How we work, and how we check ourselves.

What counts as in scope

The project looks at factual claims with a plausible public consequence: elections and civic processes, government and public policy, law and regulation, public health and safety, emergency response, conflict, economic claims with public consequence, science and the environment, and scams or fraud. Sports scores, entertainment chatter, personal conversation and jokes are out of scope.

The rules that decide scope name subjects, never sides. They can say a claim is about elections; they cannot say which party it favors.

Testing for bias

Claiming neutrality is easy. We test for it, with two instruments.

• A balanced test set. Example claims are written in opposed pairs — left- and right-associated, favorable and critical of government, favorable and critical of corporations, for and against a policy, supporting and challenging institutional consensus. The scope rules should admit each side of a pair at comparable rates.
• Mirrored claims. Sets of claims share the same structure and differ only in a politically or institutionally salient detail: the country, the faction, the institution, or the direction of an accusation. A mirrored set should be treated alike.

A material gap between opposed pairs, or a mirrored set treated differently with no evidence-based reason, counts as a finding to document and fix. Where the difference is justified — a nationwide recall matters more than one shop's product — the reason is recorded rather than counted as bias.

The same tests will be applied to each new part of the method as it is built. Results will be published with their limitations stated alongside them.

How claims are assessed

What kind of statement a claim is gets recorded separately from whether it is true. A directly verifiable figure, a contested factual claim, a legal or historical interpretation and a moral judgment each call for different handling, and a disagreement about interpretation is not treated as misinformation.

An assessment records the evidence for and against a claim and its confidence. The possible outcomes include supported, contradicted, disputed, insufficient evidence — and unknown, which is a valid result rather than a failure. No single AI model decides whether a claim is true, and consequential conclusions remain reviewable by people.

What we will publish

• The scope rules and how they perform.
• Bias-test results, with their limitations.
• The evidence behind published assessments.
• Corrections, when we get something wrong.

Questions or challenges to the method are welcome. Get in touch.