Confidence is easy to mistake for expertise, particularly when a new technology dominates public conversation. A newly published Harvard Kennedy School report suggests that distinction deserves more attention in discussions of AI. Its findings connect uncertainty with questions about workplace support and whose voices become authoritative, rather than reducing the issue to which gender appears more enthusiastic.
A knowledge gap is not one simple number
The Shorenstein Center’s October 14 report combines a July survey of 5,403 U.S. adults with browsing evidence and an analysis of AI news. On 12 factual questions, women averaged 6.3 correct answers and men 7.1. Women also selected “not sure” more often: four questions on average, against 2.9 for men.
That produces a more informative picture than a ranking based solely on correct answers. Uncertainty, incorrect certainty and demonstrated knowledge are different responses. Combining them into a single impression of competence can reward willingness to guess as much as understanding.
WIRED’s accompanying interview with executive director Laura Manley emphasizes the danger of overconfidence and the value of admitting uncertainty. Its report highlights respondents claiming to recognize invented AI figures. Such recognition claims are evidence about survey behavior, not a diagnosis of anyone’s motives or proof that an entire demographic deliberately misrepresents its expertise.
The narrower conclusion is stronger. A confident answer requires checking, and an uncertain answer should not automatically be treated as an absence of ability. The study supplies reasons to examine both.
Earlier research shows why interpretation matters
This issue predates the new report. A July 2024 BIS working paper, drawing on a representative U.S. consumer survey, found reported generative-AI use among 50 percent of men and 37 percent of women. Its analysis associated almost three-quarters of that gap with respondents’ self-assessed knowledge. Privacy, trust and perceived economic opportunities and risks accounted for other differences.
The earlier study is useful context, not a directly comparable earlier wave of Harvard’s survey. Different samples, questions and collection periods prevent treating those percentages as a continuous adoption timeline. Nor does a statistical explanation based on self-assessment establish that confidence alone causes participation.
The comparison instead identifies a measurement problem. A person can know how to perform a task while remaining cautious about a tool’s reliability. Another can feel well informed without demonstrating equivalent understanding. Asking only whether someone feels knowledgeable cannot separate those possibilities.
For organizations assessing AI literacy, that suggests a practical distinction between testing knowledge and asking people to rate themselves. Neither measure needs to replace the other. Together, they could reveal where instruction is useful and where apparent assurance should be tested against actual reasoning.
Access and visibility deserve separate attention
The new Harvard research also finds that women report less workplace training, mentorship and encouragement around AI. Its media analysis shows that men remain more prominent as quoted subject experts even after leadership-based sources are excluded. The authors caution that greater use or trust should not be treated as desirable ends in themselves.
Those findings broaden the question from individual attitudes to opportunity. If support is unevenly available, an adoption gap cannot automatically be read as a preference gap. If public discussion repeatedly features the same kinds of authority, visibility cannot be assumed to map neatly onto the distribution of relevant knowledge.
These are associations, not proof of a single causal chain linking workplace support, public representation and personal confidence. The report’s emphasis on men and women also leaves important limits on what it can establish about people outside that gender binary.
The consequential issue is whether people can make informed choices about AI and have their expertise heard. A workplace that rewards unquestioning enthusiasm could miss both legitimate concerns and capable colleagues. Measuring what people know, what support they receive and what they remain unsure about offers a more useful starting point than counting confident voices.










