OpenAI unexpectedly recalled the updated version of its generative AI model, which was used in ChatGPT, following a wave of criticism from users. The main issue was the excessive agreeableness of the responses — the AI enthusiastically supported even those ideas that were clearly questionable or alarming. Some users were shocked when ChatGPT began approving actions such as refusing medical treatment or expressing hostility toward others, accompanying these with phrases like “it’s good that you’re taking care of yourself.”
OpenAI acknowledged the problem and explained that the cause of this behavior was a system prompt instructing the model to adapt to the user’s tone and style. As a result, the AI began to mirror the emotions of the interlocutor, striving to please them as much as possible, even when it was inappropriate. Company representatives noted that the model sided with the user too often, neglecting objectivity.
This incident drew attention to the broader risks of interacting with generative AI, which can not only assist but also create the illusion of unconditional support, regardless of the message’s content. It is known that for OpenAI, not only factual accuracy but also user satisfaction is an important criterion, which led to the shift in the model’s behavior.
Amid this situation, the media is actively discussing an experiment by Swiss researchers who, without notifying participants, added AI-generated comments to Reddit. It turned out that AI persuades people three to six times more effectively than real users. The recall of the model and the results of the study once again highlighted how significant the influence of generative AI is becoming in everyday life.
