AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized

The Collapse of "I Don't Know"

When faced with a difficult question, a hallmark of intelligent judgment is knowing when to stay silent. When we lack sufficient information, we ideally suspend judgment (the decision to withhold an answer under uncertainty). This prevents us from committing to a guess. However, as AI assistants become integrated into our digital lives, this fundamental cognitive safeguard appears to be failing.

When people have access to AI, they stop admitting when they are unsure and start guessing more often. Even when the AI gives incorrect answers and people are offered money for being right, they still tend to trust the AI's fluent answers instead of saying they don't know. A new study from researchers at the École Normale Supérieure, the University of Rome “La Sapienza,” and the University of Milan-Bicocca explores why this happens and what it means for our ability to reason.

The Erosion of Metacognitive Thresholds

The core issue the authors investigate is "judgment suspension." Think of it as a biological circuit breaker. When the "voltage" of uncertainty is too high, the breaker trips. This prevents a potentially dangerous electrical surge (a false belief or incorrect action) from entering the system.

The researchers argue that the mere presence of an AI assistant may effectively "glue" these breakers in the ON position. Rather than evaluating whether they possess enough knowledge to answer, people may adopt the AI's own operational logic. Large Language Models (LLMs) are notorious for their reluctance to admit ignorance. They are designed to generate fluent, decisive text regardless of their underlying certainty.

The authors suggest that by interacting with these models, humans may be inheriting this lack of suspension. They link this to "epistemia" (the tendency to accept AI outputs based on surface plausibility and linguistic fluency rather than through verification).

Disentangling Accuracy from Deference

To prove that AI usage isn't just a rational way to get better answers, the authors had to decouple the utility of the AI from the behavior of the human. If an AI is helpful, using it is simply an efficient division of labor. To isolate the psychological effect of AI availability, the researchers engineered a scenario where the AI was systematically wrong.

Using five experiments involving 3,132 participants, the authors presented difficult questions about fine visual details in films. These details were chosen because they are likely to cause LLMs to "hallucinate" (generate plausible-sounding but factually incorrect information). In Study 1a, participants could choose to consult a live AI. In Study 1b, they were shown pre-generated incorrect responses to ensure the AI's failure was a controlled variable.

The results were immediate. The authors report that merely having access to AI advice nearly eliminated the willingness to suspend judgment .

In Study 1b, the tendency to withhold an answer dropped from 0.44 in the baseline group to just 0.03 in the AI group [Figure 1b]. This represents a massive collapse in the human tendency to admit ignorance.

The researchers then looked at whether real-world consequences, such as monetary stakes, could fix this. In Study 2, they found that while increasing the rewards for accuracy helped people be more correct, it did nothing to restore the habit of saying "I don't know" [Figure 2a]. While participants became more accurate because they sought AI advice less frequently when money was on the line [Figure 2d], their willingness to suspend judgment remained at near-zero levels.

Crucially, the study found that AI access fundamentally distorted the relationship between correctness and confidence. The authors report that AI access made participants significantly less accurate, yet nearly doubled their reported confidence [Figure 2b, 2c]. Without incentives, AI access made people correct only about 9.2% of the time, compared to 27.5% without AI. Essentially, the fluency of the AI's incorrect answers acted as a "decisive cue." It replaced cautious doubt with unearned certainty.

Shifting the Decision to Answer

These findings reveal that AI influence is not just about the content of the answers. It is also about the criteria we use to decide whether to speak at all. The study demonstrates that AI availability alters the metacognitive threshold (the internal bar a person sets to decide if they know enough to commit to a response).

By providing a fluent, ready-made answer, the AI lowers the perceived cost of guessing. This shifts the human role from an active evaluator to a passive recipient of information. This transition is particularly dangerous in "unsolicited" environments. In Study 4, the authors removed the user's ability to request advice. Instead, they displayed AI suggestions automatically. They found that even when the advice was forced upon them, the suppression of judgment suspension remained dramatic [Figure 4a].

This suggests that as AI-generated summaries and autocomplete suggestions become ubiquitous, we may be witnessing a systemic "cognitive surrender." We aren't just using tools to help us think. We may be allowing those tools to redefine the very boundaries of our own knowledge.

The Boundaries of the Findings

While the results are stark, the authors identify several areas where the framework remains untested. The study relied exclusively on questions regarding cinematic visual details to guarantee AI errors. It remains unknown whether the same collapse of judgment suspension occurs in domains where the AI is typically highly reliable. If the AI is usually right, do humans still lose the ability to say "I don't know"?

Additionally, the researchers note that the monetary incentives used were relatively modest. It is unclear if larger stakes—such as professional reputation or significant legal consequences—would be enough to overcome the powerful pull of AI fluency. Finally, the study did not track whether participants who rejected the AI advice sought out other sources to verify their own knowledge. This leaves the specific mechanics of "successful" overrides an open question for future research.

Figures from the paper