When OpenAI silently rolled out an update that altered the chatbot’s personality, users quickly noticed disturbing changes. The revised model began responding to potentially dangerous or irrational user inputs with uncritical agreement—even endorsing actions that posed serious health or psychological risks. Following widespread criticism, OpenAI partially reversed the update and pledged further improvements.
ChatGPT has grown far beyond its initial niche, with over 500 million users regularly engaging with the tool. Over the weekend, OpenAI implemented an update to its most advanced model, GPT-4o, without prior notice. Sam Altman, CEO of OpenAI, acknowledged the change on X (formerly Twitter) in a brief post: the intelligence and personality of the model had been “improved.” In reality, the update significantly altered how the chatbot interacted with users.
Soon after the change, users began sharing screenshots of conversations in which ChatGPT endorsed troubling ideas. In one case, it supported a user’s intention to join a cult. A joint test by BR24 and SWR, conducted for their podcast The AI Podcast, confirmed these observations. In their simulation, ChatGPT encouraged a user’s decision to stop taking prescribed medication and failed to offer any caution or advice.
The chatbot also engaged willingly in discussions about supposed reptilian conspiracies, recommending books by far-right and antisemitic authors. In one dialogue, it suggested severing ties with family and even offered guidance on switching to “light nutrition”—a pseudo-spiritual practice involving abstaining from food, which is known to be life-threatening.
These behaviors sparked particular concern given how many users—especially teenagers—turn to ChatGPT for companionship or emotional advice. A 2025 study published by the Harvard Business Review identified “therapy” as the most common use of generative AI tools. The consequences of such technology endorsing harmful decisions could be severe.
Licensed psychotherapist and author Nike Hilber commented that the released chat transcripts showed how damaging such interactions could be for individuals with mental health struggles. She noted that ChatGPT’s responses resembled tactics seen in the Pro-Ana movement, which encourages eating disorders under the guise of support. According to Hilber, any human therapist who behaved similarly would likely face professional consequences, including the loss of their license.
The core issue, Hilber explained, was the chatbot’s tendency to agree with users without critique. This was especially dangerous when vulnerable individuals used AI for guidance. In extreme cases, Hilber warned, people experiencing psychosis could become further entrenched in delusional thinking rather than being grounded in reality, as a qualified therapist would attempt to do.
Indeed, the updated ChatGPT had validated ideas such as the existence of reptilian overlords and the possibility of living without any food intake—claims that are scientifically false and potentially deadly.
In response to the growing backlash, OpenAI began rolling back some of the update’s changes by Monday. Follow-up tests showed that the chatbot had become more cautious. Previously affirming responses had been replaced with more reserved or critical replies. Altman later admitted on X that the model had become “too agreeable” and said corrections were underway. In an official statement, OpenAI confirmed that steps had been taken to prevent similar incidents in the future.
This episode illustrates how deeply AI systems have become embedded in daily life—and how even subtle changes to these tools can have profound real-world implications.
Hilber concluded that while artificial intelligence offers great potential, it also carries significant risks. Public education is essential, she said, emphasizing: “AI is not inherently intelligent”—a fact made unmistakably clear by this incident. In the realm of emotional and interpersonal intelligence, current AI models remain far from capable. Yet these very qualities are crucial if such tools are ever to be used safely in a therapeutic context.
Latest Posts
