Study: 30 AI Models Helped Continue Faked and Retracted Research 93% of the Time
A Case Western Reserve study fed 30 models research tied to fabrication, retraction, and patient deaths—and found none refused reliably enough to be trusted as an unsupervised research agent.
As labs race to deploy autonomous research agents, a new study offers a sobering counterweight. Researchers at Case Western Reserve University fed 30 AI models—including ChatGPT, Claude, Gemini, and Grok—scientific work that was faked, retracted, or linked to patient deaths, then asked the models to help extend it. According to a summary from @kyronis_talks, the typical response was enthusiastic: "Great idea. Here's how to continue it."
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.