Know what you're reading,
not just what it says.

Refract analyses any article for bias, funding, purpose and factual accuracy — using three independent AI models, live, as you browse. No single source of truth. No single point of view.

Free to try. No credit card required. · Why I built this

Refract - Know what you're reading, not just what it says | Product Hunt

See it check a real article

This runs automatically — no video file, just the actual interface, looping through a real analysis.

theguardian.com/politics/2026/...
Refract
Ready to analyse this page.
Bias & lean
Purpose & intent
Factual accuracy
Source credibility
Funding & ownership
Narrative pattern

Two ways to check a story

Pick the depth that matches how much the story matters to you.

⚡

Quick Scan

A trust score in about fifteen seconds.

Three models — Claude, GPT‑4 and Gemini — independently score the article for bias, purpose, funding transparency and source credibility. Refract shows you the average, and flags it clearly whenever the models disagree.

Political leanCentre-left
Factual accuracy82/100
Ownership transparency70/100
🔍

Deep Verify

Every claim, checked against a real source.

Refract pulls out each factual claim in the article, searches for the primary source behind it — a government dataset, a court record, an official report — and tells you whether the claim actually holds up, with a link to what it found.

✅ "Crime rose 18.7% in the county" — matches Home Office statistics
⚠️ "Nearly doubled" — source shows a 34% rise, not a doubling
❌ Quote attributed to a spokesperson — no matching source found

Why three models, not one

A single AI carries a single set of blind spots — trained by one company, on one slice of the internet, with one set of assumptions baked in. Asking it to judge an article's bias is really just swapping one opinion for another.

Refract sends every article to Claude, GPT‑4 and Gemini at once. Where all three agree, you can trust that's a real signal. Where they don't, we tell you outright — some stories are genuinely contested, and pretending otherwise isn't honesty, it's false confidence.

Claude 74
GPT‑4 71
Gemini 77
Spread: 6 points — high consensus

Start checking what you read

Free to start. Sign up to get your first 10 credits — on Chrome, or right in your browser on any device.