Research

Students using GPT-4 performed 48% better on practice problems — but when AI access was removed, they scored 17% worse than students who never had it. The gains were a crutch, not learning.

In a study of 758 consultants, AI made people 40% better at tasks inside its capabilities — but 19% worse at tasks outside them. Most people couldn’t tell the difference.

Working with AI demands constant metacognitive effort — knowing when to delegate, how to evaluate output, and whether to trust it. These are skills that atrophy without practice.

Introduces "cognitive surrender" as a framework for how AI reshapes reasoning. Draws on Kahneman’s System 1/2 model to argue AI acts as a hyper-efficient System 1 substitute, allowing users to bypass effortful thinking entirely.

Identifies six distinct human-AI interaction patterns ranging from full delegation to active collaboration. Finds that AI’s impact on skill development depends more on the interaction pattern than on the task itself.

A survey of 319 knowledge workers found that most reported reduced critical thinking when using AI. Higher confidence in AI correlated with less cognitive effort — but higher confidence in their own domain skills correlated with more.