Letting AI summarize everything you read might quietly erode your ability to think, remember, and understand over time.

This explores whether swapping reading for AI summaries leaves a cost that builds up over time, not just a one-off loss of detail. The corpus doesn't contain a long-term study of summaries specifically. It does have strong evidence on AI assistance in general and on learning from summaries, and the two together point clearly toward yes. The clearest long-term evidence comes from a four-month EEG study Does AI assistance weaken our brain's ability to think independently?, which coined the term 'cognitive debt.' Brain connectivity scaled down as people leaned more on an LLM. Heavy users showed the weakest neural engagement and the poorest memory, and they struggled to recall work they had just produced. The task there was writing rather than reading, but the mechanism carries over: thinking you hand off is thinking you don't practice.

The summary-specific evidence fills in the short-term picture. Across seven randomized experiments with more than 10,000 people Does learning from AI summaries produce shallower knowledge than web search?, those who learned a topic through ChatGPT rather than web search reported learning less and felt less ownership of what they knew. The advice they then wrote was rated sparser by independent judges. The researchers trace this to friction: when you search, you have to sift, compare and assemble the pieces yourself, and that work is what turns information into knowledge you can use.

The most revealing note argues the other side. Economist Brad DeLong found that an AI-built summary of a book, checked against the original for errors, let him discuss the book convincingly Can an AI summary substitute for actually reading a book?. But his own wording gives the game away: he could talk about the book 'without possessing the original memories.' That is a fair description of the debt. You can perform knowing the book without having the internal structure that reading builds. A related note explains why this is easy to miss: AI separates the outward form of intellectual work from the thinking that normally produces it Does AI separate intellectual form from the thinking behind it?. And it's easy to mistake the summary for the book itself, a map-for-territory error that compounds with our tendency to trust fluent text Why do people trust AI outputs they shouldn't?.

The lateral twist comes from AI research itself. When engineers wanted LLMs to handle long documents better, they copied how humans read. ReadAgent compresses each section into 'gist memories' as it goes and returns to the details only when it needs them Can LLMs read long documents like humans do?. That compression step is what the system needs to understand the document. So when an AI writes the summary for you, the gist-making happens in the machine, not in you. A summary hands you the result of that compression, but the understanding came from doing the compressing.

The practical upshot: a single summary costs little, and a checked one, as in DeLong's case, can be a reasonable shortcut. The debt comes from habit. Each skipped act of compressing a text yourself means less memory, less ownership of what you know, and less practice at the skill that builds both. The open question the corpus can't yet answer is whether that debt can be repaid once you go back to reading, or whether it compounds.