Fact-check

How AI-assisted fact-checking works (and when it doesn’t)

A practical look at claim extraction, retrieval, ranking, and review — including the cases where the system should decline to answer.

AI-assisted fact-checking is often described as a pipeline that turns a claim into a verdict. That description is technically accurate and practically misleading. The hard part is not generating a verdict. It is knowing when not to.

This piece walks through the five stages we specify, the failure modes at each stage, and the cases where a human must take over.

Stage 1: Claim extraction

The system separates atomic claims from rhetoric, opinion, and prediction. “The budget passed 7–4” is a claim. “The council was reckless” is an evaluation. “The plan will fail” is a prediction. Mixing them produces verdicts that look precise and mean nothing.

Extraction fails when a claim is implicit, when it spans several sentences, or when the speaker is being deliberately vague. A good system surfaces the extracted claim for the reviewer to edit before retrieval begins. An extracted claim that misstates the original is worse than no extraction.

Stage 2: Retrieval

Candidate evidence is gathered from the web, official records, wire services, and the publisher’s archive. Retrieval is where most systems quietly fail: they find plausible text, not authoritative text.

We specify source classes rather than a single search index. A court filing, a legislative transcript, and a news report are not interchangeable. The retrieval layer records what it looked for, what it found, and what it could not access.

Stage 3: Ranking

Sources are ranked for proximity, independence, recency, and reliability. Proximity asks whether the source is the original document or a summary. Independence asks whether two sources are actually one source repeated.

Ranking is a judgement, and it should be visible. The reviewer should see the sources that were rejected and why. Hiding the rejected sources makes the evidence chain look cleaner and the verdict less defensible.

Stage 4: Verification

The claim is compared with the best available evidence and assigned one of seven IFCN-aligned verdicts: True, Mostly True, Mixed, Misleading, False, Unverifiable, or Satire.

This is where false precision creeps in. A model that outputs “0.87 false” invites a confidence that the evidence does not support. We prefer a verdict plus an explicit confidence band, and we tune the system to choose Unverifiable when evidence is thin or conflicting.

Stage 5: Review

False and Misleading verdicts require a named human reviewer. The reviewer sees the claim, the sources, the passages relied on, the rejected sources, and the model version. They can accept, edit, or override. The override is recorded with a reason.

Review is not a rubber stamp. If reviewers accept 99% of suggestions, either the system is extraordinary or the review is not real. We would publish acceptance and override rates as part of the transparency report.

When it doesn’t work

Breaking news

During the first hours of an event, the best evidence may be a single official statement that later changes. A system that retrieves quickly can produce a confident verdict on incomplete information. In those cases the correct output is “too early to verify,” not a verdict.

Local and minority-language claims

Retrieval coverage is uneven. A claim about a local council in a language with little indexed material may be true and unverifiable by the system. That is a retrieval failure, not evidence of falsity.

Claims about intent

“They knew” or “they deliberately misled” are claims about state of mind. Documents rarely settle them. The system should decline to verdict them and route to reporting.

Numbers without context

A number can be accurate and misleading. “Crime rose 40%” may be true in a small district with a low base. Verification needs context, not just a source.

Manipulated media

Fact-checking a claim that relies on a synthetic image requires detection and provenance, not just text retrieval. The evidence chain must include the media analysis.

What good looks like

A good AI-assisted fact-check is boring. It shows the claim, the evidence, the reasoning, and the reviewer. It says “we could not establish this” when that is true. It does not produce a verdict because a verdict was requested.

The measure of the system is not how many claims it can label. It is whether a reviewer can defend each label six months later.

MediaMeter Editorial

Written by the MediaMeter editorial team. We publish our methods, limitations, and corrections.

https://mediameter.eu/resources/blog/how-ai-assisted-fact-checking-works/

Related reading

Newsletter

Monthly notes on journalism quality.

One email a month. Methods, limitations, and what we are learning. No product spam.