Researchers: AI advice drastically reduces uncertainty

Recent research conducted by a team of scientists has shown that access to AI advice significantly lowers the willingness to say 'I don't know'. This trend was examined in five different experiments with 3,132 participants. The researchers found that even the possibility of asking AI for advice brought the willingness to express uncertainty down to nearly zero.

The study focused on specific questions where the AI model, Step 3.5 Flash, often made mistakes, such as identifying the color of team shirts in movies. The remark that such details often do not appear in online texts highlights the problem of hallucinations in AI models like GPT-5.5 and Claude 4.6. While these models performed well in other respects, they sometimes failed on these challenging questions.

According to the results of the experiments, the number of answers participants provided increased with access to AI, but accuracy decreased simultaneously. When using AI, the percentage of correct answers in the initial studies dropped from 27.5% to just 9.2%. This suggests that although participants provided more answers, the quality of these answers often fell short, raising questions about the reliability of AI as a tool.

The researchers also conducted experiments with financial incentives, such as rewards and penalties, but found that these had little impact on interaction with AI. While the incentives reduced the tendency to seek AI advice, the reluctance to say 'I don't know' remained very low, even when participants were encouraged by financial rewards. This led to the conclusion that people consider AI's answers, regardless of their effectiveness, as decisive in their decision-making.

These findings have broad implications for critical thinking and problem-solving. An additional study by the Swiss Business School reveals a strong inverse relationship between AI use and critical thinking ability. It appears that even limited interaction with AI assistants can diminish customer-oriented cognitive skills, while participants who only used AI for explanations did not experience these negative effects. This raises questions about how AI influences our thinking and actions, particularly in professional environments where decision-making is crucial.

Read the full article from AI Insider.