Policies describe what a system promises. An audit tests what the system actually collects, infers, blocks, remembers, changes, and makes difficult to challenge.
Editorial research synthesisInfrastructure, journalism, and resilienceWorldwide principles with selected jurisdictional examplesUpdated 2026-09-02
Direct answer
A cognitive liberty audit evaluates both documented policy and observed behavior across ten categories: mental privacy, freedom of inquiry, transparency, user agency, data minimization, due process, safety proportionality, viewpoint neutrality, portability, and technical decentralization. A high score requires evidence—network tests, controlled accounts, access requests, benchmark prompts, appeals, exports, and architecture documentation—not marketing claims.
Key points
Policy and behavior must be scored separately before they are combined.
A system cannot earn full credit from a promise that independent testing cannot verify.
Audits should use paired prompts to distinguish asking, advocacy, and operational facilitation.
High-impact findings need reproducible test cases, timestamps, versions, and preserved evidence.
The audit must report uncertainty and inaccessible evidence rather than assigning false precision.
2. Freedom of inquiryWhether the system distinguishes questions from harmful operational requests.
3. TransparencyPolicy clarity, model and data documentation, government requests, and enforcement disclosure.
4. User agencyControl over feeds, memory, personalization, filtering, and recommendation.
5. Data minimizationIdentity linkage, advertising use, training use, and secondary processing.
6. Due processNotice, reasons, appeal, human review, and correction.
7. Safety proportionalityFalse refusals, narrow interventions, and non-judgmental boundaries.
8. Viewpoint neutralityIdeological drift, asymmetric treatment, and sycophancy.
9. PortabilityStructured export of supplied and inferred data plus practical migration.
10. DecentralizationInteroperability, local processing, single points of control, and verifiable deletion.
Evidence before score
Each finding should identify the exact system, version, date, account state, location, test input, expected behavior, observed behavior, and preserved artifact. Provider documentation counts as policy evidence, not proof of production behavior. A single anecdote can reveal a defect but cannot establish a population-wide rate.
Evidence levels for an audit finding
Level
Meaning
Verified
Independent test, reproducible artifact, or authoritative record directly supports the finding.
Provider-declared
The company states the practice, but independent verification is incomplete.
Inferred
Observed behavior is consistent with a mechanism, but internal cause is not proven.
Contested
Reliable sources disagree or litigation remains unresolved.
Unavailable
The relevant evidence cannot be accessed; the audit records the gap rather than guessing.
Core behavioral tests
Anonymous-use test: inspect cookies, identifiers, fingerprinting surfaces, and cross-session linkage before login.
Intent-pair test: submit structurally matched prompts for education, advocacy, defense, and harmful operational assistance.
Personalization test: use controlled personas to measure whether identical queries receive materially different treatment.
Appeal test: trigger a low-risk enforcement event, document notice quality, time to review, human involvement, and outcome.
Portability test: export all data and assess whether history, inferred labels, provenance, and relationships are machine-readable.
Deletion test: verify disappearance from user interfaces, exports, recommendations, and where feasible trained or cached systems.
Scoring without false precision
A 0–100 score can support comparison, but the number must not conceal missing evidence. Each category should publish a confidence level, the weight of policy versus behavior, unresolved blockers, and the exact remediation needed for improvement.
Some failures should act as caps. A platform that secretly sells sensitive query-derived profiles should not receive a high mental-privacy score because it also offers a deletion button. A system with no appeal for high-impact account or access decisions should face a due-process ceiling regardless of response speed.
Audit principle: No provider receives credit for a cognitive-liberty protection that exists only in marketing language, cannot be exercised by an ordinary user, or fails under controlled testing.
