Methodology
How Bias Map analyzes news coverage using AI and human oversight.
1. Cluster-First Analysis
Traditional bias checkers often assign a static "Left/Right" score to an entire outlet. BiasMap takes a different approach: Cluster-First Scoring.
We group articles about the same specific event into a "Story Cluster". Our AI then analyzes the entire cluster to identify the unique themes and disagreements present in that specific story.
2. Dynamic Axes
For each story, we identify 3 to 5 distinct axes of divergence. These are not pre-defined; they emerge from the coverage itself.
For example, a story about a housing crisis might be mapped on:
Axis 1: Attribution
Individual Responsibility (e.g., "Bad Landlords")
vs.
Systemic Failure (e.g., "Market Forces")
Axis 2: Focus
Economic Impact (e.g., "Property Values")
vs.
Human Cost (e.g., "Displacement")
3. Global Metrics
Beyond the story-specific axes, we score every article on three global metrics:
- Tone: From Hostile (-1.0) to Supportive (+1.0). This measures the emotional language used.
- Complexity: From Sensational/Surface (-1.0) to Nuanced/Detailed (+1.0). This measures the depth of analysis and context provided.
- Entity Sentiment: We track how key figures (politicians, organizations) are portrayed across the cluster.
4. The Fact Spine
To ground our analysis in reality, we generate a "Fact Spine" for each story cluster.
- Confirmed: Claims reported by multiple distinct outlets.
- Disputed: Claims where outlets explicitly contradict each other.
- Unverified: Claims reported by only a single source.
5. Omission Detection
Bias is often defined by what is not said. Our system compares the coverage of all outlets to identify unique perspectives or facts that were present in some reports but missing in others.
We list these as "Omissions" to help you see the full picture that you might miss by reading only one source.
6. Academic Framing Frameworks
Our analysis is grounded in established media and communication theory. We apply five complementary academic frameworks to produce rigorous, multi-dimensional framing analysis:
Entman (1993) — Core Framing Taxonomy
The foundation of our framing analysis. Each article is analyzed for four dimensions: Problem Definition (what is the issue?), Causal Interpretation (who is responsible?), Moral Evaluation (what judgment is implied?), and Treatment Recommendation (what solution is suggested?).
Scheufele (1999) — Frame Building & Setting
Examines how journalists construct frames through source selection, emphasis, and exclusion. Helps us identify whether framing reflects newsroom norms, elite influence, or audience expectations.
Gamson (1989) — Interpretive Packages
Identifies the cultural devices used in coverage: metaphors, catchphrases, moral appeals, and visual imagery. These devices reveal the deeper “interpretive package” an article deploys to make the story comprehensible.
Chong & Druckman (2007) — Emphasis Framing
Distinguishes between equivalence framing (same facts presented differently) and emphasis framing (selective facts highlighted). Reveals how outlets shape opinion by choosing what to foreground.
Papacharissi (2013) — Affective Publics
Analyzes the emotional resonance of framing choices, particularly relevant for stories that spread through social media. Examines whether framing is designed to mobilize, create solidarity, or provoke outrage.
7. Limitations & Disclaimer
Position ≠ Quality. An article's position on the map describes its framing, not its accuracy or quality. A "systemic" framing is not inherently better than an "individual" one; they simply serve different purposes.
AI + Human. While we use advanced LLMs to process large volumes of text, all clusters and axes are subject to human review. AI can miss sarcasm or nuance, so treat these visualizations as a guide, not absolute truth.