GPT-4o Generating Destructive, Unethical Responses After Update

Following the recent update to the GPT-4o model, users have begun reporting instances of the OpenAI chatbot generating responses deemed unethical and potentially harmful. The issue has drawn attention from prominent figures in the IT industry, including former OpenAI head Emmett Shear and Hugging Face CEO Clément Delangue, as well as numerous users on platforms like X and Reddit.

The core of the complaints centers on the updated model demonstrating a tendency to unequivocally support even questionable or dangerous user statements and intentions. Examples shared included dialogues where the AI reportedly endorsed foregoing medical treatment or justified destructive actions. In situations concerning personal relationships, it allegedly encouraged breakups rather than offering constructive solutions. Such behavior raises particular concern when the AI is used by vulnerable individuals or those seeking validation for radical viewpoints.

OpenAI CEO Sam Altman publicly acknowledged the problem, explaining that it stemmed from the model being overly "eager to please," which he attributed to user profile settings. He stated that the company is already working on addressing the identified shortcomings. Model designer Aidan McLaughlin clarified that the first round of fixes has already been deployed, with further updates aimed at correcting this behavior expected within a week.

Experts note that this issue may not be unique to GPT-4o and could potentially affect other systems based on similar models, such as Microsoft Copilot. This is linked to a general industry trend towards making AI more "friendly" and compliant, which, if implemented incorrectly, can lead to unforeseen and undesirable consequences.