Mental Privacy
Can a person inquire without unnecessary identity linkage, behavioral surveillance, or indefinite query retention?
Cognitive Liberty Audit
Audit what the system does, not what the policy promises.
A research-derived, experimental framework for evaluating AI systems, search engines, recommenders, and digital platforms across ten separate dimensions of cognitive liberty.
Audit principles
A score should make power more inspectable, not create a new authority badge.
The framework is designed to compare observable practices and expose uncertainty. It does not declare a platform morally pure, legally compliant, or safe in every context.
01Separate policy claims from observable behavior.
02Score categories independently; do not hide trade-offs in one total.
03Prefer ordinary authorized testing over intrusive probing.
04Publish evidence limits, dates, and revision conditions.
Ten-category scorecard
What should be measured?
Each category asks a different question. Strong privacy cannot excuse arbitrary moderation; good appeals cannot excuse unnecessary surveillance.
Freedom of Inquiry
Does the system preserve lawful research, criticism, education, curiosity, and controversial exploration?
Transparency
Can users understand meaningful data practices, ranking/refusal boundaries, and consequential automated decisions?
User Agency
Can users control personalization, memory, recommendations, and important defaults without coercive friction?
Data Minimization
Does the service collect and retain only what is necessary for the stated purpose?
Due Process
Are consequential restrictions explained, reviewable, correctable, and reversible?
Safety Proportionality
Do safeguards target demonstrated harm with the least restrictive effective intervention?
Viewpoint Neutrality
Are rules applied to conduct and quality rather than hidden ideological loyalty tests?
Portability
Can users export, leave, switch, and preserve their own work or memory without lock-in?
Technical Decentralization
Does architecture avoid unnecessary single points of control and make exit or interoperability technically possible?
Evidence method
Four steps before a score.
Document
Capture the public policy, feature scope, user controls, retention claims, and stated safety rationale.
Observe
Use reproducible, authorized test cases to see whether ordinary behavior matches the documented rule.
Compare
Evaluate proportionality, alternatives, errors, recourse, portability, and whether less data could achieve the same result.
Publish limits
State what was provider-reported, independently observable, inaccessible, time-sensitive, or outside the test scope.
No provider rankings yet: QAEDA is publishing the methodology before publishing comparative provider scores. A credible audit needs a repeatable test corpus, dated evidence, authorization boundaries, and enough independent observation to justify the result.
Why no single “cognitive liberty score”?
A total can be useful for comparison, but it can also hide a serious weakness. A service that scores extremely well on portability and poorly on mental privacy should not be able to average the privacy problem away. Category scores remain the primary result.
Why viewpoint neutrality is not “all content must rank equally”
Search and recommendation systems must rank, deduplicate, fight spam, remove unlawful material, and protect users from genuine hazards. The audit question is whether quality and safety rules are sufficiently clear and consistently applied that disagreement itself does not become an invisible disqualifier.
Why decentralization is only one category
Decentralized systems can still be abusive, opaque, or unsafe. Centralized systems can still provide strong privacy and due process. Decentralization is therefore evaluated as an exit and concentration-of-control property, not as an automatic virtue.
Research basis: Cognitive Liberty Audit Framework, Cognitive Liberty Risk Actor Map, and AI System Website Evaluation Research. The public implementation intentionally removes claims of universal standardization or certification that would outrun the evidence.