Can an AI summary substitute for actually reading a book?

J. Bradford DeLong, writing on his Grasping Reality Substack (2026-10-06), describes falling from "150 serious books a year" in his prime to "perhaps 75," 40% of them re-readings. To close the gap, he built what he calls a "conspectus" — "book-by-book folders... notes, briefing, and synthesis documents" generated by an AI harness ("ollama/gemma4:31b-mlx, overseen and guided by a harness constructed by anthropic/claude-opus") for books he will "never" read, such as Felix Somary's memoir The Raven of Zürich. After "spending an hour on the conspectus folder (and then half an hour checking random pieces... against the actual book itself)," his verdict is affirmative: "I think so" — the conspectus is a close enough approximation of "the memories, the brain states, that I would have had I in fact read the book" that he "can now talk about the book as convincingly as if I had read it."

DeLong names the mechanism explicitly: he and his AI systems together form a "Chinese Room" that can "rapidly simulate... the words I would have said or written if I did have the memories," without his possessing the memories themselves. What licenses his confidence is spot-checking the conspectus against the source text for fabrication ("I have run enough spot-checks... to be confident that the hallucination level here is very low"), not independent verification of its claims — a check on accuracy substituted for the deep engagement ("I would think about them. I would then go off and argue with the author") his old reading practice supplied.

This extends Does AI reshape expert work into knowledge management? with a first-person, non-professional instance of the same shift: DeLong manages and validates AI output not for a client but to patch a hole in his own reading life, performing the "is this good enough" check himself rather than the "what is true, and how do I know" work of reading. It sits in tension with Does AI assistance weaken our brain's ability to think independently?: DeLong attributes his reading shortfall to causes upstream of AI ("half of the gap comes from lower energy"), and frames the conspectus as compensation rather than cause — leaving open whether what compensates for lost reading is itself a further form of offloading. It also echoes, inverted, Does polished AI output trick audiences into trusting it?: there DeLong is both producer and sole audience of the artifact, and checks it against the source rather than taking its finished appearance on faith.

The excerpt is a single self-reported case — one person, one model harness, an unspecified number of spot-checks on one book — with no measure of how often or how badly the conspectus's details stray from the source beyond DeLong's own felt confidence. It also leaves unresolved the question DeLong poses himself: "Is that knowledge, or only a good performance of it?" He explicitly declines to answer for the "inside" view ("that is a hard question that I do not yet have an answer to"). What the excerpt supports is narrow: a motivated, literate reader who checks the output against the source can find an AI-built conspectus a convincing stand-in in conversation — not that such conspectuses reliably substitute for deep reading in general, or for readers without the background to spot-check them.