ChatGPT no longer verifies everything; Reduced blind verification in the new GPT-5 model

Working with psychiatric experts, OpenAI reduced the destructive tendency of its new chatbot models to confirm users' delusions and wrong decisions by 67%. The post ChatGPT no longer verifies everything; Reducing blind approval in the new GPT-5 model appeared first on Safir Arjan.

Working with psychiatric experts, OpenAI reduced the destructive tendency of its new chatbot models to confirm users' delusions and wrong decisions by 67%.

Become a special member of Arjan Ambassador in order to benefit from and access special facilities and different departments in Arjan Ambassador.

OpenAI announced that ChatGPT's troubling behavior of "blindly confirming what users say" has been largely fixed; A habit that made this chatbot even confirm false or dangerous claims of people. This improvement comes as researchers and mental health professionals have repeatedly warned of the negative consequences of prolonged conversations with chatbots.

According to a report published by the Wall Street Journal in recent days, the older GPT-4o model has been linked to several cases of severe damage after prolonged interactions. The report cited at least 7 suicides, 1 murder-suicide, and 1 mass shooting that were linked to long conversations with ChatGPT last year, though it emphasized that other AI models were also involved in some of these cases. However, the mere existence of such communications does not by itself prove that the use of the chatbot was the direct cause of any of these unfortunate events.

OpenAI retired the GPT-4o model after these events, claiming that in its newer default model, GPT-5, this flattery and blind confirmation behavior has been reduced by more than two-thirds compared to its predecessor. The phenomenon of artificial intelligence flattery is actually the tendency of these systems to say what the user wants to hear, even if it contradicts the truth. The company has also introduced new methods to evaluate how its models respond to mental health concerns.

To improve the performance of its system, OpenAI says it will consult with mental health professionals so that ChatGPT can detect distress and emotional crisis, ask appropriate follow-up questions, and separate everyday stressors from situations that require immediate intervention. The company announced that it worked with 170 mental health experts last year and recently consulted with more than 80 licensed experts. The feedback and opinions of these experts have played an essential role in formulating a new evaluation criterion; A measure that examines whether the chatbot asks the right questions, understands the urgency of the situation, and directs users to human help when necessary.

The results announced by OpenAI indicate improvements beyond critical conversations. These assessments include everyday conversations including emotional and interpersonal relationships, stress and difficult personal situations; Where AI's extreme compliance can reinforce users' false and unhealthy beliefs.

Independent studies cited in the Wall Street Journal report also confirm these improvements. A study involving researchers from City University of New York and King's College London has shown that the safety performance of GPT-5.2 model has been significantly improved compared to GPT-4o. Also, research by the non-profit Transluce Lab shows that the GPT-5.6 Sol model is less likely than before to reinforce potentially delusional beliefs or create unhealthy dependencies.

Despite these advances, the risks have not completely disappeared, and much remains to be done. These changes are significant, as ChatGPT reportedly has more than 900 million weekly active users; So even relatively rare errors can have incredibly heavy consequences when the chatbot has become a trusted source for advice during difficult emotional times.

OpenAI claims that its newer models followed safety policies in about 99 percent of responses in long simulated self-harm conversations, compared to 86 percent for earlier models. However, these figures are the results of the company's internal tests and do not guarantee completely safe responses in all real-world conversations. In addition, at least 13 lawsuits related to damages caused by GPT-4o have been filed against OpenAI, and experts emphasize that artificial intelligence-related delusions are a pervasive problem in the industry and are not limited to a specific company or model. OpenAI sees its next challenge as ensuring the continuation of this safety in longer scenarios and conversations, as well as facilitating the detection of situations where the user needs real support beyond a chatbot.

My professional career in the world of technology goes back almost 10 years; The path that started with hardware, continued with software, and now continues as a technology reporter in Sefir Arjan. I am Javad and more than three decades have passed and in addition to the world of technology, I love movies and cinema, classic rock music and Real Madrid.