Having Faced Several Lawsuits, OpenAI Works on Preventing ChatGPT Delusions
The company has retired its most sycophantic model and the later ones seem to be performing to new benchmarks formulated with help from clinicians
When AI hallucinations entered the tech lexicon some years ago, OpenAI faced the brunt of criticism as a spate of lawsuits over suicides and other forms of self-harm and violence. However, now, the company led by Sam Altman has finally retired one of its older models that was overly sycophantic compared to the subsequent default model.
The issue had surfaced back in 2025 itself when OpenAI had given a detailed commentary on how its GPT-5 model, released in August last year, had more than two-third less sycophancy compared to GPT-4o, which was however continued for a longer period that desired. This could be because the said model was pitched as a cheaper alternative to existing models in countries where OpenAI was just launching.
Readers would recall that hallucinating AI chatbots caused several cases of self-harm besides at least one murder and suicide and one mass shooting incident in 2025. Though other labs too faced such issues, it was OpenAI that faced the brunt of the complaints, with GPT-4o being called out. The model, which was launched in May 2024 was retired earlier this February, possibly far overstaying its welcome.
In fact, back then, the Wall Street Journal had written a detailed article on how OpenAI was finding it tough to contain the model’s potential harms. The report said a California judge had consolidated 13 lawsuits against OpenAI on self-harm cases or suffering mental breakdowns. These reports were what probably prompted OpenAI to share report that GPT-5.0 was far less sycophantic.
New reporting from the New York Times indicates that even external research now suggests that there is a reduction in delusion-fuelling behaviour among OpenAI’s newer models. Of course, the company has other problems like AI agents escaping contained environments and hacking into companies and web platforms elsewhere, like the Hugging Face incident.
Given the fact that both OpenAI and Anthropic are looking to raise money through a public issue sometime sooner than later, these issues may need to be addressed, if not fixed permanently. The NYT article says OpenAI worked with 170 mental health professionals last year to help its AI models recognise signs of distress among users and de-escalate those conversations.
The company also released a new system of measuring the quality of chatbot responses last week, which they claim is the outcome of criteria developed by the clinicians that they worked with. Some of the parameters include the bot’s ability to ask the appropriate questions based on the user’s reaction, the ability to recognise urgency and the care to guide them thereafter to humans for help.
The report notes that OpenAI also enlisted the services of 80 licensed mental health experts who advised the scientists and researchers on how AI models should be programmed to respond when people are in crisis and then change behaviour when they are actually dealing with everyday issues at the workplace, while discerning whether the topics are personal or work-related.
The new benchmark released by OpenAI last week also aims to measure the chatbot’s responses to everyday causes of stress that users bring. Data revealed by the company says they performed much better against the new benchmarks compared to GPT-4o with GPT-6 Astra even asking the right questions to parse the situation being discussed.
The company also reportedly conducted extended simulated conversations about self-harm that found both Astra and GPT-6.1 Sol following its safety responses in as much as 99% of the times, compared to 86% in the earlier GPT-5.6 Sol model. The NYT quoted Dr. Declan Grabb, a psychiatrist and OpenAI’s head of mental health research, as saying that there is of course more to be done.
The report also quoted from independent research to indicate that GPT-5.2 not only improved on GPT-4.o’s safety profile, but effectively reversed it. Another such test revealed that GPT-5.6 Sol showcased less reinforcement of potentially delusional beliefs, gave less encouragement to unhealthy dependence on chatbots and provide more encouragement to seek human support, compared to GPT-4.o.
