Generative AI's Uneven Impact: Beyond Exposure in the Freelance Market

Who Loses Work First to Generative AI? Evidence from Freelance Markets

Predictions that generative AI will reshape labor markets are now commonplace. A growing body of research also ranks occupations by “AI exposure,” estimating how closely their tasks overlap with what large language models can perform. Writing and translation tend to rank near the top.

But exposure is not the same as actual job loss.

A technology must be accessible, usable, integrated into workflows, and accepted by clients before its technical capabilities become visible in market transactions.

The 2026 paper Generative AI and the Uneven Disruption of Freelance Labor by Zeljko Tekic and colleagues examines precisely this gap. Rather than asking only what AI is capable of doing, the authors ask when those capabilities begin to alter real economic activity.

They analyze 24 months of transaction data from Profi.ru, a major Russian services marketplace, covering 11,894 individuals between February 2022 and January 2024. Six freelance categories are examined: word processing, translation, design, marketing, programming, and web development. Offline service providers on the same platform provide a comparison group.

Between “AI Exposure” and Disappearing Work

Much of the existing literature begins by measuring how strongly the tasks within an occupation overlap with AI capabilities. Writing and translation are highly exposed because language models can perform substantial portions of these tasks.

Yet such measures say little about when substitution will occur.

The authors therefore bring diffusion theory into the discussion. Technology adoption depends on access, compatibility with existing routines, learning costs, complementary skills, and the ability to integrate new tools into professional workflows.

Russia provides an unusual setting for testing this argument. During the study period, official access to ChatGPT was restricted. VPN access existed, but reliable professional use and API integration were harder. Domestic systems such as YandexGPT and GigaChat were emerging, but were still developing. Restrictions on international payments and cross-border contracting also changed the structure of the freelance economy.

Russia therefore represents a market in which awareness of generative AI expanded rapidly while access and professional adoption remained comparatively constrained.

Writing and Translation Contracted First

The results show a clear divide.

Using February 2023 as the main diffusion anchor, monthly participation fell by 6.7 percentage points for word-processing freelancers relative to the control group. Translation declined by 8.7 percentage points, while design declined by 4.4 percentage points.

Orders and income moved in the same direction.

Programming, marketing, and web development, however, showed no statistically significant contraction during the observation period.

This produces one of the paper's most interesting contradictions.

Marketing has the highest AI Occupational Exposure score in the sample, at 1.356, yet no detectable contraction appears. Design has the lowest score, at 0.527, but experiences a significant decline.

The exposure measure used in the study is centered on language-model capabilities. It therefore captures the vulnerability of text work better than the growing capabilities of image-generation systems.

More importantly, occupational exposure does not capture the structure surrounding the task.

Marketing often involves repeated client interaction, local knowledge, revisions, negotiation, and strategic judgment. Programming and web development require testing, integration, maintenance, and adaptation to client-specific systems. AI can perform parts of these jobs without removing the professional who connects those parts.

For standardized text work, the distance between client and AI can be much shorter. A client who once hired someone to rewrite, summarize, or translate a document may now perform the first version directly with an AI system.

The Same Technological Shock Arrives at Different Speeds

The timing of disruption differs across sectors.

Word processing shows something close to a level shift around the diffusion period. Activity falls and remains below its previous trajectory.

Design behaves differently. Its decline accumulates over time. The event-study results show increasingly negative order effects during the months following the diffusion anchor, while the ramp model identifies a significant downward trend rather than a single abrupt break.

This temporal difference matters.

The moment when an AI system becomes technically capable of performing a task does not necessarily coincide with the moment when market demand begins to disappear.

Interface quality, price, reliability, user skills, workflow integration, software compatibility, and output quality can create a substantial delay between technical capability and economic substitution.

Text-generation tools reached useful professional applications quickly. Design depended more heavily on the gradual improvement and diffusion of image-generation systems.

Design Lost Orders, but the Remaining Orders Became More Valuable

Design produces another unusual result.

Participation, orders, and overall monthly activity declined. Yet among months in which transactions still occurred, average order value increased by 0.111 log points. Writing and translation showed no comparable increase.

The authors are careful not to interpret this as proof that individual designers successfully moved into higher-value work.

Several other mechanisms could generate the same pattern. Lower-priced providers may have become inactive. Clients may have shifted from multiple small jobs to fewer large projects. The composition of freelancers receiving orders may also have changed.

Still, the pattern suggests an important possibility.

If generative image systems allow clients to produce basic drafts, variants, template adaptations, and simple visual outputs themselves, the work that remains for professional designers may increasingly involve interpretation, iteration, branding, coordination, and more complex deliverables.

The market can therefore contract while the average value of the surviving transactions rises.

Text services may offer less room for this kind of upward migration when the original deliverable itself is standardized.

AI Disruption Depends on Exposure, Diffusion, and the Structure of Work

The paper ultimately moves away from the idea that occupations can be placed on a simple ladder of AI vulnerability.

Three forces interact.

First, the task must overlap with what AI can perform. Second, the technology must diffuse far enough to become practically usable. Third, the surrounding structure of the work determines whether AI can replace the worker or merely become another tool inside the worker's process.

This also means that the relative resilience of programming should not be read as permanent protection. The study only shows that no robust contraction was detected through January 2024. If AI systems increasingly absorb testing, debugging, deployment, and systems integration, the boundary could shift.

Likewise, declining design activity does not demonstrate the disappearance of design as a profession. The more important question is which design tasks disappear first and which become more valuable.

The authors also acknowledge a serious limitation. The observation period overlaps with sanctions, international payment restrictions, migration of Russian technology workers, and broader geopolitical disruption. The study therefore cannot establish that generative AI directly caused every observed decline.

However, the pattern does not disappear when word processing, translation, and design are compared with other digital services such as programming, marketing, and web development. This makes a single economy-wide shock a less convincing explanation.

The question raised by the paper is therefore more precise than “Can AI do my job?”

It is:

Can the client complete the task with AI without passing through me?

And if so, what is the next part of the job for which the client will still pay?

That boundary, rather than AI capability alone, may determine how generative AI reorganizes freelance work.