How do I opt out of AI training on my images?

You can add each operator's robots.txt token to your site, such as GPTBot, ClaudeBot or Google-Extended, and write the IPTC Data Mining property into your files. These are signals that operators may honour; they do not enforce anything.

Does robots.txt stop AI from using my images?

It asks. RFC 9309 says crawlers are requested to honour the rules and that they are not access authorization, and the operators' pages describe their tokens as indicating or signalling.

Does the EU require AI developers to honour an opt-out?

In the EU, Article 4(3) of Directive (EU) 2019/790 makes a reservation count: the Article 4 text and data mining exception applies only where rightholders have not expressly reserved use in an appropriate manner, such as machine-readable means. Member States' own laws implement this, and the Article 3 research exception is separate.

Can you stop AI training on your images? The opt-out signals that exist

You can send published signals, not enforce them: a robots.txt token per crawler (GPTBot, ClaudeBot, Google-Extended, CCBot, Applebot-Extended), the IPTC Data Mining property inside the image file, and in the EU a machine-readable reservation under Article 4(3) of Directive (EU) 2019/790. Outside that EU rule, honouring them is voluntary.

Published October 30, 2026. Drafted with AI assistance and checked against the sources linked on this page. How we write our guides.

Opt out of AI training: robots.txt tokens, one per operator

Each operator below documents its own robots.txt token on its own page, and each page says what the token does for that operator only. The wording matters: none of these pages says that a token deletes anything already collected or binds anyone else. They use verbs such as "indicates" and "signals".

robots.txt tokens, as each operator describes them
Operator	Token	What the operator's page says
OpenAI	`GPTBot`	Disallowing GPTBot indicates a site's content should not be used in training generative AI foundation models. OAI-SearchBot, for search, is a separate setting.
Anthropic	`ClaudeBot`	When a site restricts ClaudeBot access, it signals that the site's future materials should be excluded from Anthropic's AI model training datasets.
Google	`Google-Extended`	Lets a site manage whether content Google crawls may be used for training future generations of Gemini models that power Gemini Apps and Vertex AI API for Gemini, and for grounding in Gemini Apps and Grounding with Google Search on Vertex AI; it does not affect inclusion in Google Search.
Common Crawl	`CCBot`	Common Crawl publishes its robots.txt instruction to prevent it crawling your site, and says it is aware of crawlers falsely identifying themselves as CCBot.
Apple	`Applebot-Extended`	Publishers can opt out of their content being used to train Apple's general purpose foundation models. Applebot-Extended does not crawl webpages.

Anthropic's and Common Crawl's pages show the same form for blocking a whole site: a line such as User-agent: ClaudeBot followed by Disallow: /. Anthropic says to add it to the robots.txt file in your top-level directory and to do this for every subdomain you wish to opt out from. We suggest one such pair of lines per token you want to use. Google's page says Google-Extended has no separate user agent string; the token is used in a control capacity. OpenAI's page says that for ChatGPT-User, used for certain user actions, robots.txt rules may not apply, because a user initiates those actions.

These tokens live on a website, not in a photo file. RFC 9309 places the robots.txt file at the top-level path of the service, for example https://www.example.com/robots.txt. In our reading, that means your own robots.txt speaks only for your own site, not for pictures you post on someone else's platform. Our guide to protecting art online looks at the wider picture.

The IPTC data mining property in the image file

The IPTC Photo Metadata Standard has a Data Mining property (XMP name plus:DataMining) described as "Data mining prohibition or permission, optionally with constraints." It was added in version 2023.1, and its value must be one of the URIs in the PLUS controlled vocabulary. IPTC's specification and User Guide list the values, which include Allowed, Prohibited, Prohibited for AI/ML training, Prohibited for Generative AI/ML training and Prohibited except for search engine indexing.

IPTC's own cautions are plain. The absence of a prohibition does not indicate that the owner grants permission, and regional laws may overrule the value you select. The property is a statement stored in the file. Nothing on IPTC's pages says which crawlers read it. Our copyright metadata tool writes Creator, Copyright notice, Web statement of rights and Licensor URL; it does not write the Data Mining property. Free, no limit. Your pixels are not touched; only the metadata changes. For the field families, see EXIF, IPTC and XMP.

Content Credentials and the data mining assertion

Earlier versions of the C2PA specification defined a training and data mining assertion. Version 1.4 (November 2023) labelled it c2pa.training-mining, and the C2PA specification no longer includes it from version 2.0 (January 2024). It is now defined by the Creator Assertions Working Group (CAWG), a working group of the Decentralized Identity Foundation, as cawg.training-mining, version 1.1, ratified May 16, 2025. It has four entries: cawg.data_mining, cawg.ai_inference, cawg.ai_generative_training and cawg.ai_training, each set to allowed, notAllowed or constrained. CAWG describes it as information for a "C2PA Manifest Consumer", inside the cryptographically signed assertions of a C2PA manifest; manifests are covered in our Content Credentials guide. The specification does not say which AI trainers read it, and our tools do not write C2PA manifests.

The EU text and data mining reservation

This is the one place where honouring a signal is not only voluntary, and it is EU law, not US law. Under Article 4(3) of Directive (EU) 2019/790 on copyright in the Digital Single Market, the text and data mining exception in Article 4(1) applies on condition that the use of the works "has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online". In the EU, that reservation is a condition of the exception for users who rely on it, as the directive's text sets it out. Article 4(4) says Article 4 does not affect Article 3, a separate exception for text and data mining for scientific research by research organisations and cultural heritage institutions. Which signals count as an "appropriate" reservation is not decided by the directive, and it is applied through each Member State's own law.

Is the noai tag a standard, and who honours it?

No standards body or AI operator page we cite defines the noai tag, and none of the operator pages above says it honours it. Google's robots meta tag documentation, last updated March 24, 2026, does not list it.

Honouring these signals is voluntary

RFC 9309, the Robots Exclusion Protocol, describes robots.txt rules as ones crawlers are requested to honour, and says they are not a form of access authorization. Outside the EU rule above, every signal on this page is a request.

This is general information, not legal advice.

Sources

Checked October 30, 2026.

Questions

How do I opt out of AI training on my images?: You can add each operator's robots.txt token to your site, such as GPTBot, ClaudeBot or Google-Extended, and write the IPTC Data Mining property into your files. These are signals that operators may honour; they do not enforce anything.
Does robots.txt stop AI from using my images?: It asks. RFC 9309 says crawlers are requested to honour the rules and that they are not access authorization, and the operators' pages describe their tokens as indicating or signalling.
Is the noai tag a standard?: No standards body or AI operator page we cite defines it, and none of the operator pages we cite says it honours it.
What is the IPTC Data Mining property?: A property in the IPTC Photo Metadata Standard that states a data mining prohibition or permission, optionally with constraints, written into the file as plus:DataMining.
Does the EU require AI developers to honour an opt-out?: In the EU, Article 4(3) of Directive (EU) 2019/790 makes a reservation count: the Article 4 text and data mining exception applies only where rightholders have not expressly reserved use in an appropriate manner, such as machine-readable means. Member States' own laws implement this, and the Article 3 research exception is separate.