4allhuman

🛡️ Don't Train On Me

Protect any creator's content from AI training in 60 seconds.

Paste your domain → get a protection score → receive ready-to-deploy opt-out artifacts:

1. robots.txt section blocking 170+ known AI crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot, Bytespider…) — doubling as a machine-readable EU DSM Directive Art. 4(3) rights reservation
2. /ai.txt — Spawning-spec machine-readable access policy
3. <meta name="robots" content="noai, noimageai"> tags for every page head
4. Legal notice of reserved rights

Why it matters (the legal hook)

• EU DSM Directive Art. 4(3): creators may expressly reserve their content from text-and-data-mining "in an appropriate manner, such as machine-readable means."
• EU AI Act Art. 53(1)(c): GPAI providers must identify and comply with those reservations — honoring this output is a regulatory obligation for them.
• US: no opt-out statute; fair use is uncertain (USCO Part 3 report). Blocking crawlers + keeping content behind access control is currently the strongest lever.

Research notes:

• research/protecting-human-content-from-ai-training.md — legal & technical landscape (EU DSM Art. 4(3), AI Act Art. 53(1)(c), USCO Part 3, Glaze/Nightshade, IETF aipref, C2PA)
• research/enhancement-research.md — verified build opportunities: Common Crawl CDX corpus check, DE-COP memorization detection, RSL licensing standard, x402 payments

What this can and cannot guarantee

No tool can make already-public content untrainable. Anyone can download what anyone can read. Our artifacts are signals + legal reservations, not force fields:

This tool sits in rows 2–3: it measures exposure, perfects the strongest legal lever available, and verifies deployment. For a true guarantee, keep content behind access control.

See research/protecting-human-content-from-ai-training.md for sources.

Design principles (first-principles trust)

The product reduces an information asymmetry between creators and AI trainers. Trust follows from three commitments baked into the UI:

1. Verifiability over assertion — every scan result links to the raw evidence (the live robots.txt we read), with a UTC timestamp. Users can re-do any check by hand.
2. Honest limits, stated up front — the "What this can / cannot do" module says plainly: signals bind compliant crawlers only; we cannot audit what a model was trained on; access control is the only complete protection. Overclaiming is the fastest way to lose trust.
3. No black boxes — fixed, published score weights; no accounts; nothing stored; open-source methodology.

Visual language follows Teenage Engineering's industrial design: flat panels, hard edges, warm-grey chassis, signature orange accents, monospace labels, LED status indicators, numbered modules — instruments, not marketing.

CLI

Run the same audit from your terminal or CI - no config, zero dependencies:

npx 4allhuman audit example.com

npx 4allhuman audit example.com [--json] [--fail-below <0-100>] [--timeout <ms>]

  --json                Machine-readable report (all layers, verdicts, score).
  --fail-below <0-100>  Exit 1 when the score is below this threshold (CI gate).
  --timeout <ms>        Per-request timeout (default 8000).

The report prints one plain-language verdict per layer (PROTECTED / PARTIAL / MISSING / NOTE) with the evidence behind it, then the total score out of 100.

Exit codes: 0 audit ran and meets --fail-below (if given) · 1 score below --fail-below · 2 usage error or the site could not be reached.

Runs directly from TypeScript via Node's type stripping - requires Node >= 22.18. The CLI and the web app share the exact same audit engine (src/lib/scanner.ts + src/lib/layers.ts), so a terminal score always matches a web score.

Run it

npm install
npm run dev        # http://localhost:3000
npm run update:crawlers   # refresh the crawler blocklist from the ai.robots.txt community dataset
npm test               # RFC 9309 + multi-layer audit tests (node:test, zero deps)
npm run typecheck      # tsc --strict, no emit

Architecture

bin/
└── 4allhuman.ts          # CLI entry point (shebang, zero deps, Node >= 22.18)
src/
├── cli.ts                # CLI core: arg parsing, report rendering, exit codes
├── app/
│   ├── page.tsx              # scan UI + artifact viewer with copy buttons
│   └── api/protect/route.ts  # orchestrator: scan → generate → score
└── lib/
    ├── crawlers.ts           # re-exports the crawler blocklist + legal text
    │   ├── crawlers.generated.ts  # 170+ crawlers, generated from the community ai.robots.txt dataset
    ├── robots9309.ts         # RFC 9309 matcher: group selection, longest-match, wildcards
    ├── layers.ts             # multi-layer audit: X-Robots-Tag, TDMRep, aipref, llms.txt, meta
    ├── scanner.ts            # live scan, per-crawler verdicts with rule-level reasons
    └── generator.ts          # artifact generation + layered protection score
tests/
├── robots9309.test.ts        # spec-case tests: ties, wildcards, group precedence
└── layers.test.ts            # multi-layer parsing + score tests

Radar (module 09) — manual probe mode

Paste an article (or its URL) → get generated probe questions → ask ChatGPT / Perplexity / Claude / Gemini yourself → paste their answers back → 8-gram plagiarism forensics run locally in your browser.

No API keys. No AI-company calls from our servers. Nothing stored. The evidence pack (.json download) seals the source text with SHA-256 and records every verdict with matched spans.

Thresholds: ≥10% word-8-gram containment = COPIED, 2–9% = SUSPICIOUS, else CLEAN. 8 consecutive shared words is the classical plagiarism-forensics threshold — chance co-occurrence is effectively zero.

The multi-layer audit

One report across every opt-out standard, each layer with its own plain-language verdict and evidence:

The same layer builders score user-pasted artifacts in verify-your-fix, so the before/after score is computed identically to a live scan.

Modules

Key references

• Directive (EU) 2019/790 Arts. 3–4 (TDM opt-out) — EUR-Lex
• Regulation (EU) 2024/1689 (AI Act) Art. 53(1)(c)–(d) — EUR-Lex
• U.S. Copyright Office, Copyright and AI Part 3: Generative AI Training (2025) — copyright.gov/ai
• DE-COP: detecting copyrighted content in training data — arXiv 2402.09910
• Beyond Public Access in LLM Pre-Training Data — arXiv 2505.00020
• Stealing Part of a Production Language Model (canary extraction) — arXiv 2403.06634
• RSL 1.0 Really Simple Licensing — rslstandard.org · RSL Collective
• IETF AI Preferences WG — datatracker · C2PA — c2pa.org
• Glaze / Nightshade — UChicago SAND Lab · Spawning DO NOT TRAIN — spawning.ai
• Common Crawl corpus & index API — commoncrawl.org

Roadmap (see TASKS.md)

• Scanner, generators, scoring, UI
• One-click deploy: open a PR to the site's repo / upload via FTP
• C2PA "Proof of Human" content signing
• Spawning DO NOT TRAIN registry submission