← Back to FactRefinery

⚙️ How FactRefinery Works

Full transparency — no black boxes

FactRefinery uses AI to read 100+ news sources every hour, strip the bias, and surface only the facts. No human editors. No injected opinion. Just distilled signal.

Step 01

Source Collection

Every hour, the aggregator fetches articles from over 100 outlets spanning the full political spectrum and every major topic category.

BBCReutersAP GuardianNYTWaPo WSJFox NewsCNN BreitbartDaily Mail ESPNSky SportsBBC Sport TechCrunchArs TechnicaThe VergeWired BloombergFTCNBCMarketWatch VarietyHollywood ReporterDeadline + 40 more
Step 02

Semantic Deduplication

The AI compares headlines and article content to identify stories that cover the same event across multiple outlets. Duplicates are merged into a single FactRefinery article — all original sources are preserved and cited.

Example: If BBC, Fox News, and CNN all report the same development, they become one article with all three sources listed and individually attributed.

Step 03

Bias Stripping

The AI reads all available versions of a story and extracts only the overlapping, verifiable facts — the elements every source agrees on. The following are removed or flagged:

Bias stripping applies to all categories — not just politics. Sports, tech, entertainment, and business coverage carry ideological framing too. FactRefinery strips it across the board.

Step 04

Purity Score

Every article receives a Purity Score from 0 to 100%, reflecting how much the original sources agree on the core facts.

ScoreWhat it means
90–100% Strong source agreement. High factual confidence.
70–89% Most sources agree. Minor differences in detail or framing.
50–69% Meaningful disagreement between sources. Read with caution.
Below 50% Highly contested. Multiple conflicting narratives. Article flagged.
Step 05

Article Generation

The AI writes a neutral summary using only the stripped facts. The model is instructed to:

The output reads as if the story has no political dimension whatsoever — because facts don't have a side.

Step 06

Source Labeling

Every article lists its original sources with political lean labels. These labels are derived from a composite of third-party media bias tracking services and updated periodically. We don't claim they are the final word — you can verify and disagree.

🔴
Hard Left
e.g. Jacobin, The Intercept, Democracy Now
🟠
Lean Left
e.g. Guardian, NYT, WaPo, CNN, NPR
Center / Wire
e.g. Reuters, AP — primarily wire services with minimal editorial framing
🟡
Lean Right
e.g. WSJ, Fox Business, NY Post
🟣
Hard Right
e.g. Breitbart, Daily Wire, Epoch Times

Note: Bias labels are a starting point, not a verdict. Outlets like BBC carry institutional lean not always captured by third-party trackers. We display labels so you can factor them in — not so you'll take them at face value. The raw sources are always linked so you can judge for yourself.

Transparency

What We Do and Don't Do

We don't

  • Use human editors to rewrite stories
  • Inject our own opinion or agenda
  • Censor stories based on political content
  • Use clickbait headlines
  • Hide which AI model or sources we use

We do

  • Show every original source with a direct link
  • Label speculation and contested claims
  • Apply the same standards to all topics equally
  • Update bias labels as outlets evolve
  • Publish this page so you can audit us