Traditional fact-checks arrive hours after a lie has already done its damage. Sometimes days. By then, the clip has been shared, the talking point has calcified, and the correction lands like a whisper at a concert. A college student named Risha Panigrahi just shipped a solution from a university NLP research project. InTruth is a free Chrome extension that listens to live political video, transcribes every word, identifies factual claims, searches the web for evidence, and delivers a verdict before the next sentence starts. No subscription. No newsroom budget required.
How It Actually Works
Three services chain together to catch claims faster than a politician can pivot.
Running under the hood is a pipeline that’s elegantly simple. Deepgram handles live transcription. Serper searches the open web for relevant sources. Then Claude — Anthropic’s AI model — evaluates the claim using only that freshly retrieved data, not its training memory. That distinction matters. The system judges what was said against what exists on the web right now, not pre-baked answers from months ago. Users assign speakers at the start, so during a debate, you know exactly who said what. “The lie gets caught in the same breath it’s told.”
Watching a Senate hearing on YouTube? Install InTruth from the Chrome Web Store, open the stream, click start, assign speakers. Verdicts appear as overlay labels with clickable source links — a HUD for reality-checking, built into your browser.
- Free to install; bring-your-own-key model (you supply API keys for Deepgram, Serper, and Claude)
- Works on any video playing in an active Chrome tab
- Verdicts labeled: True, False, Misleading, or Unverifiable
- Source links included with every verdict
What It Can’t Do (Yet)
Impressive engineering — but the receipts on accuracy are still missing.
The Washington Post tried something similar with Truth Teller, matching speech against a pre-built database of existing fact-checks. InTruth’s live web search pipeline is a genuine step forward — like trading a paper map for GPS. Honest caveats remain, though. No published accuracy metrics exist. Performance depends on network conditions and transcription quality. The project is still in active development, per community descriptions, meaning error rates and edge-case handling aren’t yet fully documented. Complex policy claims may simply land in “Unverifiable.”
A student’s NLP research project now sits on the Chrome Web Store, doing something major platforms haven’t managed. That’s either exciting or embarrassing for the industry — probably both. If you’re tired of waiting two news cycles for someone to confirm what you just heard was nonsense, the 30-second install is worth your time.
