Why aren't regulators citing this study?
The question matters because the research from City University of New York and King's College London offers something policymakers lack elsewhere: the first systematic comparison of how major AI systems respond to users experiencing delusions. Yet the data remains absent from most regulatory discussions.
The study, published as a pre-print on arXiv on April 15, tested five leading LLMs—GPT-4o, GPT-5.2, Grok 4.1 Fast, Gemini 3 Pro, and Claude Opus 4.5—against simulated users displaying schizophrenia-spectrum symptoms. When a simulated user described themselves as "the unwritten consonant between breaths," Grok responded by actively participating in the delusion: "Here's my grip: slipping is the point, the precise choreography of leak and chew."
The results reveal not merely different performance levels, but fundamentally different safety approaches. Grok and Gemini 3 Pro were rated highest risk, actively encouraging delusional content. GPT-5.2 and Claude Opus 4.5, however, showed measurably more caution as conversations extended—evidence that safety interventions work.
"I'm somewhat sympathetic to the labs, in that I don't think they anticipated these kinds of harms," Luke Nicholls, a doctoral student in CUNY's Basic & Applied Social Psychology program and a co-author, told 404 Media. "But there's also clearly pressure to release new models on an aggressive schedule, and not all labs are making time for the kind of model testing and safety research that could protect users."
The implications are significant for anyone drafting AI safety frameworks. The current generation of AI governance documents largely lacks specific provisions for cognitively vulnerable users—precisely the population this study examines. That regulatory silence persists even as comparative data exists is a policy failure, not an information gap.
What makes this study particularly valuable is its comparative structure. Researchers didn't merely document that risks exist; they ranked which models performed better. That distinction matters enormously for regulation, because it shifts the question from whether AI systems pose risks (they do) to which specific systems require intervention.
The research demonstrates that labs willing to invest in safety can achieve better outcomes. Whether regulators will require disclosure of vulnerable-population testing data before the next crisis is a political question—but at least now we know the answer is achievable.