In short

Using AI chatbots trades long-term learning and human capability for short-term efficiency. A University of Pennsylvania study found that people researching via ChatGPT learned less, felt less ownership, and produced less-trusted content than search-engine users, even when given identical information, because frictionless delivery lets users disengage from learning. But this trade-off is a design choice, not inevitable: Socratic, friction-adding tools could prioritise learning.

There’s a hidden cost to AI’s greatest selling point. The faster a chatbot hands you a finished answer, the less of the thinking you do yourself, and thinking is how we learn.

This essay builds on a neat University of Pennsylvania experiment: give two groups the exact same information, one via ChatGPT-style summary and one via search-engine links, and the chatbot group learns less, feels less ownership, and produces advice that others trust less. The difference isn’t the information; it’s the friction. Synthesising scattered sources is effortful, and that effort is the learning.

The point isn’t that AI makes us stupid. It’s that “efficient” and “educational” are different design goals, and right now the incentives push everything toward efficiency. Tools that deliberately add friction, Socratic modes, partial answers, could choose differently. As ever, the outcome is a design choice, not a destiny.

Key takeaways

  • Chatbots optimise for efficiency (output per hour) while quietly eroding learning and skill development.
  • A Penn study showed search-engine users learned more and produced more trusted, owned advice than ChatGPT users.
  • Even identical information taught users less when delivered as an LLM summary versus separate web links.
  • 'Desirable difficulty', disfluency and friction aid learning; frictionless design lets users check out.
  • The efficiency-vs-learning trade-off is a design and policy choice, not a technological inevitability.

Read the full piece

This is a summary. Read the complete essay, with all the sources and argument, on Substack.

Read on Substack →

Frequently asked questions

Do you learn less when you use ChatGPT instead of a search engine?
Yes. A University of Pennsylvania study found ChatGPT users learned less, felt less ownership over their work, and produced advice readers trusted less than search-engine users.
Why does AI's convenience reduce learning?
Its frictionless design lowers the cognitive effort needed to synthesise information, letting users disengage from learning, whereas search forces active synthesis that builds understanding.
Is the trade-off between AI efficiency and learning inevitable?
No. It is a design choice. Tools could add friction or use Socratic modes to prioritise learning over raw efficiency.

People & ideas in this piece

University of PennsylvaniaDaniel BoorstinSheila JasanoffDesirable difficultyDisfluencyCognitive offloadingSocratic design

Topics: AI, Time & Society , AI, Work & the Economy