A sentence can pass every fact-check and still lead you to believe something false. TruthForge reads the news you're reading, does the research, and hammers out the details.
Runs in your browser. No backend. Anthropic Claude or Google Gemini.
Whether a sentence is literally accurate and whether it's engineered to mislead are two different reads. TruthForge keeps them apart — a true statement can still be built to deceive, and it will say so.
“…passed with overwhelming bipartisan backing…”
Technically true“…passed with overwhelming bipartisan backing…”
Plants: near-unanimous support. Actually: a 2-vote margin.
Watch it work, then read the critique — findings streamed as white paper cards on the dark stage, each one grounded in a real source.
“…with overwhelming bipartisan backing…”
Actually: it cleared by a 2-vote margin, with most of the opposition party against it.
high confidence“…passed with 60 Senate votes…”
Actually: the recorded tally was 51–49, not 60.
medium confidenceTruthForge runs entirely in your browser and calls your chosen model directly — Claude or Gemini — with the key you paste once. Nothing routes through a TruthForge server.
Click the toolbar icon. The panel docks next to the page and reads the full article text automatically.
Your model reads the piece, runs live web searches to ground every correction, and writes a critique across eight lenses.
Each finding names the exact phrase, the belief it plants, the grounded correction, and the lean balance of its sources.
Not an open taxonomy — a deliberately fixed set of media-criticism categories. A lens with nothing to critique is dropped entirely; the checklist never shows empty rows.
Loaded word choice and slant that tilt the read before you reach a single fact.
Who gets quoted, who's left unheard, and whose account is treated as the default.
The comparison or history that would change how a number or claim actually lands.
Claims stated as settled that the reporting never actually verified.
Implying a cause or an intent the evidence on the page doesn't support.
What's true, relevant, and quietly left out of frame.
Persuasion moves that win by construction, not by argument.
Invented detail or doctored material presented as real.
These aren't marketing claims — they're constraints the engine is built around.
It doesn't issue final verdicts on contested claims. It reasons transparently — it doesn't rule from on high. A finding that holds up is stated as confidently as one that manipulates.
Whether something is literally accurate and whether it's built to mislead are independent reads — never collapsed into one true/false badge.
Each finding carries its own calibrated confidence — never an article-level score, percentage, or reliability badge. No faked certainty, no hallucinated authority.
A header control always explains what the eight-category framework is and where it structurally falls short — the same discipline it applies to the article.
Free to install. You bring your own API key — Anthropic or Google. No subscription, no backend, no data leaving your browser to us.
Add to Chrome — freeRequires an Anthropic or Google AI Studio API key. Each analysis costs tokens.