How to use this section
AI risk is not limited to dramatic future scenarios. Present systems can invent details, expose information, amplify unfair patterns, create misleading media, or shift responsibility onto people who cannot meaningfully challenge the result. These pages focus on practical safeguards.
Fluency is not evidence
A polished explanation can contain an invented source, incorrect number, or false premise. Important claims require independent verification.
Average accuracy can hide harm
Error types and consequences should be examined for relevant groups and situations, not compressed into one overall score.
Transparency must be useful
A technical description is not enough. People need to know when AI matters, what information influenced an outcome, and how to correct or appeal it.
Accountability lasts beyond launch
Data, users, threats, products, and operating conditions change. Named owners need monitoring, incident response, rollback, and retirement plans.