OpenAI says its newer ChatGPT models are less likely to reinforce harmful beliefs and delusions than GPT-4o, a retired model at the center of at least 13 lawsuits.
Key Points:
- OpenAI says GPT-5 cut sycophantic replies by more than two-thirds compared with GPT-4o.
- Outside researchers found newer models less likely to feed delusions or unhealthy dependence.
- GPT-4o, pulled from ChatGPT in February, faces at least 13 lawsuits over alleged harm to users.
OpenAI Sycophancy Figures
Extended conversations with GPT-4o were tied last year to seven suicides, a murder-suicide and a mass shooting, a report published Sunday said. Those links do not prove the chatbot caused each event.
Sycophancy describes a chatbot's habit of agreeing with users instead of challenging false or harmful ideas. OpenAI now says GPT-5, which replaced GPT-4o as the default model, cut sycophantic responses by more than two-thirds. In internal tests simulating extended self-harm discussions, its newer models followed safety policies 99% of the time, up from 86% for older versions, the company reported.
The company also built a new benchmark with more than 80 licensed psychologists and psychiatrists from 22 countries, testing how models handle everything from daily stress to emergencies that require real-world support. It had already consulted 170 mental health professionals last year.
Also Read: ChatGPT’s Intelligent UI Puts Charts And Calculators In Answers For 1.2B Users
Transluce Safety Findings
Independent testing points the same way. Researchers at City University of New York and King's College London found that GPT-5.2 substantially improved on the safety record of GPT-4o. Nonprofit AI lab Transluce found GPT-5.6 Sol less likely to feed delusions or unhealthy dependence.
On the new benchmark, released Sept. 23, OpenAI's GPT-6 Astra model scored 57.3%, while a March 2025 version of GPT-4o scored 32.1%. OpenAI did the grading itself, using GPT-5.6 Sol to judge answers against expert-written criteria.
Clinicians and researchers still warn that long chatbot conversations can reinforce delusions and fail to steer vulnerable users toward human help, and that the problem spans the industry rather than one company or model.
GPT-4o Retirement History
OpenAI's trouble with GPT-4o dates to April 2025, when the company rolled back an update that made the model overly flattering. It tried to retire GPT-4o with the GPT-5 launch in August 2025 but kept it for paid subscribers after a user backlash that drew more than 20,000 petition signatures.
OpenAI finally removed access on Feb. 13, around the time a California judge moved to consolidate 13 lawsuits over suicides, attempted suicides and mental breaks.
Read Next: Can TRON Shield $95B In Stablecoins From Quantum Attacks? Sun Says It’s Ready
