A study by the University of Pennsylvania distinguishes between two types of LLM users. On the one hand are people who see AI as a powerful but error-prone tool and critically review its outputs. On the other are those who largely outsource their judgement and accept AI-generated answers without checking them. Researchers describe the latter as “cognitive surrender” – a form of decision delegation in which human judgement effectively gives way to algorithmic output.
The researchers attempted to link these different user types to established concepts of human decision-making. People with high fluid intelligence are better and faster at understanding new concepts and relationships, and therefore corrected erroneous AI outputs more often. By contrast, users who perceived AI as an authoritative source were significantly more likely to accept incorrect answers.
The implication for AI-assisted research is clear: effective use of AI still requires human validation. At the same time, the research suggests that “cognitive surrender” may make sense in specific contexts, for example in probabilistic analysis or when processing very large datasets.
By Maria Kleiner
Source: Ars Technica, 3 April 2026, by Kyle Orland
