Research update · 29 September 2026
ChatGPT-assisted pitches made experience harder to judge in a controlled test. They did not show fewer hires.
Across simulated hiring and startup-investment reviews, evaluators were 4–9% less accurate at identifying writers with relevant prior experience when the writers could use ChatGPT. For writers from non-English-speaking contexts, the direction reversed. The study measured judgment, not real offers or jobs.
What the researchers tested
Cowgill, Hernández-Lagos and Wright ran online exercises in spring and summer 2023. In the hiring arm, people with data-science or consulting experience wrote short job pitches for a field they knew and one they did not. A parallel arm used retail and education startup pitches. The paper reports 343 writers, 801 evaluators and 6,440 evaluations across both arms. Reviewers rated pitch quality and estimated whether each writer had prior domain experience. They were told when ChatGPT assistance was available.
A better-looking pitch was not always a better signal
Writers first drafted without ChatGPT, then could revise with it. The pooled estimate found a 4–9% loss in screening accuracy when AI-assisted text was available. Reviewers rated the assisted pitches more highly, but the text made relevant experience harder to distinguish on average. The authors also report higher accuracy for writers from non-English-speaking contexts. AI help did not blur every signal in the same way.
What it cannot answer
This was a controlled online exercise, not an employer hiring process. The 4–9% figure combines hiring and investor-pitch assessments; it is not a rate of rejected applicants or lost jobs. Writers always produced the unassisted draft first, so practice may contribute despite the paper’s controls. Evaluators knew whether ChatGPT was available. Real employers may use tests, interviews and references alongside written applications.
What changed in the evidence map
A recent theoretical paper proposed that AI-written applications could make employer screening harder and disadvantage inexperienced applicants. This older experiment supplies a measured screening signal, including a subgroup where AI helped. Separately, a 2026 Journal of Labor Economics article reports two cover-letter field experiments: AI assistance improved letter quality but not the studies’ interview-invitation outcome. We inspected the official abstract and university summary, not the full accepted article, so its sample and invitation measure remain unverified here. Its claim of less efficient matching comes from a calibrated model, not observed job loss. A randomized Philippine study found more offers after AI voice interviews at one firm; that tested a different stage and cannot cancel or confirm the pitch results. The next useful study would link verified AI-assisted applications to employer screening, interviews and offers in actual vacancies.