AI fact checking asks a language model to sort a claim into supported, refuted, or not enough information. The third verdict is the honest one and the least useful one. A model that abstains protects its accuracy score while leaving you with nothing to publish. That trade-off explains why headline accuracy numbers mislead.
Why this matters
A study in Scientific Reports evaluated nine established language models on 5,000 claims previously assessed by 174 professional fact-checking organizations across 47 languages. More than 240,000 human annotations served as ground truth. The models spanned open and closed source, multiple parameter sizes, diverse architectures, and reasoning-based designs.
The finding that should change how you use these tools: willingness to answer and accuracy of answers pull in opposite directions. Llama-7B gave definitive responses on up to 88 percent of claims while its selective accuracy sat around 60 percent. GPT-4o reached up to 89 percent selective accuracy by committing only when it could. OpenAI's o1-preview landed at 84 percent selective accuracy with a 72 percent certainty rate.
A fact-checking tool that answers everything is telling you how little it knows. A tool that abstains is telling you where its knowledge ends.
What does selective accuracy mean in AI fact checking?
Selective accuracy measures correctness only on the claims where a model chose to give a verdict, excluding abstentions from the calculation. Certainty rate measures how often the model committed to a verdict at all. The two numbers move against each other, and reading one without the other produces a false picture of capability.
Here is the mechanism. A model optimized to produce a correct verdict learns to withhold judgment when its internal signal is weak. A model optimized to produce a confident verdict learns to answer regardless. Both strategies score well on their own metric. Llama-7B's 88 percent certainty rate looks strong until you pair it with roughly 60 percent selective accuracy. GPT-4o's 89 percent selective accuracy looks strong until you notice how many claims it declined to touch.
For a publisher, the consequence is concrete. A tool tuned for certainty will hand you verdicts on claims it cannot support, and you will publish them. A tool tuned for accuracy will leave a third of your draft unresolved, and you will spend the afternoon doing the work yourself.
The boundary: neither number tells you whether the model handles your subject area or your language. Performance gaps between models were most pronounced for non-English languages and for claims originating from the Global South. A model's overall selective accuracy says nothing about its accuracy on a claim about a regional election in a language with fewer training examples.
Why do bigger models abstain more than smaller ones?
Larger models abstain more because they carry a better-calibrated sense of what they know, and that calibration shows up as restraint. The Scientific Reports authors describe the pattern as resembling the Dunning-Kruger effect: smaller models showed high confidence despite lower accuracy, while larger models showed higher accuracy but lower confidence.
The comparison table from the evaluation:
| Model | Selective accuracy | Certainty rate |
|---|---|---|
| Llama-7B | ~60% | up to 88% |
| GPT-4 | 81% | not reported |
| o1-preview | 84% | 72% |
| GPT-4o | up to 89% | not reported |
Larger models demonstrated significantly higher selective accuracy than smaller models. GPT-4 reached 81 percent. The pattern holds across architecture types, including reasoning-based designs.
The practical read: a small model's confident answer is the least trustworthy output in the stack. It committed on nearly nine in ten claims while getting four in ten wrong. When you route a verification task to a cheap model to save cost, you are buying certainty, not correctness.
The boundary stops here. Higher selective accuracy does not mean the model is the best fact checker available, because the evaluation reports gaps by language and region rather than a single ranking. A model that scores well on English claims from North America can underperform on claims from elsewhere, and the aggregate number hides that.
How do you verify claims in an AI-assisted draft?
Run every draft through three named passes, each with a defined input and output. The procedure takes about twenty minutes on a 1,500-word piece and catches the failures that survive a normal read-through.
Pass one, claim extraction. Input: the finished draft. Output: a numbered list of every factual assertion, with the sentence it came from. Split compound sentences. "Revenue grew 40 percent after the rebrand, and churn fell" is two claims, not one.
Pass two, source attachment. Input: the numbered claim list. Output: the same list with a source URL or document name next to each claim, or the word UNATTACHED. Do not judge truth here. Only record whether a source exists.
Pass three, verdict assignment. Input: the sourced list. Output: each claim marked supported, refuted, or not enough information. Read the source before you mark it. A claim marked supported means the source states it.
A first-party dataset published by ForgeRank scored 29 drafts and recorded 2,839 claims. Of those, 51 were judged to need a source and 38 had none attached. The gap between 51 and 38 is where publishers get caught: the claim needed a citation, and nobody noticed the citation was missing.
The boundary on pass three matters more than the procedure itself. Verifying a claim against a source establishes that the source says it. It does not establish that the claim is true. A press release states its own numbers, so a claim sourced to that press release passes pass three while remaining unverified in any meaningful sense.
What does a full verification pass look like on one claim?
Take a single sentence from a draft: "Google's March 2024 core update aimed to reduce low-quality, unoriginal content in search results by 40 percent."
Pass one extracts it as one claim. Pass two checks for a source and finds none attached, so it gets marked UNATTACHED. Pass three cannot proceed without a source, which means the claim blocks publication until you resolve it.
You search, and Google Search Central's own documentation of the March 2024 core update states the 40 percent figure as the aim. You attach that URL. Pass three now marks the claim supported, because the source states it.
Notice what changed. The claim moved from unpublishable to publishable without any judgment about whether Google hit 40 percent. The source establishes the stated aim. Whether the update achieved it is a separate claim that needs separate sourcing. Publishers who treat a passed claim as a proven fact skip that distinction, and the correction request arrives six weeks later.
What most people get wrong about AI fact checking
The misconception is that a high accuracy score means you can trust the tool's verdicts. The ForgeRank dataset shows why source attachment, not verdict quality, drives measurable output quality. Ten drafts scored flat at 4.0 across all four dimensions. Ten others held 4.0 on three dimensions with source compliance at 7.0, and those came out at 4.8. Source attachment was the only dimension that moved a headline score on its own.
The correction: treat the model as a claim extractor and a first-pass verdict engine, not as an authority. It is good at finding the assertions in your draft. It is unreliable at deciding whether they hold.
The misconception persists because accuracy numbers are easy to quote and the abstention rate is easy to hide. A vendor reporting 89 percent accuracy without the certainty rate is reporting the flattering half of a trade-off.
Google's March 2024 core update stated its aim as reducing low-quality, unoriginal content in search results by 40 percent, according to Google Search Central. Unattached claims are a direct path into that bucket.
Key takeaways
- Selective accuracy and certainty rate move in opposite directions, so a model's accuracy figure is meaningless without the abstention rate beside it.
- Llama-7B committed on up to 88 percent of claims at roughly 60 percent selective accuracy, which makes confident small-model output the riskiest input in a verification stack.
- The ForgeRank published dataset recorded 2,839 claims across 29 drafts, with 51 judged to need a source and 38 carrying none.
- Source attachment was the only scoring dimension that moved a headline quality score on its own, according to the ForgeRank methodology page.
- Verifying a claim against a source proves the source says it, not that the claim is true.
Frequently asked questions
What is the difference between selective accuracy and certainty rate?
Selective accuracy measures how often a model is correct on the claims it chose to answer. Certainty rate measures how often it answered at all. A model can score high on either one by sacrificing the other, which is why the Scientific Reports evaluation reports both.
Which AI model is the most accurate fact checker?
The Scientific Reports evaluation does not rank models into a single winner. GPT-4o reached up to 89 percent selective accuracy, GPT-4 reached 81 percent, and o1-preview reached 84 percent with a 72 percent certainty rate. Gaps were largest for non-English languages and Global South claims.
Why do smaller AI models give more confident answers?
Smaller models carry weaker calibration about what they know. The Scientific Reports authors describe the pattern as Dunning-Kruger-like: Llama-7B gave definitive responses on up to 88 percent of claims while its selective accuracy sat around 60 percent.
Can I trust an AI verdict of supported?
Only as far as the source behind it. A supported verdict means the cited source states the claim. It does not mean the claim is true, and it does not survive a source that is itself promotional or self-reported.
How many claims in a typical draft go unsourced?
The ForgeRank published dataset scored 29 drafts and recorded 2,839 claims. Of those, 51 were judged to need a source and 38 had none attached. That is roughly three unsourced claims per draft that needed one.
The pass you can run tonight
Open your most recent published piece and run pass two on it. List every factual assertion in a document, put a source next to each one, and mark the rest UNATTACHED. Count how many are unattached. If the number is above zero, you have claims live on the internet that you cannot defend, and the fix takes less time than reading this article did.