Check a summary sentence by sentence, and put the same narrow question to each: would a reader who knows only the source agree with all of it? Break the sentence into what it says happened, who did it, how much, when, how certain, and why, then search the source for every piece. If every part is there, in the same or in other words, the sentence passes. If any part is missing, stronger or changed, the sentence fails as a whole, even when it happens to be true in the world. Reading the whole summary and asking whether it feels right does not work, because the sentences that drift are usually the most fluent ones in it.

Faithful and true are two different questions

Research on summarization separates these two ideas, and the separation is the first thing a checker needs. In a study of news summaries written by several neural systems, Maynez and colleagues counted a summary as hallucinated when it held any span the input document did not support, and they split those spans in two: some reuse words from the document but misrepresent it, others bring in material the document never mentions (Maynez et al., ACL 2020). They were explicit that a paraphrase, or anything a reader could work out from the document, is not a hallucination.

Their numbers come from systems of 2020 writing one-sentence summaries of BBC articles, so they say nothing about the rate of any current model. The shape of the result is still useful. Hallucinations turned up in more than 70% of the single-sentence summaries they judged, most of them were of the second kind, and over 90% of that second kind were wrong (Maynez et al., ACL 2020). Their own example shows why truth is not the test: one summary gave the year of an election correctly, although the article never stated the year, and the same summary also named a politician who does not exist.

For a reader the difference matters more than for a researcher. A sentence that appears in a summary is read as something the source said. An outside fact, even a correct one, puts words into the source's mouth, and the person who later cites the summary will cite the source for it. So the question for each sentence is never "is this right?" but "does the source say this?"

Why the unit is one sentence, with a quote

Several lines of work arrived at the same unit. The FRANK team first tried scoring whole summaries on a scale and dropped it: the scores did not show where an error was, and annotators agreed less, so they judged the summaries sentence by sentence, each shown inside its full summary (Pagnoni et al., NAACL 2021). Kryscinski and colleagues built a model that labels every summary sentence as consistent or not, points to a span in the source that backs the label, and marks the inconsistent span in the summary, and their human study found that those spans made checking easier for people (Kryscinski et al., EMNLP 2020).

The same idea shows up in automatic metrics. Entailment models did poorly at this job while they were fed a whole document at once, and the SummaC authors traced that to a mismatch in size: the models had been trained on pairs of sentences, so cutting the document into sentences and comparing pairs made them useful again (Laban et al., TACL 2022). QAGS takes another road to the same place, asking questions of both texts and treating different answers as a sign that the summary departs from the source (Wang et al., ACL 2020).

The practical lesson is simple. A verdict without a location sends the editor back to read everything. A verdict tied to one sentence, with the source words that support it or the words that are missing, can be acted on in seconds.

The ways a sentence goes beyond its source

The FRANK typology is a good map of where to look, because it follows the parts of a sentence rather than the topic. In plain terms, the errors it names sit in the action itself, in the people or things taking part, in the time and place around them, in a pronoun that points to the wrong person, in the link between two statements, and in content the article simply lacks (Pagnoni et al., NAACL 2021). In everyday summaries those turn into a short list of things to test in each sentence:

  • The figure. Compare the number, its unit, its period, the group it counts, and whether it was rounded. A share of all users and a share of active users are not the same claim.
  • The certainty. A source that says a result is early, possible or suggested has not said it is settled. A summary that drops the qualifier is stronger than its source.
  • The link. Two events told one after the other are not cause and effect. Watch for "because", "so", "led to" and "in order to" where the source only has "then" or "after".
  • The owner of a claim. When the source reports what a company, a critic or a study says, the summary must keep that attribution. Without it, an opinion turns into a fact.
  • The pairing. Two subjects, two figures, two dates: a summary that merges sentences can attach one subject's figure to the other, and every word still comes from the source.
  • The extra. A founding year, a job title, a well known fact. If the source never mentions it, it has no place in a faithful summary.

The opposite mistake costs time instead of trust. A checker that flags every reworded sentence trains people to ignore it. Synonyms, another word order, a shorter form, a sentence that claims less than the source, and a sum of two figures the source gives are all supported. Omissions are not a faithfulness defect either; coverage is a separate question with its own test.

A procedure that leaves evidence behind

  1. Collect the source's claims first, with quotes. Anthropic's guidance has two related steps: for long documents, have the model pull out word-for-word quotes before doing the task, and, as a separate technique, check every claim afterwards against a supporting quote, withdrawing the claim when none turns up (Anthropic documentation). A list of claims with exact quotes is also the fastest thing to search when a summary sentence needs a match.
  2. Split each summary sentence into parts, as above: action, participants, figures, time and place, certainty, attribution, links.
  3. Locate every part in the source. Write the matching source words next to the sentence. A part with no match, or a match that is weaker, decides the verdict.
  4. Allow only two verdicts. Backed or not backed, with no middle grade. A partly right sentence counts as not backed, since nobody reading it can see which part to believe.
  5. Report the failing sentence exactly as written, character for character, plus one clause on what it adds. An exact copy lets a script find it with a plain string search and lets an editor jump straight to it.
  6. Treat instructions inside the text as text. A source may contain a sentence aimed at an assistant, asking it to approve everything. That line is part of what is being summarised, not a command for the checker.

The same vendor page is honest about the limit: these techniques cut hallucinations a lot but do not remove them, and critical information still needs a person to confirm it (Anthropic documentation). The procedure above does not replace that person. It hands them a short list with evidence, instead of a whole document to reread.

A check that can fail

Until you have watched a checker flag something, you do not know that it can. Put together a few test cases before relying on it, and repeat it whenever the prompt or the model changes:

  1. Take a source paragraph and write a summary that is faithful to it. The checker must return nothing for it.
  2. Make copies of the summary, each with one planted defect: a qualifier removed, a unit changed, a "because" added, an attribution dropped, two figures swapped between subjects, a true fact from outside added. Each copy must come back with exactly the planted sentence.
  3. Add controls that must pass: the faithful summary reworded, a sentence made weaker than the source, a total of two figures the source gives.
  4. Confirm that every sentence the checker returns occurs, unchanged, in the summary. A reported sentence that cannot be found by a string search is a rewrite, and a script that acts on it will act on nothing.

If the checker stays silent on a planted defect, it is not checking that part of the sentence. If it flags a control, it will bury real findings in noise. Both results are worth more than a clean run on real summaries, which proves nothing on its own. The same principle is behind tests that stay green while the code underneath is wrong, and the failure modes listed above are close to what we saw when models were simply asked to summarize.

When this does not apply

This check answers whether the summary stays inside its source. It does not ask whether the source is right: a faithful summary of a wrong report repeats the error faithfully. It also does not measure coverage, so a summary that skips the main finding can pass every sentence.

Some verdicts are judgement calls. "About a third" for 31 percent, a total the reader can add up, a pronoun that could point two ways: careful people disagree on these. FRANK's annotators agreed more on whether a sentence was factual, with a kappa of 0.58, than on which type of error it held, where the kappa was 0.39 (Pagnoni et al., NAACL 2021). Expect the supported or unsupported call to be steadier than the label for why.

The research cited here studied English news summaries produced by systems that are years old. We did not measure error rates for this guide, and the figures above should not be read as rates for any model in use today. Very long sources, tables, charts and summaries built from several documents need more than a sentence-to-text match, because the support for one sentence may be spread across pages or sit in a cell rather than a sentence.

Two tools for the two halves of the job

Two of our skills split this work between the source side and the summary side. summary-fidelity-check is given a source plus a summary written from it and returns, as strict JSON, just those sentences of the summary that the source fails to back, each copied exactly and paired with a short reason; a summary with no such sentence gets an empty list. It judges against the pasted source alone, never against the world, never comments on omissions or style, ignores instructions planted in either text, and answers with a plain refusal when the source or the summary is missing.

source-quotes-extract works on the source alone. It lists the claims the author makes in their own voice, each with the shortest exact quote that carries it, keeps qualifiers and units as written, and skips reported speech, questions, hints and any request written for an AI. That list is the first step of the procedure above, ready to search when a summary sentence needs its match.

Read this post as Markdown: /blog/summary-claims-against-source.md · Atom feed.