University of Georgia Study Finds AI Interview Scoring Encouraged Exaggeration

AI interview screening is meant to make early hiring steps easier to process. But a University of Georgia-led study found that candidates in one-way video interviews reported and displayed more exaggeration when they believed AI, rather than people, would judge them—and the AI system tested did not penalize it.

The research, published in Information Systems Research, examined hundreds of online job seekers taking asynchronous interviews: candidates recorded answers rather than speaking live with an interviewer. The finding is not simply that an algorithm can make a weak judgment. It suggests that uncertainty about an automated evaluator can change the material the system is asked to assess.

A screen candidates tried to decode

Participants who expected AI evaluation described embellishing qualifications as necessary in an unpredictable environment. Researchers’ analysis of their recorded videos and answers supported their self-reports: candidates were more likely to stretch or embellish their experience when an AI agent was the expected evaluator than when a human was.

What the candidates were told

• Their recorded videos would be reviewed by AI rather than a human.
• In the transparency condition, the AI would assess facial expressions, verbal sentiment, keywords, teamwork, job abilities, work style and personality.

Transparency changed the behavior

The researchers tested a narrower alternative to fully revealing an interview system’s inner workings. One group was told that AI would review its videos and received specific information about the criteria. That information reduced exaggeration and increased authentic behavior.

Those candidates reported and displayed authenticity at levels comparable to people who thought a human would review them. The result challenges the assumption that keeping evaluation criteria vague necessarily prevents applicants from gaming a hiring process; in this experiment, uncertainty itself appeared to prompt more strategic behavior.

A result with a clear boundary

The evidence concerns the AI candidate-rating system used in this study, not every hiring tool or every kind of interview. It nevertheless identifies two linked failure points for automated screening: candidates may adapt their answers to an opaque process, and the system may fail to distinguish that adaptation from a candid account of the same qualifications.

What remains unresolved is how much explanation employers should provide, and whether the effect holds across other systems and hiring settings. The study’s immediate lesson is more practical: an AI interviewer’s design includes not only how it scores people, but what people think it expects from them.