How we score credibility
We believe a credibility tool should show its working. This page explains what we assess, how a score is reached, and where the method falls short. We will keep it updated as the product evolves.
What we assess
Each article is broken into its key factual claims. We then weigh a number of signals, including:
- Evidence and sourcing. Are claims supported by named, checkable sources, or asserted without backing?
- Consistency with the wider record. Do the claims align with reliable, independently reported information?
- Tone and framing. Is the language measured and specific, or emotive, vague or designed to provoke?
- Context and completeness. Does the piece give the context a reader needs, or does it omit material facts?
- Transparency. Are corrections, dates and authorship clear?
How the score is produced
An AI model based on Anthropic's Claude reads the content and assesses each claim against the signals above, then summarises them into a headline credibility score and a per-claim breakdown. Every result includes a plain-English explanation of what drove it, so you are never asked to trust a number on its own.
Deep claim verification
Any single claim can be checked in more depth. This runs a fresh review backed by a live web search, returning an updated verdict, the key evidence found, and links to the sources cited. It is a focused second look at one claim rather than a re-score of the whole article.
What the scores mean
- Higher scores indicate claims that are well sourced, consistent with the wider record and presented in context.
- Lower scores flag weak sourcing, missing context, misleading framing or claims that conflict with reliable information.
- A score is a prompt to think, not a ruling. Two reasonable people may weigh the same evidence differently.
Limits of the method
Automated analysis can make mistakes. It may misread satire, struggle with very new events where reliable reporting is thin, or reflect gaps and biases in its training. It is not a substitute for primary sources or professional advice. Always verify important claims independently. See our AI Disclaimer for more.
If you think a result is wrong, please tell us. You can challenge a result and we will review it.