About AI Use & Integrity | Legible Papers

The Short Version

Every summary here is written by a language model, and language models state things that are not true. The pipeline assumes that will happen. It restricts what the writing model is allowed to see, checks the finished text against quoted source material in code, and withholds anything that fails. What survives is then read by a person, who decides whether it is published at all. None of this makes a summary certainly correct. Each article links to the paper so any specific claim can be checked in a few seconds.

How Hallucinations Are Mitigated

1. Nothing is invented at the source

Papers arrive from the official arXiv and Hugging Face APIs. No model is ever asked to recall interesting research, which is where invented citations come from. The model that chooses what to feature sees only real titles and abstracts, and returns identifiers that are checked back against that list. A paper that does not exist cannot enter the pipeline.

2. The model that writes never reads the paper

One model reads the paper and pulls out its results, each one required to arrive with a quote copied word for word from the source. The pipeline string-matches every quote against the paper text and discards any result whose quote is not really there. A second model then writes the summary from the surviving facts alone, with no access to the paper. It cannot embellish source material it was never given. Prompts asking a model to be careful are advice it can ignore; withholding the text is not.

3. Finished text is checked, then withheld if it fails

Every figure in a new summary must appear in the quoted facts, as must every model, dataset, and benchmark named, and that covers the whole article: the one-line summary, the takeaways, the analysis, and the questions at the foot of the page. Sentences putting an opinion in the authors’ mouths must point at something the authors actually said. A separate model then reads the draft against the facts and hunts for claims they do not support. That reviewer is meant to be a model from a different company, on the grounds that a model is a poor judge of its own writing, but when that provider cannot be reached the check falls back to a model from the same company that wrote the summary. How often each of the two happens across the whole archive is counted further down this page. Each of these runs in code and has one outcome. A summary that fails any of them is discarded and the paper waits for the next run.

4. A person decides whether it is published

Everything above runs in code, and code cannot tell whether a summary is worth anyone’s time. A finished summary is not published when it is written: it waits in a queue until a human reads it and either publishes it or throws it away. Nothing reaches this site without that.

5. What the checks cannot do

Matching a quote proves the words are in the paper. It does not prove the summary drew the right conclusion from them, kept the surrounding context, or noticed the caveat two paragraphs later. Emphasis and framing are not checkable by code at all. Every article links to the paper and its PDF, and the honest use for this site is deciding what deserves your afternoon.

6. This page describes the pipeline, not every article

The checks above are what runs now, and almost everything already published was held to a weaker version of them. Until 18 August 2026 a summary’s figures and names were checked against everything the extractor produced, rather than only the parts backed by a quote, so a number or a model name the extractor introduced could pass. The one-line summary and the questions at the foot of the page were not covered by those checks at all. Separately, and for a long stretch, the independent claim review could not return a verdict, because the model it called had been retired and its fallback was being asked for the wrong kind of answer. Of the 989 summaries live on the site right now, 667 were written before that fix and carry no record of a review at all. Of the 322 that do carry one, 278 were reviewed by a model from a different company and 44 by the writing model’s own provider, which is the weaker of the two checks. Withholding a failed summary, and discarding a fact whose quote is not in the paper, did run throughout. Articles written from a paper’s abstract alone say so on the article itself. Older articles are regenerated through the current pipeline over time.

Limitations

A few things to keep in mind while reading:

Code samples: The code on each article is an educational sketch of the core idea, not a production implementation. Expect to adapt it before running it anywhere real.
Preprints are not peer reviewed: Most featured papers are preprints. Their results have not been through review yet and can change or fall apart under scrutiny.
Diagrams show the concept: Flowcharts are generated from the paper's description of its own system. They illustrate the idea and are not a formal claims chart.