, and Wu, 2026). When tasks form a weak
bundle, AI can pull one component out and narrow the remaining human job, whereas stronger task
complementarities can keep the existing bundle intact. This idea also builds on a broader task-based theory of occupational boundaries.Where the paper says it
Where the paper says it
Hui, X., Reshef, O. p. 2 Evidence from online labor markets points to both demand-side disruptions and worker adaptation. Hui, Reshef, and Zhou (2024) find declines in employment and earnings among freelancers in categories
more exposed to generative AI. Evidence from online labor markets points to both demand-side disruptions and worker adaptation. Hui, Reshef, and Zhou (2024) find declines in employment and earnings among freelancers in categories
more exposed to generative AI. Evidence from online labor markets points to both demand-side disruptions and worker adaptation. Hui, Reshef, and Zhou (2024) find declines in employment and earnings among freelancers in categories
more exposed to generative AI.Where the paper says it
Where the paper says it
Where the paper says it
Jevons, W. S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. p. 1 1 Introduction
1. 1 How AI is changing work
Public debate about AI and work often centers on a single question: how many jobs will AI replace?Where the paper says it
Noy, S., & Zhang, W. p. 2 (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025). (2024) estimate that large language models could affect tasks across a wide range of
occupations. Experimental and workplace studies find that generative AI can improve productivity
and output quality in professional writing and customer support, with especially large gains for less
experienced customer-support workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025).Where the paper says it
Where the paper says it
Ocampo, S. (2022). A task-based theory of occupations with multidimensional heterogeneity (Working Paper No. p. 3 This idea also builds on a broader task-based theory of occupational boundaries. In Ocampo (2022),
workers have discrete skill types while tasks vary continuously in their skill requirements, and occupations emerge endogenously as bundles of tasks assigned to workers.Where the paper says it
Rinz, K. (2026). The recent evolution of AI-related labor demand (Working Paper No. p. 8 This distinction between price and quantity margins is consistent with recent evidence that labormarket adjustment associated with AI can operate strongly through employment and hiring while wage
responses are smaller or more heterogeneous (Chandar and Klein Teeselink, 2026; Rinz, 2026; de Souza,
2025). The freelance setting makes this distinction particularly transparent because hourly contracts
separately reveal a requested price for worker time and the subsequent quantity of labor used.Where the paper says it
Rosen, S. (1978). Substitution and division of labour. p. 2 Which
tasks are bundled into a job reflects, in part, a classic tradeoff in the organization of work. Rosen (1978)
emphasizes how substitution possibilities across different kinds of labor shape the division of labor,
while Becker and Murphy (1992) show that greater specialization must be balanced against the costs of
coordinating complementary activities across workers.Where the paper says it
Sorrell, S. p. 8 For example, AI may increase the returns to experimentation and refinement,
encouraging clients to explore more alternatives and invest additional time in improving the quality of
the final output, or to having workers complete and “own” larger, more complex bundles of tasks. This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009). This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009). This distinction between price and quantity margins is consistent with recent evidence that labormarket adjustment associated with AI can operate strongly through employment and hiring while wage
responses are smaller or more heterogeneous (Chandar and Klein Teeselink, 2026; Rinz, 2026; de Souza,
2025).Where the paper says it
Where the paper says it
Yiu, S., Seamans, R., Raj, M., & Liu, T. (2026). Worker repositioning and technological change: Evidence from an online labor market. p. 2 Together, the results provide marketplace evidence that AI is changing
not only which activities people undertake, but also how clients assemble work into projects. 2 Related research and contribution
A longstanding literature argues that the labor-market effects of technological change are best understood
by examining tasks rather than occupations.Where the paper says it
--- Page 1 ---
THE REBUNDLING OF FREELANCE WORK IN THE AI AGE
Caroline Chin1 Shagun Kala2 Nadav Kunievsky2 Teng Liu2 Alex Martin Richmond1 Carl Shan1
November 2026
Abstract
This report studies how freelance work is changing in the AI era. We compare the activities people undertake in ChatGPT with the work clients request on Upwork and examine how clients combine tasks into projects. p. 0 THE REBUNDLING OF FREELANCE WORK
IN THE AI AGE
Caroline Chin1 Shagun Kala2 Nadav Kunievsky2
Teng Liu2 Alex Martin Richmond1 Carl Shan1
1OpenAI 2Upwork
November 2026
Abstract
This report studies how freelance work is changing in the AI era. We compare the activities
people undertake in ChatGPT with the work clients request on Upwork and examine
how clients combine tasks into projects. We compare the activities
people undertake in ChatGPT with the work clients request on Upwork and examine
how clients combine tasks into projects. AI-related postings request slightly more tasks on
average than other postings. AI-related postings request slightly more tasks on
average than other postings. Even among projects with similar numbers of tasks, AI-related
postings bring together tasks that clients less often requested together before the generative
AI era.Where the paper says it
Where the paper says it
Where the paper says it
Key findings
AI use and freelance demand have different activity mixes, as ChatGPT activities skew toward writing and advice, while Upwork postings lean toward technical implementation and visual production. This pattern is consistent with businesses continuing to hire freelancers for work that requires implementation, production, and delivery. The relationship between direct AI use and changes in freelance demand is not uniform, suggesting a continuing role for workers who provide technical implementation, creative production, and finished deliverables. AI-related work is becoming more diverse, as work within AI-related postings is becoming less concentrated in a few activities. p. 0 AI-related work is becoming more diverse. Work within AI-related postings is becoming less
concentrated in a few activities. Work within AI-related postings is becoming less
concentrated in a few activities. Among postings with at least two tasks, the average share of
historically uncommon task pairs is higher for AI-related work. AI-related postings carry larger project budgets. Posted fixed-price budgets are approximately 52
percent higher in AI-related postings than for non-AI work after accounting for work subcategory
and posting month.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
1 Introduction
1.1 How AI is changing work
Public debate about AI and work often centers on a single question: how many jobs will AI replace? Yet, technological change can reshape tasks and workflows well before occupations disappear or new jobs enter official statistics. p. 1 1 How AI is changing work
Public debate about AI and work often centers on a single question: how many jobs will AI replace? Yet,
technological change can reshape tasks and workflows well before occupations disappear or new jobs
enter official statistics. Yet,
technological change can reshape tasks and workflows well before occupations disappear or new jobs
enter official statistics. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today. Understanding AI’s labor-market effects holistically requires examining how the
content and organization of work are changing today. The first report in OpenAI Economic Research’s “Work at the Frontier” series provides evidence
of one way users are starting to broaden their work activity. It finds that 43. 5 percent of work-related
ChatGPT messages assigned to an occupation-specific activity involve tasks traditionally associated with
an occupation other than the user’s own. 5 percent of work-related
ChatGPT messages assigned to an occupation-specific activity involve tasks traditionally associated with
an occupation other than the user’s own. This “task crossover” suggests that AI is expanding the range
of activities people attempt beyond established occupational boundaries. This “task crossover” suggests that AI is expanding the range
of activities people attempt beyond established occupational boundaries. But ChatGPT data alone do
not show how demand for freelance work is changing, or how clients are combining different activities
to form projects.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Upwork Marketplace is a global platform that connects businesses with on-demand access to highly skilled independent talent, providing a demand-side view of this adjustment. p. 1 Upwork Marketplace is a global platform that connects businesses with on-demand access to highly
skilled independent talent. The platform data provide a demand-side view of this adjustment. The platform data provide a demand-side view of this adjustment. Job
postings show which activities clients seek, how those activities are bundled into projects, and the
budgets or rates clients attach to them. Job
postings show which activities clients seek, how those activities are bundled into projects, and the
budgets or rates clients attach to them. Freelance markets may reveal these changes relatively quickly
because clients can readily post emerging work, purchase one component of a larger role, or seek
combinations of tasks that previously belonged to different specialists.Where the paper says it
Where the paper says it
Where the paper says it
This report combines evidence from ChatGPT use and Upwork job postings to examine two dimensions of change. p. 1 We therefore interpret their
similarities and differences as descriptive evidence about the organization of work, with limits on what
they can establish about substitution between direct AI use and paid freelance services. This report combines evidence from ChatGPT use and Upwork job postings to examine two dimensions of change. This report combines evidence from ChatGPT use and Upwork job postings to examine two dimensions of change. First, we compare the activities people undertake directly with ChatGPT with the
work clients purchase from freelancers, and examine how the mix and stated value of Upwork demand
are changing. First, we compare the activities people undertake directly with ChatGPT with the
work clients purchase from freelancers, and examine how the mix and stated value of Upwork demand
are changing. Second, we study how activities are organized within job postings—whether AI-related
projects contain broader task bundles and whether they combine tasks in ways that had less precedent
before the generative-AI era.Where the paper says it
Where the paper says it
Where the paper says it
We find both similarities and differences in the activities people undertake with ChatGPT and those clients seek from freelancers. p. 1 Second, we study how activities are organized within job postings—whether AI-related
projects contain broader task bundles and whether they combine tasks in ways that had less precedent
before the generative-AI era. We find both similarities and differences in the activities people undertake with ChatGPT and those
clients seek from freelancers. We find both similarities and differences in the activities people undertake with ChatGPT and those
clients seek from freelancers. Within AI-related Upwork Marketplace postings, demand is spreading
across a broader range of activities and shifting toward integration (e. Within AI-related Upwork Marketplace postings, demand is spreading
across a broader range of activities and shifting toward integration (e. , integrating AI APIs into
applications), multimodal production (e. , integrating AI APIs into
applications), multimodal production (e. , creating AI-generated video and producing motion graphics),
2 and applied business workflows (e. , marketing automation and social-media campaign management).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
The content writing was particularly prominent in ChatGPT use, while freelance postings for it declined sharply, suggesting that people may use ChatGPT directly to complete drafts of blog posts or marketing copy that they would otherwise commission from a freelancer. Video & Animation illustrates the other side of this pattern, with lower AI intensity and postings that grew by about a third relative to the overall marketplace, where clients may continue to seek freelancers for video editing and animation production. p. 5 Video & Animation illustrates the other side of this pattern, with lower
AI intensity and postings that grew by about a third relative to the overall marketplace. Here, clients
may continue to seek freelancers for video editing and animation production. Here, clients
may continue to seek freelancers for video editing and animation production. Together, these examples
suggest how direct AI use may be shifting the boundary between work people perform themselves and
work they purchase from freelancers.Where the paper says it
Where the paper says it
Where the gains come from
The analysis of task composition in AI-related job postings reveals that during this early phase, demand was already shifting toward creative production and applied business workflows, with tasks such as video production, social-media campaign management, marketing automation, and AI API integration gaining share. Most of the tasks that gained share within AI-related postings describe a finished product or a business function, suggesting clients increasingly seek freelancers who can use AI as one part of delivering a broader project. p. 6 Most of the tasks that gained share within AI-related postings describe a finished product, such as
a video, or a business function, such as managing a marketing campaign. This suggests that clients
increasingly seek freelancers who can use AI as one part of delivering a broader project. Posted fixed-price budgets are approximately 52
percent higher in AI-related postings than for non-AI work after accounting for work subcategory
and posting month. The value of AI-related freelance work may lie partly in delivering broader
projects that combine many capabilities.Where the paper says it
Where the paper says it
Project Value Drivers
AI-related postings carry larger stated project budgets, as fixed-price budgets are approximately 52 percent higher in AI-related postings than for non-AI work after accounting for work subcategory and posting month. The large fixed-price budget gap, negligible hourly-rate gap, and substantially greater realized hours point to a common pattern: AI-related work is associated with greater demand for labor and larger projects, rather than primarily with a higher price for each hour of worker time. p. 8 Larger AI-related
budgets may therefore reflect larger or more intensive projects rather than simply a higher price for an
otherwise identical unit of labor. Taken together, the large fixed-price budget gap, negligible hourly-rate
gap, and substantially greater realized hours point to a common pattern: AI-related work is associated
with greater demand for labor and larger projects, rather than primarily with a higher price for each
hour of worker time. For example, AI may increase the returns to experimentation and refinement,
encouraging clients to explore more alternatives and invest additional time in improving the quality of
the final output, or to having workers complete and “own” larger, more complex bundles of tasks. This
mechanism is analogous to the rebound effect in the energy literature, where lower costs can encourage
greater use (Jevons, 1865; Sorrell, 2009).Where the paper says it
Where the paper says it
Task Bundling Novelty
AI-related jobs combine tasks in less historically familiar ways, as AI-related jobs contain combinations with lower historical prevalence, particularly pronounced in generative AI and creative production. This suggests that AI is becoming embedded in a wider range of freelance projects, often alongside traditional implementation and production work. p. 9 Thus, AI-related work is becoming less concentrated in a few dominant activities and more evenly
spread across activities. This broadening is consistent with AI becoming embedded in a wider range of
economic activities rather than remaining concentrated in a small number of early applications such as
chatbot development or text generation. , integrating AI APIs into
applications), multimodal production (e. , creating AI-generated video and producing motion graphics),
2Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
AI systems can be designed to leverage their capabilities across a broader range of economic activities, moving beyond concentration in few dominant applications like chatbot development or text generation. This allows for more versatile deployment across diverse client needs.
-
Systems can be optimized to focus on delivering finished products and applied business workflows, as these are areas showing increased demand from clients seeking freelancers, such as
integrating AI APIs into applications
orcreating AI-generated video and producing motion graphics.
-
The system can be engineered to handle projects that combine tasks in novel ways by explicitly seeking out task bundles that have
less historical precedent
before the generative-AI era. This involves prioritizing combinations of tasks, even among projects with similar numbers of total tasks, that clients are less often requested together previously. -
AI systems can be configured to manage project budgets and hours based on project scope rather than just a single hourly rate. This reflects the finding that
larger AI-related budgets may therefore reflect larger or more intensive projects rather than simply a higher price for an otherwise identical unit of labor.
-
The system can prioritize activities that require
technical implementation, visual production, and delivery,
as these are areas where Upwork postings lean, contrasting with the ChatGPT use which is skewed towardwriting and advice.
Abstract
This paper examines how freelance work is changing in the AI era. We compare the activities people undertake in ChatGPT with the work clients request on Upwork, an online freelance platform, and examine how clients combine tasks into projects. We find that activities more prevalent in ChatGPT tend to lose marketplace share on Upwork, while those less prevalent tend to gain share. Within AI-related work, demand is becoming more diverse, and job postings combine tasks that were rarely requested together before the generative AI era. AI-related postings also have substantially larger project budgets and involve more hours of work, while differences in hourly rates are small. Together, these findings suggest that AI is reshaping not only which tasks clients outsource, but also how they combine tasks into larger, higher-value projects.
Related papers
- Huge If True: Learning from the Size of the Surprise
- Dynamic investment and coordination in teams: Theory and experimental evidence
- Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing
- Generative AI at Work
- What is Work For? Labor, Politics, and Meaning in the Age of AI
- Decision Models as Scalable Economic Surrogates that Capture Human Variation