Do Not Train AI Opt-Out Notice

What Is Do Not Train AI Opt-Out Notice?

A do-not-train AI opt-out notice is a combination of machine-readable signals and written declarations that tells AI developers and their web crawlers not to use your content or personal data to train, fine-tune, or otherwise build artificial intelligence models. It is not one form you sign and file. It is a stack: a robots.txt directive that names the AI crawlers and disallows them, a rights reservation in your site terms and metadata, embedded tags inside your image and video files, and, when the material involves personal information about you, a privacy-law opt-out request sent to the company. People reach for it after they realize their blog, portfolio, photo library, or face has probably already been swept into a training set, and they want to draw a line going forward.

No single law in the United States makes the tag binding on its own. That is the uncomfortable truth most templates skip. What gives the notice teeth varies by where you are and what you are protecting. In the EU, a properly expressed reservation under copyright law actually obligates AI providers to respect it. In the US, your force comes from your website's terms of service as a contract, from technically blocking access, and from consumer-privacy statutes that let you stop the sale, sharing, or processing of your personal data. The notice works best when you treat it as a layered defense rather than a magic incantation, because any one layer can be ignored by a crawler that decides the rules do not apply to it.

Why This Matters Now

The ground shifted because the web stopped being free for the taking. On 1 July 2025, Cloudflare, which sits in front of roughly a fifth of all websites, flipped the default: new domains it manages now block AI crawlers unless the owner opts in, and it launched a Pay Per Crawl marketplace letting publishers charge AI companies per page fetched. That single infrastructure change did more to operationalize do-not-train preferences overnight than years of voluntary robots.txt entries. Around it, the standards bodies moved. The IETF chartered its AIPREF working group in January 2025 to build a common vocabulary for AI usage preferences, attachable through robots.txt and a new Content-Usage HTTP header, after everyone admitted the existing tools were a patchwork.

The legal pressure is just as recent. The EU AI Act began applying its general-purpose AI obligations on 2 August 2025, and Article 53(1)(c) ties those models to the copyright opt-out in the older DSM Directive, so a machine-readable reservation in Europe is now something providers must address in a published policy. On the privacy side, the European Data Protection Board issued Opinion 28/2024 on 18 December 2024, confirming that when an AI developer relies on legitimate interest to train on personal data, the individual's right to object under GDPR Article 21 applies. Meta paused its EU AI training in mid-2024 after regulator pushback before resuming under an objection process. And the US courts spent 2025 drawing the copyright lines in real cases, which is why an opt-out filed today sits on top of a fast-moving and unsettled body of law rather than a finished one.

The Legal Backbone

Why a do-not-train notice is a layered defense, and what each layer can and cannot do

The reason you cannot just paste one paragraph and call it done is that no single layer covers the whole problem, and each one fails in a different way. Start with the technical layer, robots.txt and access controls. This is the only layer that physically stops a compliant crawler before it ever reads your content, which makes it the most effective tool you actually control. Its weakness is that compliance is voluntary. A crawler that ignores robots.txt, or that fetches your pages through a third party like Common Crawl, walks right past it. So robots.txt is necessary and never sufficient, and pairing it with infrastructure-level blocking, the kind Cloudflare now offers by default, closes part of the gap a polite text file leaves open.

The second layer is the rights reservation, and here geography decides everything. In the EU, expressing a machine-readable Article 4(3) reservation has bite, because the AI Act now obliges general-purpose model providers to identify and respect it. In the US, the identical tag carries no statutory force over a copyrighted work, and whether scraping for training is lawful at all is being fought out case by case, with judges so far split. What the reservation does buy you in the US is contract and notice: language in your terms of service can form an enforceable agreement against a party that accesses your site, and a documented, dated reservation is evidence you objected, which matters if the law shifts or if you later sue. The drafting can be identical on both sides of the Atlantic, but in Europe the reservation commands compliance while in the US it builds a contractual and evidentiary record.

The third layer is embedded metadata, IPTC's plus:DataMining property or a TDMRep statement, which keeps your opt-out attached to a photo or video even after it is copied, reposted, and stripped of its surrounding page. That persistence is its whole value, and its limit is that the file can be re-encoded and the metadata wiped, and that downstream tools have to choose to read it. The fourth layer, the privacy opt-out, is the one most people overlook and the one with the most reliable US teeth, because it does not depend on owning a copyright. If the training data is your name, face, posts, or other personal information, the CCPA and its sister statutes let you demand the company stop selling or sharing it, delete it, and in several states honor a Global Privacy Control signal automatically. The strongest notice runs all four layers at once and matches the demand to the harm: robots.txt and Cloudflare to block the fetch, an Article 4(3) reservation for EU copyright weight, IPTC or TDMRep tags for persistence, and a privacy opt-out for anything personal. Then keep dated copies of every layer, because the value of a do-not-train notice in the US is as much about proving you objected as about any single switch you flipped.

When You Need This

You run a website, blog, or online store and you want AI crawlers blocked going forward: add the AI user-agents to robots.txt, enable infrastructure blocking, and post a machine-readable rights reservation in your terms.

You are an author, journalist, or publisher whose text is being scraped for language models: combine a robots.txt block with an Article 4(3)-style reservation, and watch the publisher litigation since your leverage in the US is still contractual and evolving.

You are a photographer, illustrator, or designer and your images are training generative models: embed the IPTC plus:DataMining opt-out in your files and add TDMRep so the reservation survives reposting.

The material is personal information about you, your name, face, voice, or social posts, not a copyrighted work: file a CCPA or state-law opt-out of sale and sharing, request deletion, and enable Global Privacy Control in your browser.

You sell into or have audiences in the EU: express the reservation in machine-readable form so it triggers the AI Act Article 53 compliance duty on general-purpose AI providers.

You want a dated record that you objected before the law settles: keep timestamped copies of every signal and demand, because in the US the proof that you opted out may matter more than any one mechanism.

How to Fill Out Do Not Train AI Opt-Out Notice

1. 1. Inventory what you are protecting and where it lives

   List the assets and classify each. A website or blog is one thing, downloadable image and video files are another, and personal data about you is a third. The right layers differ: a site needs robots.txt and a terms reservation, image files need embedded metadata, and personal information needs a privacy opt-out. Note where each asset is hosted, because your control over robots.txt and headers depends on having access to the server or CDN.

2. 2. Block the AI crawlers in robots.txt

   Add explicit disallow rules for the documented AI training user-agents, including GPTBot, Google-Extended, ClaudeBot, CCBot, and PerplexityBot, and decide deliberately whether to also block the separate search and real-time crawlers each company runs. Place robots.txt at your domain root. Remember this is a request a crawler can ignore, so do not rely on it alone, and re-check the token list periodically because vendors add and rename bots.

3. 3. Turn on infrastructure-level blocking where you can

   If your site sits behind a CDN or security provider, enable its AI-crawler controls. Cloudflare, for example, now offers to block AI crawlers by default and runs a Pay Per Crawl option. Infrastructure blocking enforces at the network edge rather than asking nicely, so it stops crawlers that disregard robots.txt, and it is the biggest practical upgrade over a text file you can make in a single setting.

4. 4. Express a machine-readable rights reservation

   State a reservation of rights for text-and-data mining in your website terms and conditions and in a machine-readable form, the approach Article 4(3) of the EU DSM Directive contemplates. Implementing TDMRep across the site, or at minimum clear terms-of-service language prohibiting AI training use, gives you the EU copyright hook and, in the US, a contractual basis and dated evidence of your objection.

5. 5. Embed opt-out metadata in image and video files

   For visual work, write the IPTC Data Mining property (plus:DataMining in XMP) into each file using a metadata editor, choosing the value that prohibits AI or ML training. Because the tag lives inside the file, it travels with the image when it is copied or reposted. Re-export carefully, since some platforms strip metadata on upload, and keep your original masters with the tags intact.

6. 6. File privacy opt-outs for personal data

   If the material is personal information, send the company a CCPA or state-law request to opt out of sale and sharing and to delete your data, and enable Global Privacy Control in your browser so the objection is broadcast automatically to sites that honor it. In the EU or UK, send a GDPR Article 21 objection where the developer relies on legitimate interest. These do not untrain an existing model, but they govern future processing and create a record.

7. 7. Send a direct notice to identifiable developers

   Where you can name the AI company collecting your work, send a written do-not-train notice citing your reservation, your robots.txt directives, and, for personal data, the applicable privacy statute. Keep it specific: identify the assets, state the basis, and request confirmation. This converts a passive signal into an addressed demand and strengthens any later claim that the use continued after notice.

8. 8. Date, archive, and re-check on a schedule

   Save timestamped copies of your robots.txt, your terms reservation, your tagged files, and every opt-out request and response. Re-verify the crawler token list and your CDN settings on a recurring basis because the user-agents and the standards keep changing. The archive is the point: in the US, where the law is unsettled, proof that you reserved your rights on a specific date is often your strongest asset.

Key Terms Defined

Text and data mining (TDM)

The automated analysis of large amounts of text, images, or other content to extract patterns and information. Training an AI model on scraped web content is a form of TDM, which is why EU copyright exceptions written for TDM in Articles 3 and 4 of Directive (EU) 2019/790 now govern AI training in Europe.

Rights reservation (Article 4(3) opt-out)

A rightsholder's express statement, required to be machine-readable for online content, that reserves their works from the general text-and-data-mining exception. Under EU DSM Directive Article 4(3) a valid reservation removes the commercial TDM exception for that work, and the AI Act obliges general-purpose AI providers to respect it.

robots.txt and user-agent

A file at a website's root, standardized as RFC 9309, that lists which automated crawlers may access which paths. A user-agent is the identifying token a crawler sends, such as GPTBot or ClaudeBot. Disallowing a user-agent in robots.txt requests that the crawler stay out, though compliance is voluntary.

plus:DataMining (IPTC property)

A machine-readable metadata field introduced in the IPTC Photo Metadata Standard 2023.1 in October 2023, embedded in a file's XMP, whose controlled values can signal that data mining is prohibited, prohibited except for search indexing, or prohibited except for AI and ML training. It keeps the opt-out attached to the image itself.

Global Privacy Control (GPC)

A browser-level signal that automatically tells every website you visit that you opt out of the sale and sharing of your personal data. Several US state privacy laws, including those in California, Colorado, Connecticut, and Texas, require covered businesses to honor it as a universal opt-out mechanism.

Algorithmic disgorgement

An FTC enforcement remedy, grounded in Section 5 of the FTC Act, ordering a company to delete AI models or algorithms developed from improperly collected data. The agency has imposed it in matters including Cambridge Analytica, Everalbum, Kurbo, and Rite Aid. It is a government remedy, not a consumer self-service right.

Related Documents

Legal Authorities & Sources

This page is grounded in primary law. The statutes and official resources below are the authorities behind the guidance above. Verify the current text of any statute before relying on it.

• EU DSM Directive (EU) 2019/790, full text (EUR-Lex)
• EU AI Act, Regulation (EU) 2024/1689, Article 53 (artificialintelligenceact.eu)
• IETF AIPREF Working Group, Progress on AI Preferences
• IETF draft, Indicating Preferences regarding content usage (aipref-attach)
• IPTC Photo Metadata Standard (Data Mining property)
• C2PA Clarification: No TDM Assertions in the Standard (22 Jan 2026)
• EDPB Opinion 28/2024 on AI models and personal data
• California Attorney General, CCPA consumer rights
• Cloudflare, Content Independence Day: no AI crawl without compensation

Frequently Asked Questions