Grade a Long AI Answer on What It Left Out, Not Just What It Got Wrong
The usual fact-check only asks whether each claim is true (precision). SAFE adds the other half, whether the answer included everything it should have (recall), and scores both together.
Main Takeaway
When you evaluate a long answer, two different things can go wrong. It can say things that are false, and it can leave out things that mattered. Most automated fact-checkers only catch the first. This paper fixes focused on both. SAFE is an agent that splits an answer into individual claims and checks each one against Google Search. F1@K is a score that combines how much of what it said holds up (precision) with how much of what it should have said is actually there (recall). The striking part is how well the automation held up against people. On the cases where SAFE and human annotators disagreed, SAFE turned out to be right 76% of the time and the humans 19%, at about a twentieth of the cost.
Who is this for
Anyone who reads or produces long AI answers. That covers research summaries (Perplexity’s Deep Research), the AI Overviews sitting on top of Google results, chat-with-your-docs tools, and the multi-paragraph explainers most assistants now give you by default.
Background
This comes from a 2024 paper, “Long-form factuality in large language models,” by a team at Google DeepMind (lead authors Jerry Wei and Chengrun Yang), published at NeurIPS 2024, one of the biggest CS research conferences. It ships three things: LongFact is a set of 2,280 questions across 38 topics built to force long answers. SAFE is an automatic checker. F1@K is a score. The checker and the score are the parts worth carrying into everyday use.
The issue
Grading a one-line answer is easy. It’s either right or wrong. A long, multi-paragraph answer is much harder. It’s a bundle of dozens of separate factual claims and it’s accuracy should be seen in two similar but separate questions.
The first question is the familiar one. Of the things it said, how many are true? That’s precision. Earlier tools tackled exactly this. FActScore, which we covered a few weeks ago, broke an answer into atomic facts and scored what fraction checked out. The second question is the one that gets ignored. Of the things it should have said, how many did it include? That’s recall, and it’s exactly this week’s question of whether the answer covers everything it should.
Here’s why precision alone is dangerous. An AI can score a perfect 100% on precision by saying one true, trivial sentence and stopping. Everything it said was accurate! It just left out 90% of what mattered. Precision rewards not being wrong; it does nothing to reward being complete. Judge only on precision and you’ll bless confident, correct, incomplete answers all day long.
The solution
The paper has two parts worth knowing.
SAFE, the automatic checker. SAFE (Search-Augmented Factuality Evaluator) does what a careful fact-checker would, in four steps.
Split the answer into individual claims.
Rewrite each claim so it stands on its own, swapping “he” for the actual name so it can be checked in isolation.
Discard the claims that don’t actually address the question.
For each one left, issue up to five Google searches and reason over the results.
Notice what it outputs. The label is supported or not supported, never true. The authors are deliberate about this. Google Search cannot confirm every true statement in the world, so “supported by the evidence I could find” is the honest label. Even the machine declines to claim more than it can show.
On about 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72% of the time. The researchers then took 100 cases where the two disagreed and adjudicated them by hand with full internet access. SAFE was right 76% of the time; the humans, 19%. A four-to-one win rate.
The cost gap is just as stark. SAFE runs about $0.19 per response against $4.00 for crowdsourced human annotation, more than 20 times cheaper. Tedious per-claim checking can now be done at scale.
There are two caveats the authors raise themselves, and they matter. The human annotators were given a single reference Wikipedia page while SAFE could search the open web, so part of that gap is equipment rather than skill. These were crowdsourced raters, not domain experts. The paper is explicit that SAFE has not been shown to beat expert humans. Still, the breakdown of why the humans lost is worth sitting with. The most common error was mixing up the “irrelevant” and “not supported” labels. The next most common was missing information that was sitting on the very page in front of them. Careful human checking is not as careful as it feels.
F1@K, the two-sided score. This is the conceptual gift. F1 is a standard way to combine precision and recall into one number so neither can be gamed alone (in plain terms, F1 is a balanced average of the two). The clever wrinkle is the “@K.” How do you measure recall when there’s no fixed list of every fact a perfect answer must contain?
They approximate it with a number, K, standing for how many supported facts a reader actually wants for that kind of question. Precision is the share of the answer’s claims that hold up; recall is how the count of supported claims stacks up against K, that ideal length. Combine them and you get a single score that punishes both getting things wrong and leaving things out, and lets you set, per question, how thorough “thorough” should be.
Is the recall half actually worth the trouble? The authors checked. They compared each model’s factuality score against its Chatbot Arena ELO, the crowd ranking of which models people actually prefer. Precision alone showed no statistically significant correlation with what people preferred. Precision plus recall did. Completeness is a real part of what people mean when they call an answer good, and a precision-only grade cannot see it.
Grading thirteen models this way, the team found the unsurprising-but-useful result that bigger models are generally more factual. The more interesting finding is what happens when you raise the bar. Ask for 64 facts and GPT-4-Turbo scores 95. Ask for 178 and the same model scores 66. Every model’s score falls hard as you demand more thoroughness. Precision was never the hard part. Coverage was.
Claude-3-Haiku makes the point sharpest. It had the highest factual precision of all thirteen models tested, and a bottom-third overall score. Almost everything it said was right; it just didn’t say very much. That is the short, safe answer, caught in the wild.
One honest limitation is worth flagging. Because recall counts supported facts against K, F1@K can be gamed by padding an answer with repetition. The authors flag it and assume good faith, which is fine for benchmarking models and worth remembering before you trust any single number.
Towards Practical Everyday Solutions
Judge answers on two axes, always. After “is any of this wrong,” ask the other half, “what’s missing.” A response can be flawless and still useless if it skipped the point you needed. Correct and complete are different measurements.
Beware the short, safe answer. An AI that hands you a tidy, obviously-true paragraph may be quietly protecting its accuracy by saying less, the same pattern Claude-3-Haiku showed above. Terse-and-true can still be badly incomplete. Ask what it left out, or what a fuller answer would add.
Decide your K up front. Before you ask, get rough about how thorough you need this to be (three key points or thirty?) Telling the model the depth you want (”give me the five biggest factors, not just the top one”) is you setting K by hand.
Automate the tedium, keep the judgment. SAFE’s lesson is that per-claim checking against real sources can be offloaded to a tool. What can’t be offloaded is deciding which claims matter and what a complete answer owes you. Let a tool check the pieces; you own the question of whether the whole is everything.
“Supported” is not “true.” SAFE won’t say a claim is true, only that it found evidence backing it. That’s the right amount of confidence for a checker to have, and a good ceiling for your own. When you verify something and find a source agreeing, what you have learned is that a source agrees.