# DPIA: AI Auto-Moderation

## Scope

Automated nudity, extreme-violence, and speech-safety screening used before
human moderation review for UGC, text chat, and voice transcripts.

## Data Categories

Account identifiers, content identifiers, moderation findings, chat text,
transcripts, preview labels, and evidence clip references.

## Risks

False positives can restrict lawful content. False negatives can expose players
to unsafe content. Cross-region processing can violate residency commitments.

## Mitigations

Human review is required for non-trivial findings, the Lilith persona policy is
used for consistent speech decisions, and each moderation action logs a DSA
statement of reasons. EU, PRC, and US account data stays in-region.

## Retention

Evidence follows the moderation retention schedule and is purged when a valid
deletion case reaches its 45-day purge point unless legal hold applies.

## Human Review

Human moderators review queued findings, appeals, and any blocked decision
before a long-duration sanction is finalized.