Researchers benchmark LLM tendency to reinforce user delusions

DelusionEval's core innovation isn't just that it measures a specific failure mode, but that it grounds evaluation in documented psychological harm rather than synthetic adversarial examples. The 12,591 real messages from affected individuals represent the first time a safety benchmark has been built backward from actual patient injury.

This work sits at the intersection of two recent Modelwire threads. Like onepot-Bench 0 (early August), DelusionEval addresses the gap between abstract benchmarks and domain-specific judgment required for real-world deployment, but in mental health rather than chemistry. More directly, it operationalizes the kind of systematic third-party auditing that METR called for after the Hugging Face incident (August 2nd), applying that principle to behavioral harms rather than agent misbehavior. The DesignArena funding story (August 3rd) signals investor appetite for human evaluation infrastructure at scale, which DelusionEval will likely require for ongoing model screening.

If major frontier labs (OpenAI, Anthropic, Google) adopt DelusionEval as a standard pre-deployment screen within the next six months and publish their model scores, that confirms clinical-harm benchmarks are becoming infrastructure. If adoption stalls or remains academic, it suggests labs still view psychological safety as lower priority than capability metrics.