AI Safety Watch covers a field where extraordinary claims are common, technical evidence can be difficult to evaluate and reasonable experts often disagree. Our standards are designed around that reality.
Evidence first
We distinguish demonstrated capabilities from forecasts, scenarios and speculation. When a company makes a claim about its own system, we identify the source of the claim and seek independent context when it is material to the story.
Disagreement is part of the story
AI safety is not a single school of thought. Researchers disagree about timelines, threat models, technical safeguards, regulation and even the vocabulary used to describe risk. We represent meaningful disagreements rather than compressing them into a false consensus. We do not manufacture balance where the underlying evidence is lopsided.
Original reporting
Our strongest work is based on interviews, papers, technical documents, incident reports and other primary material. We link to source material wherever practical so readers can inspect the evidence themselves.
Independence
AI Safety Watch is an independent publication. Founder and editor Sascha Brodsky also writes for other publications and organizations. Work first published elsewhere is clearly identified and linked to the original publication. Outside affiliations do not determine AI Safety Watch’s conclusions or coverage.
Language
We avoid treating anthropomorphic language as evidence. Words such as “wants,” “knows,” “decides” or “escapes” can sometimes be useful shorthand, but stories should explain the underlying behavior rather than imply motives a system has not been shown to possess.
AI tools
AI Safety Watch may use software and AI tools to assist with tasks such as transcription, search, organization and production. AI-generated material is not treated as a source. Factual claims and quotations must be checked against primary material or reliable reporting before publication.
Corrections
We correct substantive errors promptly and transparently. Readers can find the correction policy on the Corrections page.
AI Safety Watch 