AI reasoning studies linked higher AI literacy to less accurate self-assessment. Here is what the findings mean for executive review and judgment.
On February 1, 2026, Computers in Human Behavior published Daniela Fernandes and co-authors’ paper, “AI makes you smarter but none the wiser: The disconnect between performance and metacognition.” An October 19 analysis by AIR APAC revisited the research and its distinction between performing well and accurately judging that performance.
According to that analysis, participants using AI on logical-reasoning questions scored above a comparison group but overestimated their results. Higher AI literacy was also associated with greater confidence and less accurate self-assessment.
For executives, the useful distinction is between proficiency with a tool and evidence that a particular conclusion deserves trust. The practical leadership response is to make room for independent checks of consequential reasoning rather than treating familiarity or confidence as validation.
AIR APAC’s summary describes two online studies involving participants recruited in the United States through Prolific. The first involved 246 people. The second involved 452.
Participants worked on 20 LSAT logical-reasoning questions with AI assistance. The research examined their performance alongside their estimates of how well they had done. It therefore addressed both task accuracy and metacognition: a person’s assessment of their own thinking and performance.
Those are distinct outcomes. A score records how many answers were correct. A self-estimate records how successful the participant believed the exercise had been. Comparing them reveals whether confidence is well calibrated to the result.
The summary reports that participants overestimated their performance across skill levels. It contrasts this with the familiar pattern in which weaker performers overestimate their results while stronger performers underestimate them. In these studies, stronger performance did not remove the tendency toward overestimation.
The research also examined AI literacy. According to AIR APAC, greater literacy was associated with higher confidence but lower accuracy in judging personal performance.
The central finding is therefore more specific than a claim about whether AI makes people better thinkers. AI-assisted performance and accurate self-assessment were separate outcomes. Success on the task did not automatically come with an accurate understanding of that success.
The operational implication is to separate evidence of competence from evidence that a decision is sound. Knowing how to use AI and being comfortable discussing its output answer different questions from whether the underlying reasoning holds up.
For consequential work, a reasonable safeguard is an independent review of the reasoning. That means assessing the recommendation itself rather than relying on the presenter’s confidence or history of using the tool.
A review could ask:
These are questions about the work product. They keep the review focused on something another person can examine rather than an impression of how capable the user feels.
This distinction matters when considering how to assess judgments under pressure. Confidence may be relevant to a discussion. It should not close the discussion before the reasoning has been examined.
The same logic applies to the way leaders interpret AI training. Training participation can establish that someone has completed training. It should not stand in for an assessment of the reasoning behind a particular recommendation. Tool familiarity and decision validation belong in separate parts of the conversation.
For demanding professional lives, the practical question is where review belongs in the workflow. Rather than asking everyone to maintain constant vigilance over their own confidence, leaders can make independent scrutiny part of consequential decisions. The emphasis is on a reviewable process rather than a personal feeling of certainty.
That is also a useful lens for reflecting on the quality of past decisions. The relevant comparison is between the original assessment and the evidence that later becomes available. A persuasive explanation at the time is not the same thing as a well-supported conclusion.
AIR APAC reports that participants in the first study scored about three items higher than a norm-population comparison. At the same time, their estimates of their own results were about four items too high out of 20.
Those figures describe different comparisons. The first concerns observed scores relative to a comparison dataset. The second concerns the gap between participants’ actual scores and their estimates of those scores. They should not be combined into a single measure of AI’s effect.
The reported result allows both observations to be true: participants performed better than the comparison population and believed they had performed better than they actually had.
According to the summary, the second study replicated the reported pattern. The significance for leaders lies in that coexistence of stronger performance and inaccurate self-assessment. Better output does not make accurate self-evaluation redundant.
The first study was quasi-experimental rather than randomized. AIR APAC notes that it used an earlier dataset for comparison. Its score difference therefore should not be presented as a clean causal estimate of how much AI assistance improved performance.
The task was also abstract and carried no workplace stakes. These were online logical-reasoning exercises rather than executive decisions with organizational consequences. The findings do not directly establish how founders or employees assess AI-assisted work in professional settings.
The AI-literacy finding was an association. It does not establish that literacy caused poorer self-assessment. Nor does it justify treating AI education as harmful. The finding concerns the relationship between literacy and calibration within the reported research.
The proposed operational safeguard has its own boundary: independent review was not established here as an effective intervention. It is a prudent leadership implication drawn from the distinction between performance and self-assessment. Its effectiveness in a particular workflow needs evaluation rather than assumption.
Finally, these results provide no basis for claims about improved energy or reduced stress. They concern reasoning performance and the accuracy of self-evaluation. For ExecuFuel, the practical position is to keep those boundaries visible while translating the findings into a clear decision principle: assess consequential reasoning on evidence rather than confidence alone.
The next useful studies would test whether the same pattern appears in work with meaningful consequences. They would also examine whether independent review improves the accuracy of users’ self-assessments as well as the quality of their decisions. Those are related but distinct outcomes.
The broader concern appears elsewhere in the literature. A separate paper in AI and Ethics uses the phrase “sustained illusion of cognitive competence” in its title. That framing points toward an important research question: when does feeling more capable correspond to demonstrably better judgment?
For workplace research, a useful comparison would separate the quality of an AI-assisted answer from the user’s assessment of that answer. Testing review processes would then help establish whether a safeguard improves calibration or merely adds another confident opinion.
For leaders considering where AI reliance meets human responsibility, the editorial outlook is straightforward. AI capability and judgment assessment deserve separate attention. A useful industry shift would be to evaluate the reasoning behind consequential recommendations alongside the skills required to produce them.
The next advance worth watching is evidence that a defined review process helps people judge their AI-assisted work more accurately. That would move the discussion from confidence in the tool toward a clearer account of when its users deserve confidence.
Stay connected for research and practical guidance on executive performance, energy, focus, sleep, recovery and longevity. Ideas built for people who want to stay sharp, capable and effective for the long run.
Build habits and systems that support clear thinking, steady energy and long term capacity throughout a demanding career.
explore the Blog
