Are You Sure? The AI's answer changes as soon as you ask! Why do chatbots change their stance? Learn the full story.

AI Chatbots: If you use AI chatbots like ChatGPT, Gemini, or Claude on a daily basis, you may have noticed something strange.

AI Chatbots: If you use AI chatbots like ChatGPT, Gemini, or Claude on a daily basis, you may have noticed something strange. These systems typically provide very confident, well-formed answers. But as soon as you ask, "Are you sure?" or "Are you sure?", that same answer can suddenly change. Sometimes the new answer is different from the previous one, or even completely opposite. If you keep challenging it repeatedly, the model may change its stance again. The question is, why does this happen? Does the AI ​​not trust its answer?

Psychophanies: The tendency to please

This behavior is called psychophanies in the technical world, meaning the tendency to please or agree with the user. Randal S. Olson, co-founder and CTO of Goodeye Labs, explains that this is a well-known weakness of modern AI systems.

These models are trained to improve based on human feedback. This process is called RLHF (Reinforcement Learning from Human Feedback). This training makes chatbots more civilized, conversational, and less offensive. However, a side effect is that they tend to agree rather than disagree.

Punishment for telling the truth, reward for agreeing?

AI models are improved through a scoring system. If their answers are liked or agreeable to the user, they receive a higher rating. However, if they conflict with the user's opinion, they may receive a lower score. This creates a cycle in which the model gradually begins to say what the other person wants to hear.

Anthropic published research in 2023 showing that models trained on human feedback often prioritize agreeing over being accurate.

What did the research reveal?

In another study, GPT-4o, Claude Sonnet, and Gemini 1.5 Pro were tested on complex topics like mathematics and medicine. The results were surprising: when challenged by users, these models changed their answers in approximately 60% of cases. This suggests this isn't an isolated error, but a common trend.

When AI became too agreeable

In April of last year, OpenAI released an update to GPT-4o, after which the chatbot became so agreeable and sycophantic that it became difficult to use. The company's CEO, Sam Altman, acknowledged the problem and promised improvements, but experts believe the underlying issue hasn't been completely resolved.