{
 "acronym": "",
 "adoption": {
  "by": null,
  "level": "unmeasured",
  "probe_run": null
 },
 "aliases": [],
 "analog": "robots.txt, extended from crawl permission to AI-specific permissions (training, indexing, caching) with licensing and attribution terms",
 "collisions": [
  "At least four unrelated, mutually incompatible specifications share the \"ai.txt\" name (see dispute): the IETF Internet-Draft draft-car-ai-txt-wellknown (training/scraping/indexing/caching policy), kaylacar/ai-txt (a separate IANA-filed proposal for the same policy domain), the arXiv 'ai.txt: A Domain-Specific Language' paper (Disallow: Return / Disallow: Train directives), and aitxt.ing ('a universal standard for describing any resource to AI assistants', a content-description format unrelated to policy)"
 ],
 "discovery": {
  "mechanism": "Fixed filename at the site root, following the /.well-known/ convention on some proposals",
  "path": "/ai.txt or /.well-known/ai.txt (varies by proposal)"
 },
 "dispute": {
  "claims": [
   "The IETF draft-car-ai-txt-wellknown Internet-Draft presents itself as the well-known-URI registration proposal for AI usage policy",
   "kaylacar/ai-txt independently claims its own IANA filing and IETF draft for the same purpose",
   "The arXiv DSL paper proposes a third, syntactically different grammar for the same policy domain without reconciling with either draft",
   "aitxt.ing uses the identical filename for a content-description standard unrelated to permissions, which will collide in practice with any site trying to adopt a policy-declaration ai.txt"
  ],
  "summary": "\"ai.txt\" is used by at least four unrelated efforts: two independent IETF-style Internet-Drafts covering the same training/scraping/indexing/caching policy domain (draft-car-ai-txt-wellknown and kaylacar/ai-txt), an academic DSL paper covering similar ground with different syntax, and a commercial product (aitxt.ing) that repurposes the name for an unrelated content-description use case."
 },
 "evidence": [
  {
   "claim": "IETF Internet-Draft draft-car-ai-txt-wellknown-00 defines '/ai.txt' as a well-known URI declaring site-wide Training/Scraping/Indexing/Caching defaults (allow/deny/conditional), per-path Training-Allow/Training-Deny overrides, licensing and fee fields, and per-agent blocks (e.g. distinct rules for ClaudeBot vs. GPTBot)",
   "date": "2026",
   "label": "VERIFIED",
   "url": "https://www.ietf.org/archive/id/draft-car-ai-txt-wellknown-00.html"
  },
  {
   "claim": "A separate GitHub repository, kaylacar/ai-txt, independently proposes 'an open standard proposal for declaring a website's AI policy: training, licensing, attribution, per-agent rules', claiming IANA filing #76/#77 'under review' and an IETF Internet-Draft on Datatracker; created 2026-02-22, last updated 2026-07-18",
   "date": "2026-07-18",
   "label": "VERIFIED",
   "url": "https://github.com/kaylacar/ai-txt"
  },
  {
   "claim": "An arXiv paper, 'ai.txt: A Domain-Specific Language for Guiding AI Interactions with the Internet', independently defines its own ai.txt DSL with Disallow: Return and Disallow: Train directives, explicitly positioned as an extension of robots.txt's model of declared-not-enforced regulation",
   "date": "2025",
   "label": "VERIFIED",
   "url": "https://arxiv.org/html/2505.07834v1"
  },
  {
   "claim": "aitxt.ing markets itself as 'a universal standard for describing any resource to AI assistants using plain text, preventing hallucination' — a content-description product, not a policy-permission format, despite sharing the ai.txt name",
   "date": "2026",
   "label": "THIN",
   "url": "https://aitxt.ing/"
  }
 ],
 "form": "well-known-file",
 "governance": "None; at least three independent proposals for the same policy-declaration purpose (an IETF Internet-Draft by 'car', kaylacar/ai-txt, and an arXiv DSL paper) plus a fourth, differently-scoped product (aitxt.ing) all use the same filename with no shared authority",
 "group": null,
 "history": [
  {
   "date": "2026-09-06",
   "note": "new record added during read-step verification; reclassified to the permit step after finding the name covers multiple unrelated policy proposals plus one unrelated content-description product"
  }
 ],
 "id": "ai-txt",
 "implementation": "prototype",
 "last_verified": "2026-09-06",
 "name": "ai.txt",
 "native_status": "IETF Internet-Draft (draft-car-ai-txt-wellknown-00) presented as a formal well-known-URI registration proposal; kaylacar/ai-txt separately claims IANA filing (#76/#77, 'under review')",
 "origin": {
  "date": "2026-02-22",
  "org": "Contested; earliest identified dated artifact is the kaylacar/ai-txt GitHub proposal"
 },
 "part_of": null,
 "parties": [
  "agent-site"
 ],
 "phase": "idea",
 "probe": {
  "method": "GET",
  "on_absent": "No ai.txt under any identified convention; given active fragmentation and low visibility, absence does not indicate a deliberate policy choice",
  "parse": "200 response with a text body containing top-level Training/Scraping/Indexing/Caching directives (allow/deny/conditional) and optional per-Agent blocks, per the IETF draft's grammar; other proposals use different field names for the same purpose, or an unrelated content-description grammar (aitxt.ing)",
  "path": "{origin}/ai.txt"
 },
 "question": "May AI systems train on, scrape, index, or cache this site's content, and under what per-agent or per-path conditions?",
 "scope": "general",
 "spec_url": "https://www.ietf.org/archive/id/draft-car-ai-txt-wellknown-00.html",
 "step": "permit",
 "summary_md": "**Moved from the Read step to Permit.** ai.txt is not read-step content — every substantive proposal found (excluding the differently-scoped aitxt.ing) is a permissions declaration: whether AI systems may train on, scrape, index, or cache a site's content, comparable to robots.txt extended with AI-specific and licensing semantics [1][2][3]. It is filed here under \"Permit\" (what may the agent do here?) rather than \"Read\", and is reported with this corrected step.\n\nWhat it is, per the most detailed version found (IETF Internet-Draft draft-car-ai-txt-wellknown-00): a text file at `/ai.txt` declaring site-wide defaults for four fields — Training, Scraping, Indexing, Caching — each `allow`/`deny`/(`conditional` for Training only), with path-scoped `Training-Allow`/`Training-Deny` overrides, a `Training-License` and `Training-Fee` for conditional training, and per-agent blocks that can override the defaults for a named bot (e.g. allowing ClaudeBot to train while denying GPTBot) [1]. A second, independently authored proposal, kaylacar/ai-txt, targets the same policy domain (training, licensing, attribution, per-agent rules) and separately claims its own IANA filing and IETF draft [2]. A third, an arXiv-published domain-specific language, defines yet a different syntax (`Disallow: Return`, `Disallow: Train`) for largely the same purpose, explicitly modeled on robots.txt's declared-but-unenforced approach [3].\n\nIt solves, in principle, a real gap: robots.txt has no notion of \"may train on\" versus \"may crawl for search indexing\" versus \"may serve in a live AI answer\", and site owners increasingly want to distinguish these. Per-agent, per-path granularity and an attached licensing/fee mechanism go further than robots.txt's binary allow/disallow.\n\nIt does not currently solve that gap in a way any agent could rely on, because — as with agents.txt — there is no single ai.txt an agent can fetch and trust: at least two IETF-style drafts and one academic DSL independently claim the same well-known filename for the same purpose, with no evidence any converged or even cited one another. A fourth party, aitxt.ing, uses the identical filename for an unrelated content-description product (\"preventing hallucination\" by describing resources to AI assistants), which would collide outright with any site that adopted a policy-declaration ai.txt at the same path [4].\n\nCurrent state, as of this verification (September 2026): the earliest dated artifact identified is the kaylacar/ai-txt repository, created 2026-02-22 [2]; the arXiv DSL paper is dated 2025 [3]; the IETF draft-car draft carries no clear publication date in the fetched text. Enforcement is, by every proposal's own admission, declarative only — like robots.txt, nothing prevents a non-compliant crawler from ignoring the file.\n\nWhat is disputed or unknown: which (if any) proposal has priority, whether any has actual publisher adoption, and how the aitxt.ing name collision would be resolved in practice. This record should be treated as thin and revisited once (or if) one proposal gains distinguishable traction.",
 "track": "community",
 "version": {
  "date": null,
  "label": "draft-car-ai-txt-wellknown-00"
 },
 "url": "https://agenticweb.wiki/protocols/ai-txt/",
 "category": "permit",
 "type": "protocol"
}