Makridis’ paper, The Labor Market Effect of Generative Artificial Intelligence on Artists, finds little evidence of short-run earnings declines among more LLM-exposed artistic occupations through 2023. Yet its employment result is explicitly mixed: in some specifications, more exposed artistic occupations had weaker employment growth in 2023. The study is evidence against an immediate economy-wide collapse in creative employment, not proof that generative AI has left creative labor markets untouched.
That difference matters for IT leaders and managers evaluating whether AI is augmenting a team or quietly reducing its entry-level pipeline. A wage series can look stable while available assignments, junior hiring, freelance demand, and the mix of work all shift sharply underneath it.
The paper finds no immediate wage collapse, not no labor-market effect
The ASU story draws on Makridis’ analysis of U.S. Bureau of Labor Statistics Occupational Employment and Wage Statistics data and Census Bureau American Community Survey data, comparing artistic occupations with differing estimated exposure to large language models after ChatGPT’s November 2022 release. The research also uses Gallup Workplace Panel data to examine reported AI use and worker outcomes.
The basic result is more restrained than the university’s framing. Across the datasets, Makridis finds earnings estimates near zero for the early post-ChatGPT period, with some models showing modestly positive effects. But the statistical uncertainty is large enough that he does not present the earnings result as a firm wage premium from AI adoption.
The employment evidence deserves equal weight. The paper reports weaker employment growth in 2023 for more exposed artistic occupations in some analyses. That is neither a clean demonstration of broad job destruction nor a result that can be safely condensed into “jobs stayed.” It is an early indication that occupational change may appear in hiring and employment growth before it appears in average wage data.
This is a familiar measurement problem in technology transitions. If a company stops filling junior design, copywriting, support, QA, or development roles while retaining senior staff, occupational wage averages may remain steady—or even rise—because the lower-paid jobs disappear from the sample. Stable average pay is therefore a poor standalone test for whether the number and accessibility of career opportunities are holding up.
Exposure to an LLM is not the same as a task being automated
Makridis’ most useful point is that occupations are bundles of tasks rather than single activities. A designer may ideate, create assets, revise work with clients, maintain a brand system, prepare files, negotiate a scope, and defend creative choices. A developer may write code, review pull requests, reproduce bugs, validate security assumptions, translate requirements, run deployments, and take responsibility when an automated suggestion breaks production.
ChatGPT and Copilot can assist parts of those jobs. They can generate rough drafts, boilerplate, alternate concepts, test cases, scripts, documentation, summaries, and snippets. They cannot, by themselves, establish whether a generated answer meets a customer’s requirements, complies with a contract, avoids infringing material, survives a production incident, or fits into an existing software and security process.
That distinction should not be mistaken for a guarantee of job security. Task substitution can still lead organizations to reprice work, shrink project budgets, ask fewer people to deliver the same output, or expect junior workers to produce at a more advanced level. The task bundle survives, but the staffing model changes.
For Windows and enterprise IT teams, the practical question is therefore not whether Copilot or another model can write a PowerShell function, generate an Excel formula, draft a policy, or suggest C# code. It is whether the organization has shifted the responsibility for reviewing, testing, securing, and maintaining that output onto the remaining employees—and whether it is still training new staff to perform those judgment-heavy tasks.
The study’s data source mix is a strength, but its window is short
Makridis uses both establishment-based and household-survey data, which is important for a creative workforce that includes employees and self-employed workers. The Bureau of Labor Statistics’ OEWS program, however, does not include self-employed workers in its employment estimates; the American Community Survey helps cover part of that blind spot.
Even with that broader approach, the study’s strongest labor-market estimates only reach the early post-ChatGPT period. ChatGPT launched publicly at the end of 2022, and the paper’s principal earnings findings run through 2023. That allows limited time for employers to change procurement rules, redesign job ladders, consolidate vendors, retrain workers, and decide whether an AI-assisted workflow actually holds up.
The study’s survey component also measures reported AI use rather than a controlled intervention. Workers who adopt AI may already work in better-paying organizations, have more technical confidence, or be assigned tasks that reward experimentation. A correlation between AI use and slightly stronger earnings growth cannot establish that the tool caused the income difference.
Makridis acknowledges the uncertainty, characterizing the evidence as cautiously positive and emphasizing that adoption is still evolving. That caveat is not boilerplate. It is central to interpreting the paper: the research measures a shock’s first phase, when adoption is uneven and many organizations are still experimenting rather than restructuring around AI.
Freelance markets show why occupation averages can miss disruption
Other research reaches less reassuring conclusions in a different part of the labor market. A study by researchers affiliated with Imperial College London, Harvard Business School, and DIW Berlin examined nearly two million job posts on a major online freelancing platform and found a 21% decline in postings for automation-prone writing and coding work relative to more manual-intensive categories in the months after ChatGPT’s release. It also reported lower demand for image-related work after image-generation tools arrived.
A separate study published in Organization Science found that freelancers in highly exposed occupations saw reductions in both employment and earnings after the introduction of ChatGPT and image-generation models. These studies do not directly contradict Makridis’ work because they observe a specific online contract market rather than occupation-wide U.S. employment and wages. But they do show why broad labor statistics can lag behind changes in the demand for discrete, readily specified tasks.
The distinction is especially relevant to software work. A full-time systems administrator, Windows developer, or enterprise security engineer is embedded in infrastructure, approvals, operational knowledge, and accountability. A one-off request to write a script, create a landing page, draft product copy, generate an icon set, or build a simple app feature is easier to replace, bundle, or demand at a lower price.
Organizations can therefore experience both realities at once: stable employment in established roles and a contracting market for freelance, junior, or routine project work. Calling that outcome “no job loss” overlooks the workers trying to enter the field or sustain themselves on smaller assignments.
Managers should measure the work that has disappeared from view
The ASU article is right to challenge obsolete productivity measures. Counting lines of code was already a weak proxy for software engineering; AI-generated code makes it worse. The relevant measures are whether a release meets its requirements, whether it reduces incidents, how many regressions it introduces, whether security review catches flaws before deployment, and how quickly a team can resolve failures.
There is a parallel measurement problem for labor. Executives introducing Copilot or comparable generative AI tools should track more than headcount and average compensation. They should examine whether task volume is rising without corresponding time, whether outsourced work is being reduced, whether junior roles are being eliminated, whether review burdens are concentrating on senior staff, and whether quality or incident rates change after automation.
A useful operating checklist is short:
- Teams should record which tasks are assisted by AI, which are fully automated, and which still require human approval.
- Managers should separate output gains from workload transfer, particularly when senior employees are expected to review larger volumes of generated material.
- Security and engineering teams should treat generated code, scripts, and configuration changes as untrusted until they pass the same testing and change-control requirements as human-authored work.
- Organizations should protect training paths for junior workers, because removing routine tasks can also remove the work through which people learn judgment and systems context.
Makridis’ research offers a defensible early conclusion: generative AI has not yet produced a clear, broad collapse in artists’ earnings. It does not establish that creative employment has been unaffected, nor that the risks facing freelancers, junior workers, and task-based contractors have failed to materialize.
For IT departments, the near-term consequence is straightforward. AI deployment should be assessed as a redesign of tasks, review responsibilities, hiring routes, and risk ownership—not as a simple headcount-saving tool or a productivity dashboard metric.