# DPIA: In-Game Commentary

## Scope

AI-generated match commentary, live spectator commentary, highlight narration,
post-match analysis, consent-gated personal voice-clone commentary, and
post-season podcast-style recaps that reference gameplay events and public match
context.

## Data Categories

Account identifiers, friend account identifiers for explicit voice-clone
consent, match identifiers, broadcast identifiers, gameplay events, selected
replay segments, generated commentary, consent-artifact IDs, revocation state,
bias-review state, localization review state, and opt-out state.

## Risks

Commentary can infer sensitive traits, amplify harassment, expose player
behavior beyond the expected match context, introduce biased treatment based on
gender, region, and playstyle, or misuse a friend's voice likeness without clear
ongoing consent.

## Mitigations

Commentary generation uses opt-out state, avoids sensitive trait inference,
limits replay context to match events, and routes abuse reports into the same
moderation and appeal workflow as other player-facing content. Live broadcast
commentary also requires `BiasReview.AICommentary.Live` review for player
gender, region, and playstyle before generated lines can be broadcast. Personal
voice-clone commentary requires the player opt-in, the friend's explicit voice
likeness consent, watermarking, provenance embedding, minor-account blocking,
and a revocation-ledger check before any cloned-voice narration can be queued.
Post-season podcast recaps use grounded season statistics, replay highlights,
and public balance-ledger data, then receive human review before publication.
Localization nuance review is performed by native-speaker reviewer queues with
protected-trait inference disabled and glossary locks enforced.

## Retention

Generated commentary follows replay retention and is removed from eligible
archives during DSAR export refreshes or deletion purges.

## Human Review

Reported commentary can be reviewed, removed, and included in transparency
reporting when moderation action is taken.