Playbook — Disclosing AI Use in Entertainment Works¶
Built from 44 cases (film, music, anime/manga: 30; AI characters: 14): what separated disclosures that were accepted from those that drew backlash. Legal labeling duties differ by country (→ comparison by country). This guide covers disclosure that goes beyond the legal minimum to protect the trust of fans and creators. Japanese edition: ../../foundations/ai-disclosure-guide.md
In one line¶
What decided the backlash was less whether AI was used than when, who, what, and how it was said. Disclosure that comes early, voluntarily, specifically, and together with the human work tends to be accepted. Disclosure that comes late, after discovery, vaguely, and downplaying the human work draws the strongest backlash.
Four dividing lines from the cases¶
| Dividing line | Accepted | Backlash |
|---|---|---|
| When (timing) | Voluntarily, at announcement or before release (A04 Twins Hinahima, A06 Cyberpunk: Peach John, F02 The Eternaut) | Only after being discovered post-release (F07 Late Night with the Devil, C14 Level-5, A09 Crunchyroll) |
| What (specificity) | Which process, how much, and why (F02 "blockbuster-level VFX within budget", A04 "assisted in 95% of cuts") | Just "we used AI", or "labor shortage" as the only reason (A01 The Dog & The Boy) |
| Human work (credits) | Names and roles of the people involved (A04, F10 Here) | People credited as a small "+Human" (A01), or AI marketed as a "replacement for people" (F05 Tilly Norwood, M10 TaTa) |
| Consent (voice, likeness, works) | Consent from the person or rights holder (M06 Randy Travis, M02 Grimes, F04 Disney × OpenAI) | Unauthorized voices or art styles (M01 AI Drake, A07 voice actors' voices, A08 "Ghibli style") |
Timing and context matter too. During labor negotiations or strikes (F01 Secret Invasion), or when the industry is especially sensitive to AI, the same use draws more backlash.
The disclosure process (5 steps)¶
Step 1. Inventory where AI was used¶
Record, process by process, what used AI and what did not. Include subcontractors (in A09 a vendor used AI in breach of contract).
| Category | Examples | Disclosure priority |
|---|---|---|
| Content the audience or players see or hear | Art, backgrounds, voices, music, lines, NPC dialogue generated at runtime | High (also most likely to be legally required) |
| Correcting or restoring human performance or work | Stem separation, de-aging, pronunciation correction | Medium (high if it touches what awards evaluate — F08 The Brutalist) |
| Tools for production efficiency | Coding assistants, scheduling, internal draft translation | Low (Steam also does not require it to be declared) |
Step 2. Check the legal minimum¶
- EU (AI Act Art. 50, from August 2026): notice that users are interacting with AI, machine-readable marking of outputs, disclosure of deepfakes. Artistic works may disclose in a way that does not hamper enjoyment
- China (Labeling Measures, from September 2025): both visible labels and metadata
- South Korea (AI Basic Act, from January 2026): advance notice of generative AI services and labeling of outputs
- US state laws: AI notices for AI companions (New York: all users, every 3 hours)
- Store and platform rules: Steam's generative AI disclosure (including runtime generation), AI labels on Spotify and Deezer → Details: comparison by country
Step 3. Decide what to say (the four-point set)¶
- Where: the process or scene ("part of the backgrounds", "one building-collapse shot")
- Why: the creative reason ("an expression impossible on this budget", "hair movement"). Never only "labor shortage" or "cost cutting"
- What people did: human judgment and finishing by the director, art staff, voice actors. Credit them by name
- Consent and rights: consent of people whose voice, likeness, or works were used; where training data came from
Step 4. Decide when and where to say it¶
- At announcement, before release, voluntarily, through official channels (press release, official site, store page)
- Inside the work: end credits, work information. In games, the character profile screen or the first conversation (AI characters → ai-characters P30)
- Machine-readable marks: metadata and watermarks on outputs (for the EU, China, Korea)
- In sensitive periods (e.g., during labor negotiations), be especially careful about how you announce
Step 5. Prepare how to respond when questioned¶
- Keep the Step 1 records so you can explain the facts immediately
- Answer with "what, why, what people did". A bare denial, or only a grand plan, leaves distrust (A03 Toei)
- Be ready to withdraw output whose quality fell short (A10 Prime Video's AI dubbing) — but make sure withdrawal does not remove the alternative altogether
Example wording¶
△ Likely to draw backlash
This work makes use of the latest AI technology. (Backgrounds: AI + Human)
○ Likely to be accepted
The background art is based on concept art by art director ___. In some scenes, an image generation model was used to add detail, and art staff ___ and ___ finished every cut by hand. This was our choice to achieve the density of the ___ scenes within a limited production schedule. The model was trained only on art materials whose rights we hold.
For an AI character (in-game)
's (character name) conversations are generated on the spot by generative AI. The character's setting and voice are supervised by scenario writer . Conversation content is stored for ___ days and can be reviewed and deleted from the "Memories" screen.
Common mistakes¶
- "It's minor, so no need to say anything" → even three still images triggered boycott calls (F07)
- Leading with "more, faster, cheaper with AI" → read as lower quality (A02 Orange)
- Writing people as assistants to the AI → the credit format itself causes backlash (A01)
- Saying "we don't use AI" without checking vendors → no contract terms or checks in place (A09)
- Marketing AI as a "replacement for people" → criticized on both employment and representation (F05, M10)
Checklist¶
- Recorded AI use per process, including subcontractors (Step 1)
- Checked legal labeling duties for each country and store you ship to (Step 2)
- Wrote all four points: where, why, what people did, consent and rights (Step 3)
- Plan to disclose voluntarily, before release, through official channels (Step 4)
- Credited the people involved by name
- Confirmed consent of people and rights holders whose voice, likeness, or works were used
- Prepared an explanation for when questioned, and criteria for withdrawal (Step 5)
Open questions¶
- How much does the wording of a disclosure actually change reception? (💡 testable with surveys or A/B tests)
- Does labeling a work as "made without AI" (human-made labels) function as value for the work?