Summarizing books with human feedback
OpenAI
OpenAI got an AI to summarize entire books by having it summarize chunks, then summarize the summaries. The trick isn't the summarizing — it's teaching AI on tasks humans can't fully check themselves.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI's latest research project sounds almost mundane on its face: teach a language model to write decent summaries of full-length books. But the actual point has nothing to do with book reports. It's about a much thornier problem — how do you train an AI system to do something well when no human reviewer has the time, patience, or memory to verify the output directly?
Nobody can read a 300-page novel in the time it takes to grade a single training example, and nobody wants to. So instead of trying to get a human to judge a summary of an entire book against the source text, the team broke the job into pieces small enough for people to actually check. The model summarizes short sections first, then summarizes those summaries, then summarizes that, climbing up a ladder of condensation until an entire book gets boiled down to a page or two. At each rung, a human can plausibly evaluate whether the summary of a chunk holds up — a task that's tedious but doable, unlike trying to fact-check a whole novel's worth of plot against a single paragraph.
This approach is sometimes called recursive task decomposition, and it's less about summarization specifically and more about a general strategy for supervising AI on jobs that outstrip human evaluation capacity. If you can't judge the final output directly, you judge the smaller steps that built it. It's the same logic that shows up in debates about aligning far more powerful future systems: humans may not be able to verify a superhuman AI's conclusion, but they might be able to verify each piece of its reasoning along the way.
The books themselves were, in a sense, beside the point. What mattered was whether the technique scaled — whether feeding human feedback into each layer of the hierarchy produced summaries that stayed coherent and faithful the whole way up. OpenAI's writeup frames the whole exercise as an early, modest test case for a much bigger question the field is still wrestling with: how do you keep a human in the loop when the loop keeps getting bigger than the human?
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous but genuinely important side of AI safety work, the kind that doesn't trend on social media because it's about grading rubrics and chunked feedback loops rather than chatbots doing party tricks. I'd rather see ten more papers like this than another splashy demo, because the scalable-oversight problem is exactly what breaks down first if these systems keep getting smarter faster than we get better at checking their work.
Read more about this at: OpenAI