How BiasRubric Works
A transparent look at our 4-model consensus methodology, bias scale, and data practices.
The 4-Model Consensus
Rather than trusting any single AI view of political bias, BiasRubric runs every article through four leading language models simultaneously: Claude (Anthropic), ChatGPT/GPT-4o (OpenAI), Gemini 1.5 Pro (Google), and Grok-2 (xAI). Each model independently scores the article, then the four scores are averaged into a consensus.
The consensus approach reduces the risk that any one model's training biases dominate the result. When models diverge significantly, that divergence itself is informative — and we display each model's score individually so you can judge.
The Bias Scale
Strongly Left
Strongly progressive framing, loaded left-coded language, exclusively left-leaning sources.
Moderately Left
Left-leaning perspectives given prominence; centrist and right sources underweighted.
Center / Neutral
Balanced sourcing, neutral language, multiple perspectives represented proportionally.
Moderately Right
Right-leaning perspectives given prominence; centrist and left sources underweighted.
Strongly Right
Strongly conservative framing, loaded right-coded language, exclusively right-leaning sources.
What AI Models Analyze
Each model evaluates:
- Loaded language — emotionally charged words favoring one side
- Source selection — which experts, officials, or studies are cited
- Framing — how issues are contextualized and which aspects are foregrounded
- Omission of perspectives — which viewpoints are absent from the piece
- Headline tone — whether the headline matches the body neutrality
- Attribution patterns — who gets authoritative quotes vs. skeptical framing
Data Policy
We never store article text.
Only bias scores, model reasoning summaries, and the article URL (if provided) are stored. The full article body is analyzed in memory and immediately discarded. Your reading is private.
Honest Limitations
AI models are trained on human-generated data and may reflect biases present in that data. BiasRubric is a tool for prompting critical thinking — not an authoritative arbiter of truth.
- Scores can vary run-to-run; cache results are served for 24 hours for consistency
- Satire, opinion pieces, and highly technical reporting may be scored less accurately
- Non-English content is analyzed but with lower accuracy
- The 1–5 scale is a simplification; real bias exists on multiple axes