Inference, validation, appeal
AI Surveillance
AI surveillance turns ambiguity into rankings, clusters, summaries, matches, flags, and risk scores. Each output is a claim that needs source data, validation, human authority, appeal, and deletion.
Publication information
Demand A Public-Power Model Card
Ask what the system does, what it does not do, what data it uses, who approved it, where it is deployed, what thresholds trigger action, and which decisions humans can override.
Validate Against Real Harm
Look for false positives, dialect or neighborhood bias, protected-activity chilling effects, stale data, demographic skew, automation bias, and whether claimed benefits justify civil-liberties burdens.
Human Review Must Have Power
A reviewer needs source access, time, authority to disagree, documented reasons, override power, escalation routes, and a duty to correct downstream records when the model is wrong.
Procurement Is The Audit Trail
Request model purpose, data sources, validation reports, bias assessments, human-review rules, audit rights, retention terms, appeal policy, vendor lock-in, renewal clauses, and termination rights.
Do Not Let AI Launder Uncertainty
Summaries, entity matches, and risk rankings should preserve uncertainty rather than smoothing it away. Public pages should disclose what is confirmed, inferred, disputed, stale, or unknown.