SoftTech AI Registry
Meta · Facebook and Instagram

High-Confidence AI Content Violation Removal

Meta uses AI models to predict whether an individual piece of user-generated content on Facebook or Instagram violates a specific policy; a separate enforcement system applies Meta's policies and confidence requirements to that prediction and can delete the content automatically without individual pre-removal human approval. Meta focuses these proactive systems on supported illegal and high-severity enforcement areas and has tuned them to require a much higher degree of confidence before content is taken down. Other content is sent for human review, some removals require multiple reviewers, and enforcement based solely on matching previously identified violating content is excluded.

Recorded characteristics

Function
Controlled object: one individual piece of user-generated content on Facebook or Instagram (not an account, Page, Group, profile, network or user generally). Causal chain, stage by stage: (1) user-generated content is evaluated by Meta's AI/ML enforcement architecture; (2) an AI model predicts how likely the individual content item is to violate a particular policy; (3) Meta's separate enforcement technology applies Meta-defined policy and confidence requirements to that prediction; (4) where the qualifying automatic-enforcement conditions are satisfied, the content can be deleted automatically; (5) no individual human approval is required before the qualifying automatic removal. Architecture: content → AI model prediction → separate enforcement system applying Meta policy/confidence requirements → automatic deletion. Meta: "an AI model predicts whether a piece of content is hate speech or violent and graphic content. A separate system—our enforcement technology—determines whether to take an action, such as deleting, demoting or sending the content to a human review team for further review." Confidence boundary — Meta determines: platform policies; scope of proactive enforcement; required confidence level for automatic enforcement. The AI/model determines: the content-specific assessment; how likely the individual item is to violate a particular policy. The AI does not write Meta policy, does not choose its own enforcement threshold, and does not independently own the complete removal decision; the qualifying authority is the model-derived content judgment being operationalised by Meta's standing enforcement architecture. Policy scope: supported illegal and high-severity enforcement areas; Meta's examples are "terrorism, child sexual exploitation, drugs, fraud and scams" (examples only; not every category is claimed to use an identical classifier or enforcement architecture; not all Community Standards violations).
Data access
User-generated content on Facebook and Instagram; Meta states AI operates on URLs, text, images, audio and videos. Exact production features are not publicly established.
Actions
Can take actions
External actions
No
Human confirmation
Not required
Permission basis
Not established
Administrative control
Meta sets platform policies, the scope of proactive enforcement and the required confidence: "we're going to tune our systems to require a much higher degree of confidence before a piece of content is taken down." Human review boundary — Automatic qualifying path: AI/model-derived assessment reaches Meta's automatic-enforcement requirements → content can be removed without individual pre-removal human approval (human confirmation not_required applies only to this path). Other enforcement paths: content can be sent to human review; some removals require multiple reviewers ("in more cases, we are also now requiring multiple reviewers to reach a determination in order to take something down"). Appeals: occur after the original action and are not pre-action confirmation. Meta has not removed humans from content enforcement. Outside action recorded "no": the action occurs inside the product boundary (Meta enforcement system → Meta-hosted Facebook/Instagram content); the consequence for a third-party user does not by itself make it an outside action. Permission basis not_established: Meta acts through its own platform authority, which is not mapped to any user, delegated, agent, administrator or service-account mechanism. Default state not_established: the Registry's default-state methodology is deployment/customer oriented and does not cleanly describe this platform-governance architecture.
Default state
Not established
Availability
Operates on Facebook and Instagram; Meta publishes separate EU DSA transparency reports for each describing the same AI enforcement architecture. Threads excluded. Geographic scope of the January 2025 enforcement changes is not publicly established.
Licensing
Not applicable: platform enforcement operated by Meta on its own services; no customer licensing.
External model or provider
Meta's own AI models and machine-learning classifiers. No third-party model provider is established.
Limitations and uncertainty
Reversibility: Meta says people are often given an opportunity to appeal and ask Meta to review an enforcement action again ("People are often given the chance to appeal our enforcement decisions and ask us to take another look, but the process can be frustratingly slow and doesn't always get to the right outcome"). Not claimed: universal appeal availability, universal human appeal review, universal restoration, automatic restoration after appeal, or automatic reversal of strikes or other consequences. Matching exclusion: automatic enforcement based solely upon matching against previously identified violating content is excluded (exact hashes, perceptual hashes, previously adjudicated copies, near-identical content matching, known violating URLs, known-content databases); matching evidence is not used to establish this capability. Downstream penalties excluded: strikes, warnings, account restrictions, posting restrictions, feature restrictions, suspensions, disabled accounts, Page penalties, Group penalties — this capability stops at removal of the individual content item. Exclusions: (1) known-content matching as the sole decision mechanism; (2) exact/near-identical matching; (3) demotion/reduced distribution; (4) labels; (5) warnings; (6) recommendation/ranking changes; (7) human-review-only enforcement; (8) lower-confidence review routing; (9) strikes; (10) account restrictions; (11) suspensions; (12) account disablement; (13) Page penalties; (14) Group penalties; (15) Threads; (16) newer advanced-AI/LLM systems where current production authority is not established; (17) policy writing; (18) confidence-threshold selection by the model. Not publicly established: (1) exact production classifier architecture; (2) exact models used for each policy category; (3) exact confidence thresholds; (4) whether thresholds vary by country; (5) whether thresholds vary by language; (6) precise model weights; (7) complete training corpus; (8) exact current policy-category coverage of automatic AI removal; (9) exact classifier/model version associated with an individual removal; (10) feature attribution for an individual removal; (11) numerical confidence associated with an individual removal; (12) exact relationship between classifier, rules and ensemble components; (13) universal percentage of removals attributable to automatic AI classification; (14) universal appeal availability; (15) whether every automatic-removal appeal receives human review; (16) appeal SLA; (17) complete internal decision log; (18) retention period for enforcement records; (19) universal restoration behaviour; (20) relationship between restoration and downstream strikes/penalties; (21) maximum simultaneous content-action blast radius; (22) precise Threads parity; (23) exact production role of Meta's newer advanced-AI/LLM enforcement systems; (24) whether every high-severity policy category permits classifier-driven automatic removal; (25) where the January 2025 enforcement changes apply geographically; (26) share of automatic removals originating from AI classification versus known-content matching; (27) exact technical relationship between the AI model's prediction and Meta's separate enforcement system; (28) which removals require multiple human reviewers; (29) what substantive content, if any, changed in the 6 February 2026 update to the newsroom article. Monitoring: technically monitorable, but automated monitoring deliberately not enabled because of potentially applicable automated-data-collection restrictions. The newsroom article passed the two-fetch test (HTTP 200 both times, no redirect, 16,788 cleaned characters each, identical fingerprint 8b7b1a7eb50c00029fcc6d3c0a33d0ad810f419b0293bb8d38ee7b7c33c894a9); Meta's Automated Data Collection Terms state "You will not engage in Automated Data Collection without first obtaining Meta's express written permission". Transparency Center pages render client-side and return almost no content to the Registry fetch.

Evidence