Sourcing standards
Part of the AI Wiki policies
A citation is a promise: follow this link and you will find this fact. These standards govern which links we are willing to make that promise with.
Order of preference
- Primary sources. Peer-reviewed papers and preprints, technical reports, model cards, official documentation, company announcements, regulatory filings, court documents, and benchmark leaderboards for their own results.
- Established technical press and analysis. Reporting and peer analysis with named authors and editorial standards, used especially for events, context, and claims a primary source would self-servingly frame.
- Everything else, sparingly. Personal blogs and social media posts only when the author is the subject or a recognized authority (a lab announcing on X, a researcher explaining their own paper), cited as what they are.
SEO content farms, uncredited aggregators, and AI-generated summary sites are not acceptable sources at any position. Marketing claims are attributed as claims ("the company states"), not repeated as facts.
How citations are wired
- Claims carry inline markers like [3] pointing into the numbered References list at the end of every article, so a specific sentence maps to a specific source.
- Statistics keep their date and origin in the text ("as of March 2026", "according to the v3 technical report") so they age visibly rather than silently.
- Cited links are monitored for rot; when a source goes offline we repoint the reference to an archived snapshot rather than leaving a dead promise.
For contributors
Suggesting an edit that adds a fact? Add the source with it (the editor toolbar has an insert-citation button). Reviewers check suggestions against these standards before publishing; unsourced substantive claims are declined, however plausible they sound.