Bias Screening

Detect editorial bias across 6 categories with severity classification and neutral rewriting patterns.

Overview

Neutral reporting is not the same as centrist positioning. Neutral means presenting facts without editorial opinion, loaded language, or framing that privileges one interpretation. Centrist means taking a position in the middle of a political spectrum — which is still a position. This skill focuses on neutrality: the absence of bias, not the presence of a particular viewpoint.

Core Value: Systematic bias detection that catches patterns human editors miss, with actionable neutral rewrites rather than vague "this seems biased" feedback.

Core Capabilities

• 6 Detection Categories — Loaded language, partisan framing, false balance, editorializing-as-reporting, selection bias, attribution asymmetry
• Severity Classification — Flag (minor, note for awareness), Warning (moderate, should fix), Block (severe, must fix before publication)
• Neutral Rewriting Patterns — For each flagged passage, provide a neutral alternative
• Neutral vs Centrist Distinction — Explicitly differentiates neutral reporting from centrist positioning
• Loaded Terms Dictionary — Reference dictionary of common loaded terms by domain with neutral alternatives

Quick Start

1. Read the article or poll to screen
2. Scan for each of the 6 bias categories
3. Classify each finding by severity
4. Provide neutral rewrite for each flagged passage
5. Report findings with overall assessment

6 Detection Categories

1. Loaded Language

Words or phrases with emotional connotation that substitute for neutral description.

Loaded	Neutral
"slashed" (budgets)	"reduced" or "cut"
"reckless spending"	"the spending proposal"
"crackdown"	"enforcement action"
"controversial" (without context)	[describe the actual disagreement]
"admit" (implies wrongdoing)	"said" or "stated"

See references/loaded-terms-dictionary.md for the complete dictionary.

2. Partisan Framing

Presenting facts through a lens that favors one political or ideological position.

Indicators:

• Using one side's terminology exclusively ("pro-life" vs "anti-abortion" vs the neutral "abortion opponents")
• Framing a policy by its stated goals (favorable) rather than its mechanisms (neutral)
• Presenting one side's argument in detail and the other in summary

3. False Balance

Giving equal weight to unequally supported positions, implying they're equally valid.

Indicators:

• "Some scientists say X, while others say Y" when Y has overwhelming consensus
• Quoting one expert per "side" when one side has thousands of experts
• Using "debate" or "controversy" when the scientific community is in consensus

Note: Not all balance is false balance. When legitimate disagreement exists among qualified experts, presenting multiple perspectives is good journalism.

4. Editorializing-as-Reporting

Inserting opinion or judgment into what appears to be factual reporting.

Indicators:

• "Unfortunately, the policy..." (value judgment)
• "The smart move would be..." (prescription disguised as reporting)
• "Clearly, this shows..." (interpretation presented as fact)
• "Not surprisingly..." (implies something should be obvious)

5. Selection Bias

Choosing which facts to include or exclude in a way that supports a narrative.

Indicators:

• Reporting only positive or only negative aspects of a topic
• Omitting relevant context that complicates the narrative
• Cherry-picking data points or time ranges

Note: All editorial selection involves choices. This category flags patterns where the choices consistently favor one narrative.

6. Attribution Asymmetry

Treating sources differently based on which "side" they represent.

Indicators:

• "Experts say X" vs "Critics claim Y" (experts vs critics implies credibility difference)
• Named, credentialed sources for one side, anonymous "some say" for the other
• "According to" (neutral) for one side, "alleges" (skeptical) for the other

Severity Classification

Severity	Meaning	Action
Flag	Minor bias indicator. Could be incidental.	Note for awareness. Fix if easy, skip if it would make prose awkward.
Warning	Moderate bias. Pattern of framing or loaded language.	Should fix before publication. Neutral rewrite provided.
Block	Severe bias. Editorializing disguised as reporting, or systematic one-sided framing.	Must fix before publication. Passage cannot go out as-is.

Neutral vs Centrist

Neutral	Centrist
"The bill would increase spending by $50B"	"The bill takes a moderate approach to spending"
"Supporters say X. Opponents say Y."	"The truth is somewhere in the middle"
Reports facts without evaluation	Evaluates and takes a middle position
No position	A position (the middle one)

Key principle: Neutral reporting lets the reader form their own view from the facts. Centrist reporting forms a view for the reader — it just happens to be a middle-ground view. This skill targets neutrality, not centrism.

Best Practices

• Screen after transformation, not during — get the facts right first, then check for bias
• Apply to both articles and polls (poll options are a common source of bias)
• When flagging, always provide a neutral rewrite — "this is biased" without an alternative is unhelpful
• Some loaded terms are context-dependent — "crisis" is loaded for a policy disagreement but neutral for an actual emergency
• Check attribution patterns across the entire article, not just individual sentences

Reference Guides

• loaded-terms-dictionary.md — Dictionary of common loaded terms organized by domain with neutral alternatives

Integration

Core skill for fact-checker agent. Also consumed by editorial-writer (awareness during drafting) and poll-writer (balance checking for poll options).