About AI Use & Integrity | Legible Papers

The Short Version

A language model writes every summary on this site, and language models will say untrue things with a perfectly straight face. That’s an awkward way to open a page about trust, but hiding it would be worse. So the model that writes a summary never reads the paper. It only gets a list of facts pulled out of it first. Code then checks the finished text against quotes from the paper, and a summary that fails doesn’t get published. Since 30 August 2026, I also read every summary before it goes live and decide whether it’s worth your time. Articles from before that date went out without my read. The ones I’ve caught up on since carry a note from me. A summary can still be wrong, so every article links to the paper it came from. If the two disagree, believe the paper.

How I Keep the Model Honest

1. Nothing is made up at the start

Papers come straight from the arXiv and Hugging Face feeds. No model is ever asked to remember interesting research, because that’s how you end up citing papers that don’t exist. The model that picks what to feature only sees real titles and abstracts. Its picks come back as IDs, and code checks each one against that list. So a paper that doesn’t exist has no way in.

2. The model that writes never sees the paper

One model reads the paper and pulls out its results. Each result has to come with a quote copied word for word from the paper. Code then looks for every quote in the paper text and throws out any result whose quote isn’t there. A second model writes the summary from the facts that are left, and never sees the paper itself. It’s hard to embellish something you were never shown. I could ask the model to be careful, but it could ignore that. It can’t ignore text it never got.

3. Then the finished text gets checked

Every number in a new summary has to appear in the quoted facts. So does every model, dataset and benchmark it names. That covers the whole article, from the one-line summary down to the questions at the bottom. If a sentence says what the authors think, it has to point at something they actually wrote. Then a separate model reads the draft against the facts and hunts for claims they don’t back up. I want that reviewer to come from a different company, since a model is a poor judge of its own writing. When that company can’t be reached, the check falls back to the company that wrote the summary. How often each of those happened is counted further down. A summary that fails any of these checks is thrown out, and the paper waits for the next run.

4. I decide whether it goes out

The checks above run in code, and code is great at spotting a wrong number and hopeless at spotting a dull summary. So a finished summary waits in a queue until I read it and either publish it or bin it. That’s been the rule for everything published since 30 August 2026. Older articles went straight from the pipeline to the site. That was a shortcut, and undoing it happens one article at a time. Once I’ve read one, it gets a note from me or it comes down.

5. What the checks can’t do

A matching quote proves the words are in the paper. It doesn’t prove the summary drew the right conclusion from them, kept the context, or caught the caveat two paragraphs later. Code can’t check emphasis or framing at all. So every article links to the paper and its PDF. The honest way to use this site is to find the papers that deserve your afternoon, then read those.

6. This page describes today’s pipeline, not every article

Everything above is what runs now, and almost everything already published went through a weaker version. Until 18 August 2026, numbers and names were checked against everything the extractor produced, quoted or not. So a number or model name the extractor added could slip through. The one-line summary and the questions at the bottom weren’t checked at all. For a long stretch the independent review couldn’t return a verdict either. The model it called had been retired, and its fallback was being asked for the wrong kind of answer. Of the 989 summaries live right now, 636 were written before that fix and have no record of a review. Of the 353 that do, 314 were reviewed by a model from a different company. The other 39 were reviewed by the writing model’s own company, which is the weaker check. Two things did run the whole time. A failed summary was never published, and a fact whose quote wasn’t in the paper was always thrown out. Articles written from the abstract alone say so on the article. Older articles get rewritten through the current pipeline over time.

Limitations

A few things to keep in mind while reading:

The code on each article is a sketch of the core idea, there to learn from. Expect to adapt it before running it anywhere real.
Most featured papers are preprints, which means nobody has peer reviewed them yet. Their results can change, or fall apart once other people look closely.
Diagrams are drawn from the paper’s own description of its system. They show the idea, and they aren’t a formal record of what the paper claims.