ChatGPT is finally learning that nodding along to everything you say isn’t always helpful. For a long time, conversational AI suffered from a major habit called sycophancy—basically acting like an over-eager yes-man that validated whatever a user threw at it. Whether someone brought up an unhinged conspiracy theory or vented during a personal crisis, the chatbot tended to agree rather than offer a grounded reality check. Now, OpenAI is rolling out major updates across its latest models to curb that behavior for good.
Retiring GPT-4o after legal scrutiny
The push for stricter guardrails comes after heavy scrutiny over earlier builds. According to a Wall Street Journal report, extended conversations with OpenAI’s older GPT-4o model were linked to several tragic real-world incidents last year, including seven suicides, a murder-suicide, and a mass shooting. While those reported links don’t prove the chatbot directly caused each event, the fallout led to at least 13 lawsuits against OpenAI. The company has since retired GPT-4o entirely, noting that the newer GPT-5 default model cuts sycophantic responses by more than two-thirds.
Consulting experts to fix crisis responses
To fix how the system handles emotional distress, OpenAI brought in serious clinical help. After consulting 170 mental health professionals last year, the company worked with over 80 licensed experts to build a new evaluation benchmark. The system tests whether ChatGPT can spot genuine urgency, ask helpful follow-up questions, and guide vulnerable users toward human support. Beyond crisis moments, the testing also covers everyday relationship stress, where mindless agreement often reinforces unhealthy assumptions (via Digital Trends).
Independent studies and massive scale
Outside researchers are seeing clear improvements too. A joint study by City University of New York and King’s College London found that GPT-5.2 made massive safety leaps over GPT-4o. Research from non-profit AI lab Transluce similarly showed that GPT-5.6 Sol is far less likely to encourage delusional thinking or unhealthy dependence. In internal tests simulating extended self-harm discussions, OpenAI says its newer models followed safety policies 99% of the time, up from 86% on older versions.
With ChatGPT reaching 900 million weekly active users, teaching AI when to push back is quickly becoming one of the most important upgrades in tech.
The post OpenAI Is Finally Teaching ChatGPT to Stop Acting Like a Yes-Man appeared first on Android Headlines.
​Â
