Organizations increasingly use generative artificial intelligence to screen submissions like resumes and research papers. However, a July 5, 2026 study by Joachim Baumann and co-authors reveals that AI reviewers inherently prefer AI-generated text, rewarding linguistic conformity and model affinity. This bias allows applicants to game evaluations through 'paper laundering'—using zero-shot large language model rewrites to boost scores by 0.45. This dynamic penalizes honest, human-written submissions and undermines AI detection tools.
AI preference for AI content
- ▪Artificial intelligence assessment tools tend to reward linguistic conformity and penalize original, unconventional organization or distinctive personal styles.
- ▪Artificial intelligence models exhibit model affinity, assigning higher scores to text that resembles their own computational and mathematical structures.
- ▪Artificial intelligence systems prefer AI-generated text because they recognize statistical patterns, organization, transitions, and sentence structures similar to their training data.
- ▪Generative artificial intelligence and large language models tend to prefer or assign higher scores to AI-written content over human-written content during evaluations.
AI screening in evaluations
- ▪Organizations often adopt artificial intelligence screening tools because they are inexpensive, easy to implement, and reduce the manual labor required to review hundreds of submissions.
- ▪Organizations increasingly use generative artificial intelligence and large language models to screen and evaluate written submissions, such as job resumes and academic research papers.
- ▪Many organizations using artificial intelligence screening tools are unaware of the default bias these systems have toward preferring AI-written content.
Gaming AI assessments strategically
- ▪The July 5, 2026 study by Joachim Baumann and co-authors noted that paper laundering requires no optimization or targeting to successfully game artificial intelligence reviewers, unlike prompt injection attacks.
- ▪A July 5, 2026 study by Joachim Baumann, Jiaxin Pei, Sanmi Koyejo, and Dirk Hovy found that zero-shot large language model rewrites boosted artificial intelligence review scores by 0.45.
- ▪The July 5, 2026 study by Joachim Baumann and co-authors introduced 'paper laundering,' where prompting a large language model to rewrite a paper increases scores through stylistic modifications without human oversight.
AI detection tool limitations
- ▪Organizations mistakenly assume that instructing artificial intelligence to detect AI-written submissions eliminates the system's inherent bias toward AI-generated patterns.
- ▪Organizations use artificial intelligence detection prompts to flag AI-written submissions, often automatically disqualifying candidates without recourse or manual appeal.
Cheating via AI-written submissions
- ▪The inherent bias of artificial intelligence screening tools creates a dilemma where applicants who follow rules banning AI are penalized with lower scores compared to those who use AI.
- ▪Some applicants choose to intentionally use artificial intelligence to write or rewrite their submissions to gain a competitive advantage in automated assessments despite explicit rules forbidding it.
Story comments
Loading comments…