Educational companion dossier · Fact, interpretation, lived experience, clinical education, fiction, and mechanics are labeled separately. Scope & safety
PsychologicalWar.orgHistory before hype.

COGNITIVE LIBERTY · SECTION 05 OF 07

Safety needs procedure

Moderation, Suppression, and Appeal

Distinguishes legitimate harm prevention from opaque suppression, examines language and context errors, and proposes procedural safeguards for automated moderation.

A · DEFINITION & SCOPE

What this section means—and what it does not

Definition

Content moderation includes rules and interventions used to address illegal material, credible threats, exploitation, fraud, harassment, privacy violations, and other concrete harms. Suppression describes the rights risk when lawful expression is invisibly or discriminatorily restricted without adequate notice, context, proportionality, or remedy.

Outside scope

A rights-preserving framework is not content absolutism. It does not require platforms to amplify every post or expose victims to threats, child exploitation, nonconsensual intimate imagery, or targeted abuse.

B · WHY IT MATTERS

The rights and governance problem

Automated systems operate at a scale where small error rates can affect many people. Dialect, reclaimed language, political context, documentation of abuse, and identity terms can be misread. During conflict or crisis, asymmetrical errors can distort whose testimony, grief, or evidence remains visible.

ExpressionEqualityContextual reviewEffective remedyResearch access

C · KEY DISTINCTIONS

Do not collapse unlike things

Policy rule

The published norm describing what is prohibited or restricted.

Detection

Automated or human identification of potentially governed material.

Enforcement

Removal, restriction, demotion, labeling, monetization change, or account action.

Remedy

Notice, reason, human review, appeal, restoration, and correction of downstream effects.

D · CLAIM REGISTER

Three bounded claims with twenty evidence stages each

Each claim preserves the difference between an artifact, its availability, audience exposure, belief, conduct, and downstream effect. “Not assessed” is retained rather than converted into an implied result.

CL-013NORMATIVE

Normative proposal

Moderation can be legitimate when addressing concrete harms, but restrictions should be specific, proportionate, consistently governed, and open to effective remedy.

Source scope
This is a rights-preserving governance principle, not a claim that all platforms or jurisdictions use the same standard.
Do not infer
Do not present the framework as requiring amplification of illegal or seriously harmful content.
Competing explanations
Where outcomes are discussed, ordinary ranking changes, user choice, market incentives, security requirements, model error, institutional process, and non-AI causes remain possible unless claim-specific evidence excludes them.
Affected-person context
Public appeals and policy review show concrete remedy and proportionality questions; lawful safety purposes remain separately recognized.
Rights and privacy implications
Potential implications include freedom of thought or expression, mental and data privacy, equality, autonomy, identity, notice, due process, and effective remedy; legal scope remains jurisdiction-specific.
Correction trigger
Revise after legal review, affected-community review, or new evidence on remedy effectiveness.

Claim-specific sources

WIP.54 evidence overlays

Review all twenty evidence stages
Artifact or event existence
OWNER_SUPPLIED NORMATIVE PROPOSAL PRESERVED
Content status
NORMATIVE POLICY OR DESIGN POSITION
Coordination
NOT APPLICABLE
Actor identity
PROJECT EDITORIAL POSITION IDENTIFIED
Sponsorship or direction
OWNER SUPPLIED AND EDITORIALLY INTEGRATED
Intent
PUBLIC EDUCATION AND GOVERNANCE ADVOCACY
Output
PUBLIC PRINCIPLE OR PROPOSAL
Distribution
WEBSITE PUBLICATION
Availability
PUBLICLY AVAILABLE AFTER DEPLOYMENT
Reach
NOT MEASURED
Exposure
NOT MEASURED
Attention
NOT MEASURED
Recall
NOT MEASURED
Comprehension
NOT MEASURED
Credibility
NORMATIVE; NOT PRESENTED AS SETTLED LAW
Belief or attitude
NOT CLAIMED
Intention
NOT CLAIMED
Behavior
NOT CLAIMED
Operational outcome
NOT CLAIMED
Strategic effect
NOT_ESTABLISHED
Questions for specialist review
  1. Is the claim phrased no more strongly than the cited sources support?
  2. Are legal scope, exceptions, and currentness accurately bounded?
  3. Does the claim preserve the distinction between inference, exposure, belief, behavior, and effect?
CL-014EMPIRICAL

Reviewed empirical evidence

Automated moderation can make context, language, dialect, and identity-related errors; performance and disparate impact must be evaluated claim by claim.

Source scope
The sources support documented concern and remediation work; they do not prove identical bias across every model, language, or platform.
Do not infer
Do not generalize one conflict, language, or product finding to every moderation system.
Competing explanations
Where outcomes are discussed, ordinary ranking changes, user choice, market incentives, security requirements, model error, institutional process, and non-AI causes remain possible unless claim-specific evidence excludes them.
Affected-person context
Contextual errors involving reclaimed language, Arabic usage, and crisis speech are bounded examples, not universal error-rate estimates.
Rights and privacy implications
Potential implications include freedom of thought or expression, mental and data privacy, equality, autonomy, identity, notice, due process, and effective remedy; legal scope remains jurisdiction-specific.
Correction trigger
Update with audited error rates, independent studies, and platform-specific currentness evidence.

Claim-specific sources

WIP.54 evidence overlays

Review all twenty evidence stages
Artifact or event existence
PEER_REVIEWED_OR_AUTHORITATIVE_RECORD_LOCATED
Content status
CLAIM_REVIEWED_AT_CITATION_LEVEL
Coordination
NOT_APPLICABLE_OR_NOT_CLAIMED
Actor identity
RESEARCH_OR_REPORTING_BODY_IDENTIFIED
Sponsorship or direction
SOURCE_SCOPE_RECORDED; INDEPENDENCE_NOT_ASSUMED BEYOND SOURCE
Intent
RESEARCH_OR_GOVERNANCE PURPOSE RECORDED
Output
PUBLIC REPORT OR STUDY CONFIRMED
Distribution
PUBLICATION CONFIRMED
Availability
PUBLICLY AVAILABLE
Reach
NOT A PERSUASION REACH CLAIM
Exposure
NOT ASSESSED
Attention
NOT ASSESSED
Recall
NOT ASSESSED
Comprehension
NOT ASSESSED
Credibility
SOURCE AND METHOD BOUNDED
Belief or attitude
NOT ESTABLISHED BEYOND REPORTED STUDY
Intention
NOT ESTABLISHED
Behavior
NOT ESTABLISHED UNLESS CLAIM TEXT STATES OTHERWISE
Operational outcome
CONTEXT DEPENDENT
Strategic effect
NOT_ESTABLISHED
Questions for specialist review
  1. Is the claim phrased no more strongly than the cited sources support?
  2. Are legal scope, exceptions, and currentness accurately bounded?
  3. Does the claim preserve the distinction between inference, exposure, belief, behavior, and effect?
CL-015LEGAL

Primary legal text located

The EU Digital Services Act provides a procedural model requiring reasons and redress for certain visibility and monetization restrictions; it is not a universal global moderation code.

Source scope
Primary legal text establishes bounded duties within EU scope and defined service categories.
Do not infer
Do not imply every user, platform, or jurisdiction receives identical rights.
Competing explanations
Where outcomes are discussed, ordinary ranking changes, user choice, market incentives, security requirements, model error, institutional process, and non-AI causes remain possible unless claim-specific evidence excludes them.
Affected-person context
EU implementation aggregates document formal redress use, not individual accessibility or effectiveness for every affected person.
Rights and privacy implications
Potential implications include freedom of thought or expression, mental and data privacy, equality, autonomy, identity, notice, due process, and effective remedy; legal scope remains jurisdiction-specific.
Correction trigger
Update after implementation guidance, enforcement decisions, amendments, or court interpretation.

Claim-specific sources

WIP.54 evidence overlays

Review all twenty evidence stages
Artifact or event existence
CONFIRMED_BY_PRIMARY_LEGAL_TEXT
Content status
PRIMARY_TEXT_REVIEWED_AT_BOUNDED_CLAIM_LEVEL
Coordination
NOT_APPLICABLE
Actor identity
LEGISLATIVE_OR_TREATY_BODY_IDENTIFIED
Sponsorship or direction
PUBLIC_LEGISLATIVE_OR_INTERNATIONAL_PROCESS
Intent
BOUNDED_TO_STATED_LEGAL_PURPOSE
Output
ENACTED_OR_FORMALLY_PUBLISHED_TEXT
Distribution
OFFICIAL_PUBLICATION_CONFIRMED
Availability
PUBLICLY_AVAILABLE
Reach
NOT_ASSESSED
Exposure
NOT_ASSESSED
Attention
NOT_ASSESSED
Recall
NOT_ASSESSED
Comprehension
NOT_ASSESSED
Credibility
LEGAL_AUTHORITY_IS_JURISDICTION_AND_SCOPE_SPECIFIC
Belief or attitude
NOT_APPLICABLE
Intention
NOT_APPLICABLE
Behavior
IMPLEMENTATION_NOT_MEASURED
Operational outcome
ENFORCEMENT_OR_COMPLIANCE_OUTCOME_NOT_ASSESSED
Strategic effect
NOT_ESTABLISHED
Questions for specialist review
  1. Is the claim phrased no more strongly than the cited sources support?
  2. Are legal scope, exceptions, and currentness accurately bounded?
  3. Does the claim preserve the distinction between inference, exposure, belief, behavior, and effect?

E · AFFECTED-PERSON & COMMUNITY EVIDENCE

Whose experience is represented—and whose remains missing

These records are public, consent-aware, and bounded. Illustrative accounts are not converted into prevalence estimates or universal community views.

CLAE-006-CREATOR-RECLAIMED-LANGUAGECreator appeal involving reclaimed identity languageIllustrative individual case.
Source role
First-person platform appeal summarized by independent oversight body
Consent/privacy boundary
Use only details made public in the decision; do not infer identity, income, or location beyond the record.
Supports
Context-sensitive language error, discoverable appeal, correction, restoration, and possible creator visibility/livelihood consequence.
Does not establish
Platform-wide error rate, measured lost revenue, or that every use of a reclaimed slur is allowed.
Selection and nonresponse limits
Selected appeal; not a representative sample of moderation decisions.
Risk boundary
Do not amplify slurs gratuitously or enrich the creator's identity.
Correction/withdrawal
Correction path: /corrections; public decision remains the source authority.
CLAE-007-MULTILINGUAL-SPEECHArabic and multilingual communities affected by overbroad rulesSystemic policy analysis with explicit scope limits.
Source role
Policy advisory opinion with stakeholder and platform evidence
Consent/privacy boundary
Use public aggregated findings and avoid identifying individual speakers or conflict-affected users.
Supports
Language-context error, disproportionate burden, and need for context-aware review while maintaining violence-prevention rules.
Does not establish
That every removal is erroneous, every use is benign, or all platforms share the same policy.
Selection and nonresponse limits
Policy advisory process rather than representative community survey.
Risk boundary
Avoid doxxing, religious inference, or conflict-position attribution.
Correction/withdrawal
Correction path: /corrections; reopen for implementation updates and language-specific accuracy data.
CLAE-008-CRISIS-SPEECH-AUTOMATED-APPEALCrisis-context speech removed and appeal rejected automaticallyIllustrative crisis case.
Source role
Independent oversight decision on one affected user's appeal
Consent/privacy boundary
Use public decision facts only; do not infer user identity or reproduce graphic material.
Supports
Failure mode of automated appeal, public-interest context, and the need for authorized human review in crises.
Does not establish
The truth of every conflict claim, platform-wide frequency, or strategic effect of the removal.
Selection and nonresponse limits
Selected case under expedited oversight; not representative.
Risk boundary
Avoid graphic reproduction, conflict-party profiling, and exposure of the appellant.
Correction/withdrawal
Correction path: /corrections; reopen for platform implementation reports or superseding decisions.

F · SCIENTIFIC & LEGAL CURRENTNESS

Measurement validity and jurisdiction remain separate questions

A law may regulate a system without validating its scientific claims. A model may detect a signal without validly inferring an emotion, intention, personality, loyalty, or vulnerability.

CLSCI-010-DECISION-CONSEQUENCEFrom inference output to consequential actionCONSEQUENCE_AND_REMEDY_ARE_SEPARATE_FROM_MODEL_ACCURACY
Construct validity
A score must measure the decision construct rather than a convenient proxy.
Generalization
A model valid in one institution or period may not transfer.
Calibration/base rates
Decision thresholds must reflect error cost, legal duties, and uncertainty. Low-base-rate adverse events can produce many false flags.
Error and disparate-impact burden
Track denial, accusation, removal, discipline, and missed opportunity separately. Audit outcomes by protected and access-relevant groups where lawful and ethical.
Action, override, remedy
Record who acted, what rule applied, and whether the model was determinative or advisory. Human review must be independent, informed, and empowered. Specific reasons, evidence access, correction, restoration, compensation, and propagation to downstream systems.
CLLAW-007-EU-DSAEuropean Union · Statement of reasons, complaint, out-of-court dispute, recommender transparency, and systemic riskENACTED_AND_OPERATIONAL_WITH_AGGREGATE_IMPLEMENTATION_EVIDENCE
Enacted text
Regulation (EU) 2022/2065 establishes bounded procedural and systemic obligations for covered intermediary services.
Effective date
General application from 2024-02-17, with earlier designated-service obligations.
Implementation/guidance
Delegated acts, Commission enforcement, national Digital Services Coordinators, and transparency databases support implementation. Commission implementation materials report more than 1,800 first-half 2025 out-of-court disputes and 52% reversal among closed cases; this is not an all-decision error rate.
Enforcement/ruling
Commission investigations, fines, and court review must be recorded separately by case.
Scope limit
Not a universal global moderation code; a formal channel does not prove effective or accessible remedy for every user.

G · VISIBILITY ACTION & REMEDY

Identify the intervention, then test whether the remedy can repair it

Ranking differences are not automatically censorship; technically hosted content is not automatically discoverable. Effective remedy requires more than a nominal appeal form.

CLVIS-001-REMOVAL

Removal

Content is no longer available through the service under the relevant account or URL.

Notice/reason
A usable notice should state the action, rule, content, decision mode where required, and appeal path.
Evidence/appeal
Preserve the affected content and evidence sufficiently for challenge without creating new privacy harm.
Alternative explanations
Deletion by user, account change, link rot, jurisdiction restriction, or ordinary audience change.

CLVIS-002-ACCESS-RESTRICTION

Access restriction

Content remains stored but access requires login, relationship, warning acknowledgment, or other condition.

Notice/reason
State condition, scope, duration, and appeal route.
Evidence/appeal
Affected person should be able to inspect the basis unless lawful confidentiality applies.
Alternative explanations
Privacy settings, age/account status, network error, or user choice.

CLVIS-008-DEMONETIZATION

Demonetization or monetization change

Advertising, subscription, tipping, recommendation, or revenue eligibility changes while content may remain hosted.

Notice/reason
State the rule, affected revenue stream, duration, and appeal.
Evidence/appeal
Revenue and eligibility records should be available to the affected account.
Alternative explanations
Advertiser demand, market rates, copyright claims, product changes, or audience shift.

CLVIS-009-LABELING

Labeling or contextualization

A warning, fact-check, provenance, sensitivity, or context label is attached to content.

Notice/reason
Explain why the label appears and whether it affects distribution.
Evidence/appeal
Link evidence and correction process.
Alternative explanations
Publisher metadata, user settings, legal notices, or accessibility context.

CLVIS-012-ACCOUNT-PENALTY

Account-level or content-level penalty

A strike, reduced functionality, posting limit, suspension, or reputation penalty applies to content or account.

Notice/reason
State the triggering content, rule, penalty, duration, and appeal path.
Evidence/appeal
Preserve access to the challenged item and account history.
Alternative explanations
Security lock, compromised account, rate limit, payment, or user setting.

  1. CLREM-003-SPECIFIC-EXPLANATION

    Specific explanation

    Effective when: Explains the principal reasons, rule, evidence, uncertainty, and role of automation sufficiently to challenge the outcome.

    Weak or failed when: Model complexity, trade secrecy, or a boilerplate code substitutes for an actual reason.

    Evidence to retain: Reason specificity, consistency with record, automation role, and understandable alternatives.

  2. CLREM-006-HUMAN-AUTHORITY

    Authorized human review

    Effective when: A trained reviewer can inspect evidence, consider context/accommodations, overturn the result, and repair downstream effects.

    Weak or failed when: The reviewer repeats the model score, lacks authority, or the appeal is closed automatically.

    Evidence to retain: Reviewer role, authority, evidence inspected, decision, rationale, and conflict controls.

  3. CLREM-007-INDEPENDENT-APPEAL

    Independent and discoverable appeal

    Effective when: The channel is easy to find, accessible, free or proportionate, and reviewed independently from the initial decision path.

    Weak or failed when: The appeal repeats the same classifier, is unavailable in the person's language, or cannot change the outcome.

    Evidence to retain: Discovery path, completion rate, reviewer independence, reversal rate, and reasons—not reversal rate alone.

  4. CLREM-008-TIMELINESS

    Response time and interim protection

    Effective when: Urgency, livelihood, education, liberty, safety, and election/crisis context shape deadlines and interim relief.

    Weak or failed when: A successful appeal arrives after the event, job, exam, benefit, or audience opportunity has passed.

    Evidence to retain: Submission, acknowledgment, review, decision, restoration, and propagation timestamps.

  5. CLREM-009-RESTORATION-REPAIR

    Restoration, compensation, and downstream repair

    Effective when: The remedy restores access or opportunity, removes erroneous strikes/labels, corrects downstream records, and addresses measurable loss where authorized.

    Weak or failed when: Content returns but recommendation eligibility, reputation, pay, grade, or third-party records remain impaired.

    Evidence to retain: Restored state, downstream systems, monetary/equitable relief, and residual harm.

  6. CLREM-010-AUDIT-REPEAT-PREVENTION

    Audit logs and repeated-error prevention

    Effective when: Systems preserve accountable logs, investigate root causes, update rules/models/training, and test whether the error recurs across languages and groups.

    Weak or failed when: A single case is fixed without identifying systemic causes or affected peers.

    Evidence to retain: Version, trigger, reviewer path, root cause, corrective action, regression test, and aggregate outcome.

  7. CLREM-011-ACCESSIBILITY-LANGUAGE

    Accessibility, language, and advocate support

    Effective when: Notice and remedy work with assistive technology, narrow screens, plain language, relevant languages, and authorized representatives.

    Weak or failed when: The formal channel is unusable because of disability, literacy, language, identity verification, cost, or device barriers.

    Evidence to retain: Languages, formats, assistive-technology tests, representative support, and failure/abandonment data.

  8. CLREM-013-TRANSPARENCY

    Aggregate public transparency

    Effective when: Aggregate reports disclose action types, reasons, automation, appeals, reversals, timing, language/region, and limitations without exposing individuals.

    Weak or failed when: A single total hides mechanisms, groups, or whether users could obtain remedy.

    Evidence to retain: Denominators, definitions, coverage, missingness, subgroup privacy, and changes over time.

H · OUTCOME & DOWNSTREAM REPAIR

Documented reversals, restoration, relief, deletion, and implementation gaps

A required or announced remedy is not treated as proof that copied data, ranking signals, lost income, delayed access, reputation effects, or repeated errors were repaired.

CLOUT-003-DSA-OUT-OF-COURT-AGGREGATEDSA out-of-court dispute outcomes in the first half of 2025 AGGREGATE_PROCEDURAL_REMEDY_OUTCOME

The Commission reports more than 1,800 disputes reviewed in the first half of 2025 and platform decisions reversed in 52% of closed cases, restoring content or accounts.

Institution
Certified EU out-of-court dispute settlement bodies, platforms, national coordinators, and European Commission reporting
Notice and reason
DSA processes require reasons and redress routes, but accessibility and discoverability vary by service and person. Case-specific reasons are handled within disputes; the aggregate source does not publish every reason or evidence file.
Source/inferred-data access
The DSA supports statements of reasons and procedural access, not universal source-code disclosure.
Explanation
Aggregate figures demonstrate use and reversal, not the quality of every explanation.
Correction and deletion
Closed cases had a reported 52% reversal rate in the cited aggregate; individual categories and denominator details remain source-bounded. The aggregate does not establish deletion of every moderation profile, strike, or copied ranking signal.
Human authority and appeal independence
Certified bodies provide external process; authority and enforceability differ by mechanism and case. Out-of-court bodies are structurally external to platforms, but this record does not evaluate each body’s practical independence.
Repair
Restoration of content and accounts is reported; compensation, reach recovery, and reputational repair are not established.
Downstream propagation
No aggregate proof shows that all strikes, recommender signals, mirrors, search caches, or monetization records were corrected.
Accessibility, language, and support
Cross-language and disability access are not fully measured in the cited aggregate.
Unresolved harm
Lost time, reach, income, audience trust, and copied enforcement signals may outlast restoration.
Boundary
A successful appeal in one or many submitted cases cannot be generalized to all moderation decisions.
Reopening trigger
Reopen when certified-body datasets publish category, service, language, timeliness, accessibility, and downstream-repair details.
CLOUT-004-BREAST-CANCER-RESTORATIONBreast-cancer-awareness content: fifteen acknowledged enforcement errors CASE_BUNDLE_RESTORATION_AFTER_EXTERNAL_ESCALATION

Meta restored all fifteen breast-cancer-awareness posts after the Board brought the appeals to the company.

Institution
Meta and the Oversight Board
Notice and reason
Users reached the Board appeal process; the source does not establish that every affected creator could discover or access it. The bundle concerns mistaken enforcement against medical-awareness imagery under nudity-related rules.
Source/inferred-data access
Creators had their own posts and appeal records; model features and full automated decision traces were not published.
Explanation
Meta acknowledged errors after escalation; the record identifies content types and enforcement context.
Correction and deletion
All fifteen appealed posts were restored. No evidence establishes deletion of strikes, model features, or derived enforcement signals beyond the documented correction.
Human authority and appeal independence
External escalation changed the outcome; ordinary first-line reviewer authority is not demonstrated. The Board is structurally separate from Meta but depends on the platform’s case framework and implementation.
Repair
Post restoration is confirmed; reach, campaign timing, health-information access, and monetization repair are not measured.
Downstream propagation
No public proof shows correction of every recommender, strike, cache, or duplicate signal.
Accessibility, language, and support
Cases spanned multiple countries; comprehensive language/accessibility support is not established.
Unresolved harm
Time-sensitive awareness reach and audience trust may not be recoverable after restoration.
Boundary
Restoration confirms correction of these decisions, not the full causal chain of lost reach or health outcomes.
Reopening trigger
Reopen on Meta implementation evidence, repeat-error data, or creator-reported downstream repair outcomes.
CLOUT-005-SOMALILAND-JOURNALISM-RESTORATIONSomaliland journalism page, four posts, and strike restored MULTI_LAYER_ACCOUNT_CONTENT_AND_STRIKE_RESTORATION

Meta republished a Somali-language journalism page, restored four posts, reversed the account strike, and later reinstated additional Somaliland content it acknowledged was removed in error.

Institution
Meta and the Oversight Board
Notice and reason
Four post appeals received repeated human review; the page appeal was automatically closed without prioritized review before Board escalation. The page and posts were incorrectly treated as violating Hateful Conduct despite public-interest journalism context.
Source/inferred-data access
The creator could inspect their content and outcomes; internal classifier/reviewer evidence was only partially described publicly.
Explanation
The Board documents the page-level, post-level, strike, language, and review-path failures.
Correction and deletion
Page, posts, and strike were restored; ten additional Somaliland appeals were also reported as errors and reinstated. No public record confirms deletion of all prior risk labels or copied moderation signals.
Human authority and appeal independence
Six human reviews upheld errors; external Board escalation prompted reversal, showing that human review alone did not guarantee remedy. The Board supplied external review after internal and automated appeal paths failed.
Repair
Content, page, and strike restoration are concrete; lost audience contact, news timeliness, revenue, and journalist safety effects remain unmeasured.
Downstream propagation
No complete propagation record covers search visibility, follower feeds, recommendations, mirrors, or future review queues.
Accessibility, language, and support
The content and decision include Somali-language context; comprehensive appeal-language access remains a review question.
Unresolved harm
News timeliness, safety, source trust, audience reach, and future account risk may persist after reinstatement.
Boundary
The case establishes documented error and restoration, not full downstream or population-level effect.
Reopening trigger
Reopen on Meta implementation updates, repeated Somali-language error data, or creator/press-freedom outcome evidence.
CLOUT-006-KENYA-SLUR-CURRENTNESSKenyan political speech restored after slur-list currentness review POLICY_CLASSIFICATION_CORRECTION

The Board overturned removal of a Kenyan political comment and found the contested term should not have qualified as a slur at the time of posting.

Institution
Meta and the Oversight Board
Notice and reason
The user reached external appeal; ordinary users without escalation may face different notice and access conditions. Hateful Conduct / slur designation was applied too broadly to evolving political language.
Source/inferred-data access
The public decision describes the language-list classification and contextual use; internal list governance is not fully exposed.
Explanation
The decision explains temporal, political, and contextual reasons for reversal.
Correction and deletion
The removal decision was overturned and content restored. No proof establishes deletion of all policy-risk labels or downstream ranking effects.
Human authority and appeal independence
External review changed the result; reviewer authority and policy-list governance remain distinct. The Board provides independent judgment but is not a court or universal public regulator.
Repair
Content restoration and policy correction are documented; political reach and debate timing were not restored measurably.
Downstream propagation
No published evidence confirms every language list, classifier, strike, recommender, or reviewer tool was updated.
Accessibility, language, and support
Local-language expertise was central; broader linguistic access is not quantified.
Unresolved harm
Lost election-period attention, account trust, and self-censorship may persist.
Boundary
Contextual reversal does not mean the term is never harmful or never regulable.
Reopening trigger
Reopen on public evidence that Meta updated the designation process and tested downstream language effects.

WIP.55 FIELD REALISM

Reports linked to this rights question

These owner-supplied reports add outcome, validity, currentness, lived-experience, or repair evidence. Exact source identity is preserved, while independent citation and specialist review remain open.

  1. REAL-01-ALGORITHMIC-REMEDYAlgorithmic Remedy Outcomes and Downstream RepairRemedy and downstream repair
  2. REAL-02-MACHINE-UNLEARNINGMachine Unlearning, Derived-Data Correction, and the Right to ChangeData lineage, correction, and deletion
  3. REAL-04-COGNITIVE-LIBERTY-LAWComparative Cognitive Liberty Law, Regulation, and Enforcement AtlasJurisdiction-specific law and implementation
  4. REAL-05-INVISIBLE-EDITOR-OUTCOMESThe Invisible Editor Outcome CasebookVisibility actions and observed outcomes
  5. REAL-06-AFFECTED-COMMUNITYAffected Person and Community Evidence in AI GovernanceAffected-person and community evidence
  6. REAL-12-DEMOCRATIC-COGNITIVE-DEFENSEDemocratic Cognitive Defense Without Domestic ManipulationBehavior-based democratic defense
  7. REAL-13-PREDICTIVE-DEPLOYMENTSPredictive Population Management: Deployments, Feedback Loops, and RemediesPredictive deployment reality and decision consequence

Open the complete field-realism evidence bridge

I · SAFEGUARDS & RESEARCH GAPS

What rights-preserving practice would require

Safeguards

  • Publish rule-specific enforcement statistics and error rates.
  • Evaluate performance by language, dialect, disability, identity, and conflict context.
  • Use qualified human review for high-impact or ambiguous cases.
  • Restore reach, monetization, and account standing when an appeal succeeds.

Open questions

  • How can platforms measure demotion error when users are not notified?
  • What independent data access is necessary for public-interest auditing?
  • How should urgent crisis moderation preserve evidence and testimony?
CONTEMPORARY / ONGOING CLAIM — NOT SETTLED HISTORY

J · SOURCES & REVIEW STATUS

Exact reports and claim-specific external records

Owner reports are shown with exact filename, size, and SHA-256. External records are linked where a public source is available. Public presentation never exposes protected repository paths or internal memory links.

  1. CLSRC-OWNER-04

    The Invisible Editor: AI Censorship, Algorithmic Suppression, and the Right to Know

    Owner-supplied research report · 2026-07-27 · Primary project input; independently review consequential claims

    Exact source: The Invisible Editor AI Censorship, Algorithmic Suppression, and the Right to Know.md · 35,621 bytes · SHA-256 8de48a8e5a90d2792c789185d26e476308df7286b65979c04f38a091dbdce0ee

    Supports
    Supports the public information architecture, issue taxonomy, rights framing, proposed safeguards, and source-recovery agenda for this section.
    Does not establish
    Does not independently establish every embedded citation, current legal conclusion, causal effect, platform practice, or universal right.
    Review status
    EXACT_SOURCE_PRESERVED_AND_EDITORIALLY_REVIEWED · Owner source received and preserved on 2026-07-27.
  2. CLSRC-OWNER-06

    Algorithmic Suppression and AI-Driven Censorship

    Owner-supplied research report · 2026-07-27 · Primary project input; independently review consequential claims

    Exact source: AI Moderation and Suppression Research.md · 56,439 bytes · SHA-256 46900cf830ecd5a36b4f574c23d133b898e91256177b55ca1626f1b9107c9430

    Supports
    Supports the public information architecture, issue taxonomy, rights framing, proposed safeguards, and source-recovery agenda for this section.
    Does not establish
    Does not independently establish every embedded citation, current legal conclusion, causal effect, platform practice, or universal right.
    Review status
    EXACT_SOURCE_PRESERVED_AND_EDITORIALLY_REVIEWED · Owner source received and preserved on 2026-07-27.
  3. CLSRC-EXT-08-EU-DSA

    Regulation (EU) 2022/2065 — Digital Services Act

    European Union · 2022-10-19 · Primary legal text

    Supports
    Requires clear reasons and redress paths for certain platform decisions, including visibility and monetization restrictions.
    Does not establish
    Does not eliminate moderation error, mandate identical platform ranking, or operate as a global speech code.
    Review status
    PRIMARY_TEXT_LOCATED · Primary text and full-application status checked on 2026-07-27.
  4. CLSRC-EXT-13-META-HRDD

    Meta Final Update: Israel and Palestine Human Rights Due Diligence

    Meta · 2025-12-01 · First-party record describing independent due-diligence follow-up

    Supports
    Documents recommendations and implementation work involving granular policy rationales, appeals, language and dialect routing, and transparency.
    Does not establish
    Does not independently resolve all claims of bias, prove equal outcomes, or cover every Meta product and conflict context.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Final update reviewed on 2026-07-27.
  5. CLSRC-EXT-17-EU-AI-ACT-TIMELINE

    AI Act regulatory framework and implementation timeline

    European Commission · 2026-07-27 · Primary EU institutional currentness source

    Supports
    Supports current phased AI Act application dates and records that the targeted 2026 AI Omnibus amendments were adopted and entered into force on 2026-07-27.
    Does not establish
    Does not make all obligations immediately applicable, eliminate exceptions, prove provider compliance, or provide legal advice for a particular deployment.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official Commission page reviewed 2026-07-28; the Omnibus is enacted, not merely proposed.
  6. CLSRC-EXT-18-DSA-IMPACT-APPEALS

    How the Digital Services Act enhances content moderation transparency and appeals

    European Commission · 2026-03-27 · Primary EU institutional implementation source

    Supports
    Supports DSA reason and redress mechanisms and the Commission's aggregate that first-half 2025 out-of-court bodies reviewed more than 1,800 disputes and reversed 52% of closed cases.
    Does not establish
    Does not supply an all-decision denominator, platform-wide error rate, universal accessibility finding, or proof that every downstream strike, ranking, cache, income, or audience effect was repaired.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official Commission implementation page checked 2026-07-28; aggregate remedy outcomes remain case-selection dependent.
  7. CLSRC-EXT-21-NIST-FRVT-DEMOGRAPHICS

    Face Recognition Vendor Test Part 3: Demographic Effects (NISTIR 8280)

    National Institute of Standards and Technology · 2019-12-19 · Primary NIST technical evaluation

    Supports
    Supports measured demographic differentials in many face-recognition algorithms and the need to track false-positive and false-negative burdens by application and dataset.
    Does not establish
    Does not establish that every algorithm has identical error patterns, that identity matching reveals emotion or intent, or that laboratory results automatically predict every field deployment.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Stable NIST publication and current demographic-effects index checked 2026-07-28.
  8. CLSRC-EXT-24-ICO-SELDOM-HEARD-VOICES

    Seldom Heard Voices: ethnic minority groups and gig economy workers' experiences

    UK Information Commissioner's Office / Thinks Insight & Strategy · 2026-07-01 · Primary public research report with explicit qualitative sample limits

    Supports
    Supports lived-experience evidence from 43 participants, including 15 gig workers, about data sharing, discrimination concerns, language access, inaccurate data, work opportunities, and barriers to exercising information rights.
    Does not establish
    Does not provide a representative prevalence estimate for all ethnic-minority groups or gig workers, prove platform intent, or establish the outcome of a specific appeal.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official commissioned report published July 2026 and reviewed 2026-07-28.
  9. CLSRC-EXT-25-EEOC-ITUTORGROUP

    iTutorGroup to pay $365,000 to settle EEOC discriminatory hiring suit

    U.S. Equal Employment Opportunity Commission · 2023-09-11 · Primary agency litigation and settlement record

    Supports
    Supports a resolved federal case in which the EEOC alleged automated rejection thresholds based on age and sex, with monetary and non-monetary relief.
    Does not establish
    A settlement does not establish every alleged fact through trial, represent all automated hiring systems, or prove that every older applicant was affected in the same way.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official EEOC settlement record checked 2026-07-28.
  10. CLSRC-EXT-26-FTC-RITE-AID

    Rite Aid facial-recognition case and modified order

    U.S. Federal Trade Commission · 2024-03-05 · Primary agency enforcement record

    Supports
    Supports a documented FTC case alleging harmful false matches and inadequate safeguards, and an order imposing a five-year surveillance-use prohibition plus deletion, notice, assessment, and complaint-response duties.
    Does not establish
    Does not prove every allegation through a contested trial, establish the error rate of every face-recognition system, or extend the order beyond its parties and terms.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official FTC case page and modified order checked 2026-07-28.
  11. CLSRC-EXT-29-OVERSIGHT-DRAG-RECLAIMED

    Reclaimed Term in Drag Performance

    Oversight Board · 2025-04-23 · Independent platform oversight record; one illustrative appeal

    Supports
    Supports an affected creator's appeal, Meta's acknowledged context error, restoration, and the reported visibility and livelihood relevance of the removed post.
    Does not establish
    Does not provide a platform-wide error rate, measure lost income, or establish that every reclaimed-term removal is wrongful.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public decision reviewed 2026-07-28.
  12. CLSRC-EXT-30-OVERSIGHT-SHAHEED

    Referring to Designated Dangerous Individuals as 'Shaheed'

    Oversight Board · 2024-03-26 · Independent platform oversight analysis with stakeholder input

    Supports
    Supports evidence that a blanket rule could over-enforce multilingual and contextual speech and disproportionately burden Arabic speakers and other language communities while legitimate safety goals remain.
    Does not establish
    Does not bind all platforms, establish every removal's intent, or prove that every use of the term is benign.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public policy advisory opinion reviewed 2026-07-28.
  13. CLSRC-EXT-31-OVERSIGHT-ALSHIFA

    Al-Shifa Hospital

    Oversight Board · 2023-12-19 · Independent platform oversight record; one crisis-context case

    Supports
    Supports a documented case in which an initial removal and appeal rejection were automated without human review and the Board found public-interest speech had been removed incorrectly.
    Does not establish
    Does not establish a universal platform pattern, determine every factual claim in the underlying conflict, or prove strategic effect from the removal.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public decision reviewed 2026-07-28.
  14. CLSRC-EXT-34-CFPB-ADVERSE-ACTION

    Consumer Financial Protection Circular 2022-03: adverse action notification when creditors use complex algorithms

    Consumer Financial Protection Bureau · 2022-05-26 · Primary agency interpretation of existing adverse-action duties

    Supports
    Historically documents the CFPB's 2022 interpretation that covered creditors could not use model complexity as an excuse for failing to provide specific principal reasons under ECOA and Regulation B.
    Does not establish
    The circular was withdrawn on 2025-05-12, is not current CFPB guidance, does not govern every sector, and does not repeal or fully define the underlying statutory and regulatory duties.
    Review status
    ARCHIVED_WITHDRAWN_GUIDANCE_RETAINED_FOR_HISTORY_AND_UNDERLYING_LAW_CONTEXT · Official CFPB withdrawal index checked 2026-07-28; cite as withdrawn historical guidance only.
  15. CLSRC-EXT-35-NIST-POST-DEPLOYMENT

    Challenges to the Monitoring of Deployed AI Systems (NIST AI 800-4)

    National Institute of Standards and Technology · 2026-03-09 · Primary NIST post-deployment evaluation guidance

    Supports
    Supports the need to complement controlled pre-deployment evaluation with ongoing field monitoring for functionality, human factors, security, impacts, and changing context.
    Does not establish
    Does not certify any particular system, define settled best practice for every sector, or prove that monitoring alone prevents harm.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Official NIST publication checked 2026-07-28.
  16. CLSRC-EXT-38-OVERSIGHT-BREAST-CANCER-2025

    Breast Cancer Awareness Content

    Oversight Board · 2025-05-15 · Public independent oversight record with case-specific restoration outcome

    Supports
    Supports that Meta restored all fifteen breast-cancer-awareness posts after the Board brought the appeals to the company, and that the bundle documents acknowledged enforcement errors.
    Does not establish
    Does not establish a platform-wide error rate, complete downstream reach repair, compensation, or long-term prevention of repeat errors.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public decision checked 2026-07-28; restoration outcome is case-specific and not precedential.
  17. CLSRC-EXT-39-OVERSIGHT-SOMALILAND-2025

    Reporting on Somaliland Current Affairs

    Oversight Board · 2025-10-30 · Public independent oversight record with page, post, and strike restoration

    Supports
    Supports that Meta republished a Somali-language journalism page, restored four posts, reversed a strike, and separately reinstated additional appealed Somaliland content after acknowledged error.
    Does not establish
    Does not establish complete repair of audience, income, reputation, or chilling effects, or a platform-wide prevalence rate for Somali-language enforcement error.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public decision checked 2026-07-28; the case records multiple human-review failures and restoration after external escalation.
  18. CLSRC-EXT-40-OVERSIGHT-KENYA-SLUR-2025

    Comment on Kenyan Politics Using a Designated Slur

    Oversight Board · 2025-12-09 · Public independent oversight record with policy-list currentness finding

    Supports
    Supports that the Board overturned removal of Kenyan political speech and found the designated term should not have qualified as a slur when posted.
    Does not establish
    Does not establish that every use of the term is harmless, that every language list is inaccurate, or that restoration repaired all prior visibility and participation effects.
    Review status
    LOCATED_AND_REVIEWED_AT_CITATION_AND_SCOPE_LEVEL · Public decision checked 2026-07-28; the finding is contextual and time-sensitive.

This section is educational and non-operational. It is not legal advice, clinical guidance, human-rights certification, or authorization to conduct monitoring, influence operations, censorship, or psychological targeting.

Cognitive-liberty overviewAI PSYOPS taxonomyEvidence methodSubmit a correction

Page complete Moderation, Suppression, and Appeal Page label: CONTEMPORARY / ONGOING CLAIM — NOT SETTLED HISTORY