AI · GAMEDEV · STEAM FLAGS
Decision guide · July 2026

AI in game
development

A use-case guide with Steam disclosure flags.

For indie & AA teams shipping to PC / console on Unreal Engine 5 or Unity 6. Every way a studio can use AI, mapped to Steam’s disclosure rules — so you can decide what to embrace, what to disclose, and what to avoid.

Exempt Disclose · Pre-gen Disclose · Live-gen Gray area Prohibited
The dividing line

“Does AI-created content ship in the game or marketing — and get consumed by players?”

NO → EXEMPT

AI only helped you work faster; its output never reaches the player’s screen or speakers.

YES → DISCLOSE

AI-made content ships and is seen, heard, or read by players — pre-generated or live.

Steam’s rules

How the disclosure rules work

PRE-GENERATED

Content made with AI during development that ships and is consumed by players.

You affirm no illegal or infringing content and consistency with your marketing. Valve reviews AI output the same as everything else, and it’s declared in the Content Survey — shown publicly on your store page.

LIVE-GENERATED

Content made with AI while the game runs — LLM NPCs, runtime image / voice, dynamic quests.

= Pre-generated rules, plus:
+ describe the guardrails blocking illegal output
+ players get a Steam Overlay report button
+ highest review scrutiny
Adult-Only sexual content is prohibited

What changed in January 2026

01

Efficiency tools are explicitly exempt. Code assistants, generative fill on internal drafts, AI debugging — “not the focus of this section.”

02

“Consumed by players” is the line. Internal art, AI-assisted code and brainstorming don’t need flagging.

03

Marketing assets count. AI content in store pages, capsule art, screenshots or trailers is pre-generated disclosure.

04

Live-gen got new machinery. A Steam Overlay button lets players report illegal AI-generated content.

Disclosure is necessary — not sufficient

Valve has rejected games on training-data provenance grounds even with full disclosure. Image and story generation are the most legally fraught.

DLC & updates are covered. The survey only edits via Steam Support once approved.

Vagueness is the failure mode. Specific, honest disclosures are a far smaller problem than vague ones.

The community cost is real. The AI label can roughly halve a game’s review count and skew it negative.

Part two

Use cases, discipline by discipline

The full tables from the guide. Use the filter to isolate everything at a given flag level across all disciplines.

Filter
01

Studio administration & management

None of it ships to players — so nothing here needs Steam disclosure. Watch confidentiality and hiring law instead.

Meeting notes, transcription & summaries
EXEMPT
Confidentiality of unreleased details; check retention terms
Email drafting & internal comms
EXEMPT
Scheduling, calendar & resource planning
EXEMPT
Budgeting, forecasting & spreadsheet models
EXEMPT
Verify formulas; AI arithmetic errors compound
HR: job descriptions, question banks, onboarding
EXEMPT
Bias review; some regions regulate AI in hiring decisions
Contract & publisher-agreement first-pass review
EXEMPT
Not a lawyer substitute; don’t paste NDA’d terms into consumer tools
Pitch decks & publisher / funding materials
EXEMPT/GRAY
Steam doesn’t govern; AI “target renders” can poison publisher trust
Grant / fund applications
EXEMPT/GRAY
Many funds have their own AI-use declarations — check each
Production: sprint planning, backlogs, risk registers
EXEMPT
Internal documentation & wiki maintenance
EXEMPT
Telemetry & playtest data analysis
EXEMPT
Player-data privacy (GDPR / CCPA) if using cloud tools
Translation of internal docs
EXEMPT
02

Game design

Ideation is free. The line is whether generated design data ships — baked at build time or at runtime.

Brainstorming mechanics, themes & names
EXEMPT
Ideas rewritten by humans aren’t “shipped AI content”
Drafting GDDs, design & systems specs
EXEMPT
Internal documents only
Level design ideation (internal refs)
EXEMPT
Exempt only while it stays internal reference
Classic procgen — noise, WFC, rules-based
EXEMPT
Not “generative AI” under the policy; don’t over-disclose
Balancing via simulation / ML playtest agents
EXEMPT
Agents test the game; nothing they create ships
Economy tuning & spreadsheet-driven balance
EXEMPT
Player-facing strings written alongside are pre-gen
Accessibility design review
EXEMPT
AI-generated level / dungeon layouts shipped
DISCLOSE
If the generated layout data ships in the build
Quest / mission structure baked at build time
DISCLOSE
Ships with the game → pre-generated
Quest / mission generation at runtime
LIVE-GEN
Guardrails required
03

Programming & engineering

Tooling is explicitly exempt — even shipped code. Flags appear only when a model generates content at runtime.

Code assistants / pair programming
EXEMPT
You own quality, security & license contamination
AI-assisted gameplay / engine code that ships
EXEMPT
If AI-built end-to-end, consider a brief voluntary note
Debugging, crash triage, log analysis
EXEMPT
Build / CI automation, scripting, tooling
EXEMPT
Editor extensions & internal tools
EXEMPT
Performance optimization & netcode assistance
EXEMPT
Security-review anything auth / payment related
Shader generation (HLSL / shader graph)
GRAY
Code is exempt; if the visual result is AI art, lean disclose
Unity 6: AI assistant, generators, Sentis runtime
DEPENDS
Assistant exempt · shipped generator assets disclose · runtime gen live
UE5: ML Deformer, Learning Agents, NNE
DEPENDS
Approximation of your own work exempt · runtime generative model live
04

Art & animation

· highest scrutiny

The most disclosure-sensitive discipline. Disclosure ≠ safety — provenance is the risk. Prefer licensed / indemnified-data tools for shipped art.

Concept art / moodboards (internal only)
EXEMPT
Becomes disclosable if it appears in an artbook or store page
Generative fill / retouch on internal drafts
EXEMPT
Exempt if the draft never reaches players
Retopology, auto-UV, auto-LOD, auto-rig
EXEMPT
Efficiency tooling on your own assets
DLSS / FSR / XeSS upscaling & frame-gen
EXEMPT
Rendering technology, not generative content
AI upscaling of your own artwork
GRAY
Enhancement vs generation; cautious teams disclose one line
Motion matching / learned locomotion
GRAY
Blending your mocap exempt; runtime synthesis leans live
Generative fill / AI features on shipped assets
DISCLOSE
Any AI contribution to a final shipped image counts
Final 2D art: characters, portraits, UI, backgrounds
DISCLOSE
Highest provenance risk; prefer Firefly-class licensed data
Textures, materials & 3D model generation
DISCLOSE
Check tool license terms for commercial use
Motion synthesis / text-to-motion clips shipped
DISCLOSE
Physics-assist on artist keyframes is efficiency (exempt)
Style transfer applied to shipped assets
DISCLOSE
Avoid mimicking living artists’ styles
Runtime image generation (player-prompted)
LIVE-GEN
Guardrails + overlay; NSFW filtering essential
05

Audio, voice & music

Enhancing human work is free; generated audio that ships is disclosed. Any voice needs documented consent.

Cleanup, de-noise, dialogue isolation
EXEMPT
iZotope RX, Adobe Podcast
Temp / placeholder audio during development
EXEMPT
Only if fully replaced before ship — audit for leaks
Adaptive audio via Wwise / FMOD rules
EXEMPT
Rules-based layering, not ML generation
AI-assisted mixing / mastering of human music
EXEMPT
Enhancement of human work; conservative teams add a line
Music composition shipped in game
DISCLOSE
Check licenses; music provenance litigation is active
Sound-effects generation shipped in game
DISCLOSE
ElevenLabs SFX, custom models
TTS voice lines / voice cloning baked to files
DISCLOSE
Written consent to clone; SAG-AFTRA & likeness laws apply
Runtime TTS (dynamic narration / accessibility)
GRAY
Reading human text a feature; voicing AI-written text is live-gen
Runtime voice conversion / generative music (ML)
LIVE-GEN
Moderation + console online-safety cert
06

Narrative, writing & localization

Runtime LLM dialogue is the flagship live-gen case — and story generation is a legally fraught category.

Grammar / spell-check & copyediting
EXEMPT
Editing assistance is efficiency work
Human-translated loc with AI QA / consistency checks
EXEMPT
Brainstorming plots, characters, lore (rewritten)
EXEMPT
If AI text survives lightly edited, disclose
Subtitles auto-generated from human VO
GRAY
Transcription of human content; verify accuracy (lean exempt)
AI localization of in-game text (baked in)
GRAY
AI content consumed by players; lean disclose — “MT, human-reviewed”
Shipped dialogue, barks, item & lore text by AI
DISCLOSE
Story generation flagged as legally fraught
Runtime LLM NPC dialogue & dynamic quest text
LIVE-GEN
Describe guardrails specifically; budget for API costs
AI chatbot companion inside the Steam build
LIVE-GEN
External Discord bots are outside Steam’s scope
07

QA & testing

Nothing QA makes reaches players — effectively all exempt. Anonymize player data before using cloud tools.

Automated playtesting bots
EXEMPT
Unity ML-Agents, UE5 Learning Agents, modl.ai
Bug triage, duplicate detection & summarization
EXEMPT
Test-case generation & test code
EXEMPT
Crash log & telemetry anomaly analysis
EXEMPT
Playtest feedback synthesis & localization QA
EXEMPT
Anonymize survey / interview data first
ML-trained bots shipped as in-game opponents
GRAY
Behavior, not “content”; most treat as exempt. Player-visible output → reassess
08

Runtime & player-facing AI

· live-gen zone

Adopt only if ready for the guardrail, review, and moderation obligations. This is where AI is the feature.

LLM-driven NPC conversation
LIVE-GEN
Filters, prompt-injection resistance, topic limits, logging
Dynamic quest / story generation
LIVE-GEN
Validate output before display; constrained templates reduce risk
Runtime image / texture / voice generation
LIVE-GEN
NSFW / CSAM filtering is non-negotiable
Player-prompted AI-UGC shared with other players
HIGHEST RISK
One player’s generation reaches others — moderation + reporting
AI moderation of player chat
EXEMPT
Moderates, doesn’t create; auto-written visible messages are AI content
Live-generated Adult-Only sexual content
PROHIBITED
No exception, regardless of guardrails
09

Self-publishing, marketing & community

Steam marketing assets count. Off-Steam is Valve’s blind spot — but players compare, so stay consistent.

Capsule / header / library art made with AI
DISCLOSE
Explicitly covered; most scrutinized asset — highest backlash
Screenshots / trailer footage w/ AI in-game content
DISCLOSE
Follows from the content itself being disclosed
AI voiceover in trailers hosted on Steam
DISCLOSE
Consumed by players on your store page
Artbook / soundtrack DLC containing AI pieces
DISCLOSE
DLC is covered; internal concept art becomes disclosable if sold
AI-edited trailers (cutting real footage)
EXEMPT
Editing real gameplay is efficiency; AI-generated footage leans disclose
Store page description copy & localization
GRAY
Marketing text wasn’t the focus; ensure claims are accurate
Press, social, community mgmt & review responses
EXEMPT
Off-Steam; keep a human voice — templated replies read as dismissive
ASO, analytics, wishlist & pricing forecasting
EXEMPT
Not investment advice — sanity-check models
Shipping it

Writing a good disclosure

It appears verbatim on your store page. Name the tools, name the content categories, describe the human role — and for live-gen, describe guardrails concretely.

✓ GOOD · PRE-GEN

“Character portraits were created using Stable Diffusion and hand-edited by our artists for consistency. All other art, writing, and audio was created by our team.”

✓ GOOD · LIVE-GEN

“Optional NPC freeform chat uses an LLM. All outputs pass a moderation filter blocking hate speech, sexual and illegal content; NPCs are restricted to in-world topics; players can disable it.”

✗ BAD

“Some AI tools were used during development.”

Vagueness is the failure mode.

From audit to submission

01

Inventory the pipeline. List every AI touchpoint per discipline; assign a flag.

02

Classify by the test. Ships & consumed → disclose; efficiency → exempt; runtime → live-gen.

03

Decide what to embrace. Weigh provenance, community sentiment & review friction.

04

Keep provenance records. Tools, versions, licenses, key prompts, human-edit history.

05

Fill the survey precisely. It locks after approval — edits need Steam Support.

06

Re-audit every update. New AI assets post-launch, and every other storefront’s rules.

The decision

The whole guide, one decision tree

Did AI create or contribute to content?
→ No →No disclosure
Does it ship & get consumed by players?
→ No →EXEMPT
Is it generated while the game runs?
→ No →PRE-GENERATED
Yes — runtime generation
LIVE-GEN · guardrails + overlay + review
… and it’s Adult-Only sexual content?
PROHIBITED
Recommendation

Where most teams land

FULL EMBRACE

All exempt efficiency uses — the productivity is free and invisible to players.

SELECTIVE

Disclose-required uses with licensed / indemnified-data tools, disclosed specifically.

ONLY IF CORE

Live-gen only when AI is the feature — budget for guardrails & review.

Decide per use case. Disclose specifically. Keep the receipts.