Media Literacy × AI

Inside an AI Fact-Checker: From Spoken Claim to Verdict

Browser extensions now promise live fact-checking while you watch YouTube. Useful — but only if you understand what's happening under the hood, and where it fails. Feed a claim into the pipeline below and watch each station do its job.

Run the Pipeline

Pick a claim. A cube carries it through four stations: extract → retrieve → score sources → verdict. The verdict lamp at the end turns green (supported), amber (needs context), or red (refuted). Drag to orbit.

drag to rotate
1. Claim extraction — speech-to-text, then an LLM isolates checkable factual statements from opinion.
2. Evidence retrieval — search across encyclopedias, journals, fact-check databases (ClaimReview).
3. Source scoring — weight by reliability, recency, independence of sources.
4. Verdict + citation — supported / refuted / needs-context, with links.
Select a claim to begin

The verdicts, explained

Where AI fact-checkers fail (and how to compensate)

  1. Fresh events: retrieval lags breaking news by hours; early "verdicts" often flip. Treat anything under 24 hours old as provisional.
  2. Checkable ≠ true/false: many viral claims are framing ("the economy is failing") — no pipeline can verdict a vibe. Good tools label these "opinion," bad ones guess.
  3. Circular sourcing: if 50 articles all trace to one press release, that's one source. Station 3 (independence scoring) is the hardest to automate.
  4. Your move — lateral reading: professional fact-checkers open new tabs and check who is behind the source before reading the source. Stanford studies found this single habit beats checklist methods dramatically.
Rule of thumb: use AI fact-checkers as a first-pass radar, not a judge. When a verdict surprises you, click through to the cited primary source before repeating it.
Enjoy this tool? Build your own with Super