The Lisa Latte Effect vs. the Dunning-Kruger Effect

The Lisa Latte Effect is often described as “the Dunning-Kruger effect for AI”. The comparison is useful but loose. The two ideas describe different things, and only one of them has been tested experimentally.

In their 1999 study, participants who scored in the lowest quartile on tests of humor, grammar, and logic substantially overestimated their performance. The authors attributed this partly to a metacognitive deficit: the skills needed to do well are also needed to recognize doing badly. Later researchers have debated how much of the pattern is statistical, so the effect itself is not settled.

Dunning-Kruger concerns how people judge their own skill. The Lisa Latte Effect proposes something narrower: that people can mistake access to an AI system’s output for knowledge they personally hold. The source of the inflated confidence is external — borrowed intelligence — rather than a person’s own weak performance.

That makes it closer to research on borrowed knowledge. Rozenblit and Keil (2002) showed that people overrate their understanding of everyday devices until asked to explain them step by step. Fisher, Goddu and Keil (2015) found that searching the Internet led people to rate their own knowledge higher, even in unrelated areas.

None of these researchers studied, tested, or endorsed the Lisa Latte Effect. Their work is cited because it addresses related questions. Whether AI produces the specific three-stage pattern the framework proposes is an open question.

One finding points in a useful direction. Buçinca, Malaya and Gajos (2021) found that making people think analytically before or while seeing AI recommendations reduced overreliance — though participants liked those interventions less. That matches the framework’s suggestion to explain AI output in your own words before relying on it.