Artificial intelligence may be reshaping not just how we solve problems but how we evaluate our own abilities — and not necessarily for the better. A new multinational study led by Finland’s Aalto University reports that common AI chatbots almost eliminate the well-known Dunning-Kruger effect, a cognitive bias in which low-skill individuals overrate their abilities while high-skill individuals underestimate theirs. But instead of improving accuracy, AI use appears to push everyone toward overconfidence, regardless of actual skill.
Researchers from Finland, Germany, and Canada found that when participants used AI to tackle logical-reasoning questions, they placed excessive trust in the chatbot’s answers. Surprisingly, those with the greatest AI experience displayed the highest misplaced confidence. The findings, published in the February 2026 edition of Computers in Human Behavior, challenge the assumption that increased AI literacy leads to more accurate self-assessment.
The Dunning-Kruger effect is traditionally revealed through cognitive tests that measure attention, judgment, decision-making, and language skills. But with AI in the loop, the study shows that people bypass the normal feedback mechanisms that help them judge their own performance. As AI tools become more ubiquitous, the researchers warn this shift could have significant repercussions.
To test the impact, the study assigned 500 volunteers reasoning problems sourced from the Law School Admission Test. Half of the participants worked unaided, while the other half could use ChatGPT. After completing the tasks, both groups were asked to evaluate how well they thought they performed, with bonus compensation offered for accurate self-assessment. Results showed widespread “cognitive offloading” among the AI group — users often accepted the chatbot’s first answer without deeper reflection, leading to weaker metacognitive monitoring, or the ability to evaluate their own reasoning.
The consequences were clear: everyone who used AI thought they performed better than they actually did, and the performance-perception gap between high and low-skill participants narrowed dramatically. While AI helped raise accuracy overall, it simultaneously eroded users’ ability to judge that accuracy. The researchers say this trend risks creating a climate of miscalibrated decision-making as people grow increasingly confident in answers they have not critically examined.
The study arrives at a moment when experts are debating whether modern large-language models have become too agreeable, often reinforcing user assumptions rather than challenging them. According to the Aalto University team, this dynamic could exacerbate users’ tendency to overestimate their abilities, ultimately weakening critical-thinking habits and diminishing our capacity to evaluate the reliability of information.
To counteract this erosion of metacognitive skills, the researchers suggest redesigning AI systems to actively prompt deeper reasoning. This might include chatbots offering confidence scores, posing reflective questions such as “What might you have missed?” or guiding users to verify answers before accepting them. Such measures, they argue, could help restore healthy skepticism and preserve critical-thinking skills as AI becomes more deeply embedded in daily tasks.
The findings reinforce an emerging global consensus that AI education must go beyond technical skills and emphasize critical thinking. As the Royal Society recently argued, training users to question, verify, and reflect may prove just as essential as teaching them how to operate advanced AI tools.
