@@ -54,6 +54,7 @@ To add a starter stack (one pick per need for a kind of game), follow [`stacks/R
 - `subcategories`: lower-case kebab-case (`base-meshes`, `field-recordings`). Reuse a value already in the catalog before inventing one, and do not add a variant of an existing value that differs only in plural or spelling. Where both forms were in use, the more common one was kept (`characters`, `environment`, `interior`, `tileset`, `vectors`, `pixel`, `impulse-responses`, `base-meshes`). Retired spellings are listed in [`site/value-aliases.json`](site/value-aliases.json), which the validator enforces (V19); add a line there when you merge two values.
 - `formats`: what you actually get, in one of four kinds. **File formats**, as commonly written: usually upper case (`PNG`, `FBX`, `JSON`, `VOX`), tool-specific ones as the tool writes them (`gdshader`, `tmx`, `ktx2`). **Engine or language targets**, as the product writes them (`glTF`, `Godot`, `Unity`, `Python`, `React`). **Delivery types** for software with no file format of its own, lower-case kebab-case (`godot-addon`, `blender-extension`, `npm`, `cli`, `library`, `desktop-app`, `mobile-app`, `middleware`, `model`). And `various` for an aggregator whose files come in too many formats to list. Platforms (`windows`, `macos`, `ios`) and descriptions (`heightfield`, `examples`) are `tags`, not formats. The allowed values are the closed list in [`site/format-vocabulary.json`](site/format-vocabulary.json), which the validator enforces (V20): reuse a listed spelling, and add a genuinely new format there in the same change that first uses it. Its `groups` drive the site's Format filter, so a new format that belongs to a group (another glTF or font container) goes into that group too. A format outside every group is still searchable and shown on the card, but the Format filter cannot select it.
 - `tags`: lower-case kebab-case words describing the content (`low-poly`, `modular`, `platformer`). Do not restate another field: the licence (`cc0`, `mit`, `ofl`, `open-source`, `royalty-free`), the credit requirement (`no-attribution`) or the entry's own `publisher` (`kenney`) are rejected (V19 via `site/value-aliases.json`, V21), because search and the filters read those fields directly and a copied tag drifts. A `varies` entry lists its licence mix in the body, not in tags.
+- `ai-generated` tag: required when the source says any of its content is AI-generated (V37; the body sentence "some sounds are AI-generated" without the tag fails). Add the entry to the list in [`docs/ai-assets.md`](docs/ai-assets.md) in the same change; the validator checks that list too.
 - Frontmatter is one `key: value` per line; lists are `[a, b]` or indented `- item` lines. A key given twice is an error. Entry files may not contain emoji.
 - `grid_dimensions` / `camera_perspective` / `hardware_tags`: optional metadata for 2D and UI entries. `3d` and `characters` entries leave `camera_perspective` out: it describes a 2D camera, and those entries have none.
 - `maintenance`: optional, `archived` | `inactive` (V22). Set `archived` when the source repository is archived, and `inactive` when it has had no push for over three years, with a dated `- Maintenance:` line in Notes saying what you read. Leave it out otherwise. The weekly link check reports a GitHub source whose state no longer matches.
 
@@ -8,7 +8,7 @@ license: custom
 commercial: true
 attribution_required: false
 formats: [WAV, MP3]
-tags: [loops, one-shots, api]
+tags: [loops, one-shots, api, ai-generated]
 added: 2026-10-26
 verified: 2026-10-26
 status: needs-review
 
@@ -1,6 +1,6 @@
 # AI-generated assets — catalog policy
 
-Last reviewed 2026-10-23. Court rulings and platform rules on AI content change quickly: **re-verify** dates and rulings before relying on them.
+Last reviewed 2026-11-03. Court rulings and platform rules on AI content change quickly: **re-verify** dates and rulings before relying on them.
 
 ## Practical catalog rules
 
@@ -12,14 +12,45 @@ Last reviewed 2026-10-23. Court rulings and platform rules on AI content change
 
 ## Platform notes (verify live)
 
-- **Steam**: research notes player-facing generative assets may require store disclosure; internal coding assistants may not. Check current Valve guidelines.
+- **Steam**: the Content Survey you fill in before review has a section on generative AI.
+  The live [Content Survey page](https://partner.steamgames.com/doc/gettingstarted/contentsurvey)
+  (read 2026-11-03) says it "is concerned with the use of AI in creating content that ships
+  with your game, and is consumed by players. This includes content such as artwork, sound,
+  narrative, localization, etc.", and that AI tools used only for development efficiency
+  are "not the focus of this section". It splits AI content into "Pre-Generated: Any kind of
+  content that ships with your game and is consumed by players that is created with the
+  help of AI tools during development" and "Live-Generated", content made while the game
+  runs, for which "you'll need to tell us what kind of guardrails you're putting on your
+  AI to ensure it's not generating illegal content."
 - **US copyright**: research cites human-authorship requirements for registration; pure machine output may be unprotected — that does **not** clear third-party training-data or likeness claims.
 - **Voice/likeness**: generating a real performer’s voice without rights is high risk regardless of model license.
 
+## What an `ai-generated` entry obliges you to disclose
+
+An entry tagged `ai-generated` holds content that its source says was made with AI tools,
+in whole or in part. If you ship any of it, it is content "created with the help of AI
+tools" that players consume, so on Steam it belongs in the Pre-Generated part of the
+Content Survey, whoever ran the generator. Describe which assets and where they came
+from. The same holds for a library that mixes human and AI files, such as Soundza, if a
+file you use is labelled AI on its page. Other stores have their own rules; read them
+before you submit.
+
 ## Mixed catalogs (human CC0 vs generator)
 
 [3DTexel](../catalog/shaders-vfx/3dtexel-decals.md) ships a **human-authored CC0 decal library** on the same site as paid premium files and AI generators. The catalog entry is the human CC0 set only. Footer (2026-08-29): free assets CC0; premium and AI assets use a separate 3DTexel License. Do not treat generator output as that CC0 grant.
 
+## Entries tagged ai-generated
+
+These entries carry the `ai-generated` tag; the validator keeps this list complete (V37).
+The [live search](https://tmhsdigital.github.io/Free-Game-Dev-Assets/?q=ai-generated) shows
+the same set.
+
+- [BitNuker Fantasy Environments Sampler](../catalog/2d/bitnuker-fantasy-environments-sampler.md)
+- [BitNuker Workshop Sampler](../catalog/2d/bitnuker-workshop-sampler.md)
+- [BudgetPixel Sound Effects](../catalog/audio/budgetpixel-sfx.md)
+- [SFXMint](../catalog/audio/sfxmint.md)
+- [Soundza](../catalog/audio/soundza.md)
+
 ## Related entries
 
 - [`catalog/shaders-vfx/3dtexel-decals.md`](../catalog/shaders-vfx/3dtexel-decals.md)
 
@@ -360,6 +360,36 @@ export function checkPreviewVocabulary(vocab, allow = PREVIEW_LICENSES, credits
   return errors;
 }
 
+/* ----------------------------------------------------------------- V37 */
+/**
+ * AI-generated content stays labelled. An entry whose body says some of its
+ * content "is" or "are" AI-generated carries the `ai-generated` tag, and
+ * docs/ai-assets.md links every tagged entry, since Steam's Content Survey
+ * asks a developer to disclose AI content they ship. "No generative AI" and
+ * "not AI-generated" do not count.
+ */
+const AI_CLAIM_RE = /\b(?:is|are)\s+(?:partly\s+|partially\s+|mostly\s+)?AI[- ]generated\b/i;
+
+export function checkAiTag(rel, meta, body) {
+  const tags = Array.isArray(meta.tags) ? meta.tags.map(String) : [];
+  if (tags.includes("ai-generated")) return [];
+  for (const sentence of String(body).split(/(?<=[.!?])\s+|\n/)) {
+    if (AI_CLAIM_RE.test(sentence) && !/\b(?:not|no|never)\b/i.test(sentence)) {
+      return [`${rel} says content is AI-generated ("${sentence.trim().slice(0, 80)}") but has no ai-generated tag`];
+    }
+  }
+  return [];
+}
+
+/** docs/ai-assets.md links every entry tagged ai-generated (repo paths relative to docs/). */
+export function checkAiGuide(guideText, taggedRels) {
+  const errors = [];
+  for (const rel of taggedRels) {
+    if (!String(guideText).includes(`(../${rel})`)) errors.push(`docs/ai-assets.md does not link ${rel}, which is tagged ai-generated`);
+  }
+  return errors;
+}
+
 /* ----------------------------------------------------------------- V36 */
 /**
  * `limits`: a short line for a restriction the licence fields cannot carry,
 
@@ -38,6 +38,8 @@ import {
   checkTitleMatchesName,
   checkValueAliases,
   checkValueSpellings,
+  checkAiGuide,
+  checkAiTag,
   checkCreditPlaceholders,
   checkLimits,
   checkKnownKeys,
@@ -901,6 +903,15 @@ rejects("V34 rejects an allowlisted licence missing from the vocabulary", checkP
 rejects("V34 rejects a required licence the caption would not credit", checkPreviewVocabulary({ licenses: { "OGA-BY-3.0": { attribution: "required" } } }, ["OGA-BY-3.0"], () => false), "would lack the creator credit");
 rejects("V34 rejects a credit on a licence that does not ask for one", checkPreviewVocabulary({ licenses: { MIT: { attribution: "notice" } } }, ["MIT"], () => true), "carry a credit the licence does not ask for");
 
+/* V37: AI-generated content stays tagged and listed ------------------------ */
+accepts("V37 accepts a tagged entry", checkAiTag("ok.md", { tags: ["ai-generated"] }, "Some sounds are AI-generated."));
+accepts("V37 ignores a denial", checkAiTag("ok.md", { tags: [] }, "itch metadata: No generative AI was used. The audio is not AI-generated."));
+accepts("V37 ignores a mention that is not a claim", checkAiTag("ok.md", { tags: [] }, "Built for AI-generated pixel art that is off-grid."));
+rejects("V37 rejects an untagged claim", checkAiTag("bad.md", { tags: ["loops"] }, "- Some library sounds are AI-generated and labelled."), "has no ai-generated tag");
+rejects("V37 rejects a partly-AI claim", checkAiTag("bad.md", {}, "The pack is partly AI-generated."), "has no ai-generated tag");
+accepts("V37 accepts a guide that links every tagged entry", checkAiGuide("- [S](../catalog/audio/s.md)", ["catalog/audio/s.md"]));
+rejects("V37 rejects a guide missing a tagged entry", checkAiGuide("nothing", ["catalog/audio/s.md"]), "does not link catalog/audio/s.md");
+
 /* V36: limits is one short plain line ------------------------------------- */
 accepts("V36 accepts a short plain line", checkLimits("ok.md", { limits: "Tree packs are Unreal-only; no standalone redistribution" }));
 accepts("V36 ignores a missing field", checkLimits("ok.md", {}));
 
@@ -10,6 +10,8 @@ import {
   checkEntryBasics,
   checkPreviewVocabulary,
   checkFeatured,
+  checkAiGuide,
+  checkAiTag,
   checkCreditPlaceholders,
   checkLimits,
   checkKnownKeys,
@@ -313,6 +315,7 @@ function main() {
     errors.push(...checkKnownKeys(rel, meta));
     errors.push(...checkCreditPlaceholders(rel, meta));
     errors.push(...checkLimits(rel, meta));
+    errors.push(...checkAiTag(rel, meta, body));
     // V24 types commercial and attribution_required (a quoted "true" is text).
     errors.push(...checkFieldTypes(rel, meta));
     errors.push(...checkEntryUrl(rel, meta));
@@ -371,6 +374,13 @@ function main() {
     errors.push(...checkComparisonLicences(cat, readText(readme), entriesByRel));
   }
 
+  errors.push(
+    ...checkAiGuide(
+      readText(path.join(ROOT, "docs", "ai-assets.md")),
+      records.filter(({ meta }) => Array.isArray(meta.tags) && meta.tags.map(String).includes("ai-generated")).map(({ rel }) => rel)
+    )
+  );
+
   const measured = {};
   for (const { meta } of records) {
     const cat = String(meta.category);