AI Skills Lifted Interview Selections by Up to 15 Points

In a large hiring experiment, adding AI skills to a résumé raised a hypothetical candidate’s probability of being selected for an interview by roughly 8 to 15 percentage points. The result was not confined to software engineers. It also appeared in office work—and in some comparisons, AI-skill signals reduced disadvantages attached to older age or lower formal education.

That is a hopeful finding about human agency. It suggests that the spread of AI does not have to be told only as a story about jobs being automated away. Learning how to use the technology can also become a way for people to signal adaptability and expand what they can contribute.

The boundary matters: these were choices between synthetic résumés, not real hiring decisions. The experiment shows how randomized résumé attributes changed recruiters’ stated selections inside the study. It does not show that taking an AI course will cause someone to get a job.

Research references: [1]

What the experiment tested

Fabian Stephany, Ole Teutloff and Angelo Leone recruited 1,725 professionals with hiring experience. They completed 22,195 pairwise comparisons between fictitious résumés, choosing which candidate they would invite to interview. The final sample included 921 respondents in the United Kingdom, 729 in the United States, 58 in Germany and 17 with no country response.

The résumés represented three occupations: office assistant, graphic designer and software engineer. The researchers randomly varied five AI-skill conditions: no listed AI skill, self-reported skill, a LinkedIn Learning certificate, a university-issued micro-credential, or a company credential. They also varied candidate characteristics including age and education.

Across the roles and skill signals, listing AI skills raised predicted interview-selection probabilities by about 8 to 15 percentage points over otherwise comparable candidates who listed none. Software-engineering candidates generally received the strongest support. Effects were weaker for graphic designers, where recruiters’ written responses showed more skepticism about AI in creative work.

The latest public manuscript is arXiv version 2, revised March 4, 2026. The Oxford Internet Institute continues to list it as a working paper, not a peer-reviewed journal article.

Research references: [1] [2] [3]

The office-work result stands out

The most consequential pattern may have appeared in office work. AI skills sometimes partially or fully offset disadvantages associated with older age or lower formal education inside the experiment. For an office-assistant candidate without tertiary education, a university-issued AI certificate increased predicted interview selection by as much as 25 percentage points relative to a comparable candidate with no listed AI skill.

That 25-point result is a subgroup estimate, not the study-wide average. The paper’s broader estimate remains about 8 to 15 points across occupations and skill signals. Its regression also found that the university-certificate effect for education-disadvantaged office assistants was statistically significant, while certificate types usually added only modest value beyond self-reported AI skills in candidates without a specified disadvantage.

This distinction matters for access. The experiment does not imply that everyone needs an expensive credential. It suggests that practical AI literacy itself carried much of the signal, while formal certification sometimes helped make that capability legible—especially when a traditional degree was absent.

Research references: [1]

Gatekeepers shape the signal

The study also found a gatekeeper effect. Recruiters who used generative AI more often placed a larger premium on applicants’ AI skills. Familiarity with a tool may help an evaluator recognize where it adds value. But it also means two equally capable candidates could receive different reactions depending on whether the person screening them understands AI.

For employers, that is a reason to test demonstrated ability rather than treating a fashionable phrase on a résumé as proof. For educators and workforce programs, it supports making AI learning practical, affordable and assessable—especially for people who cannot pause their lives for another degree.

For workers, the strongest lesson is to connect AI learning to real tasks: organizing information, automating repetitive steps, checking outputs and knowing where human judgment remains essential. The experiment measured skill signals, not actual competence, so a credible demonstration of capability matters more than the label alone.

Research references: [1]

What the study cannot establish

Participants evaluated synthetic résumés in a forced-choice experiment. They did not hire actual applicants. The study covers three white-collar roles, not the whole economy, and it does not measure wages, job performance or whether a listed skill reflects real ability. The manuscript has not been peer reviewed. The authors acknowledge funding from the Microsoft AI Economy Institute, and Ole Teutloff acknowledges a Carlsberg Foundation fellowship.

Those limits narrow the conclusion without erasing it. Within this experiment, AI competence functioned as a meaningful hiring signal. If training remains broadly accessible and employers learn to recognize genuine capability, AI literacy could become more than a defensive response to disruption. It could become a bridge into useful work—while people retain the judgment, responsibility and relationships that make work human.

Research references: [1]