# What is Opt-out (AI training)?

> An AI training opt-out is a way for a content owner, customer or individual to say their content or data should not be used to train AI models. In EU copyright law it is an express reservation of rights against text and data mining, made by machine-readable means for content published online.

Source: https://sellmydata.com/glossary/opt-out/
Published: 2026-10-30

An AI training opt-out is any signal, setting or legal right used to keep content or data out of AI training. The term covers several mechanisms with different legal weight, and which one applies depends on whether the question is copyright, privacy or a software contract.

## Copyright: reserving rights against text and data mining

In the EU, Article 4 of the Copyright (DSM) Directive allows text and data mining (TDM, meaning software analysing content to find patterns) unless the rightsholder has expressly reserved their rights through machine-readable means ([arXiv](https://arxiv.org/pdf/2512.02047)). Recital 18 names metadata and website terms and conditions as examples, and says that in other cases rights can be reserved by contract or a unilateral declaration ([Kluwer Copyright Blog](https://legalblogs.wolterskluwer.com/copyright-blog/the-new-copyright-directive-text-and-data-mining-articles-3-and-4/)). The separate research exception in Article 3, which covers research organisations and cultural heritage institutions, cannot be overridden by an opt-out ([arXiv](https://arxiv.org/pdf/2512.02047)).

Courts are still settling what counts. In October 2025 the Danish Maritime and Commercial Court held that a scraping or data mining ban in a publicly accessible HTML privacy or data policy could qualify as an appropriate reservation, and in December 2025 the Hamburg Higher Regional Court gave the first substantive judicial treatment of machine-readable reservations in an AI context ([Kluwer Copyright Blog](https://legalblogs.wolterskluwer.com/copyright-blog/laion-round-2-machine-readable-but-still-not-actionable-the-lack-of-progress-on-tdm-opt-outs-part-1/)).

## Technical signals

The most common signal is robots.txt, a plain text file telling crawlers which pages they may visit. Its standard, RFC 9309, describes its rules as requests crawlers are asked to honor, not access authorization ([IETF](https://datatracker.ietf.org/doc/html/rfc9309)), and not a substitute for security measures ([RFC Editor](https://www.rfc-editor.org/info/rfc9309/)). It was not designed for TDM and cannot tell AI training apart from search indexing ([Springer](https://link.springer.com/article/10.1007/s40319-025-01664-8)). The EU General-Purpose AI Code of Practice, published 10 July 2025, calls for crawlers that follow robots.txt and other appropriate machine-readable protocols ([Freshfields](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-26-the-final-general-purpose-ai-code-of-practice-a-short-g-102ksv0)). The IETF's AIPREF working group is standardizing ways to express preferences about AI use of content ([IETF](https://datatracker.ietf.org/doc/charter-ietf-aipref/00-01/)); an early draft treats AI training as a form of TDM ([IETF draft](https://www.ietf.org/archive/id/draft-ietf-aipref-vocab-00.html)), and its vocabulary milestone was 31 August 2026, though whether it was met is unconfirmed ([IETF](https://datatracker.ietf.org/wg/aipref/about/)).

## Privacy law

In Opinion 28/2024 (17 December 2024), the European Data Protection Board says that when a controller relies on legitimate interest, the GDPR right to object applies, and names an opt-out list as one possible safeguard ([EDPB](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf)). California's CCPA gives consumers a right to opt out of the sale or sharing of their personal information ([CPPA](https://cppa.ca.gov/regulations/pdf/ccpa_statute_eff_20260101.pdf)), including through a global privacy control ([California AG](https://oag.ca.gov/privacy/ccpa)). It isn't specific to AI training; it applies whenever a business sells or shares personal information with third parties.

## Vendor settings

Software vendors set their own rules. In May 2024, Slack's privacy principles let workspace owners opt out of global model training by emailing a request ([SecurityWeek](https://www.securityweek.com/user-outcry-as-slack-scrapes-customer-data-for-ai-model-training/)), with no stated processing time ([The Register](https://www.theregister.com/2024/05/20/slack_ts_and_cs_update/)); current terms may differ. In February 2024, the FTC warned that quietly and retroactively changing terms to allow AI training may be unfair or deceptive ([FTC](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive)).

## Why it matters to a company selling data

An opt-out is a refusal; a license is a negotiated permission. Before licensing, it helps to know which opt-outs touch your data: promises in your privacy policy, limits in customer contracts, and settings in the tools you use. SellMyData runs a [rights check](https://sellmydata.com/rights/) covering contracts, privacy policy, employee communications and third-party content before anything is listed. This is general information; ask counsel how these rules apply to you.

## Frequently asked questions

### Is robots.txt a legal AI training opt-out?

robots.txt tells crawlers which pages they may visit, but its standard says these rules are requests, not access authorization, and it cannot separate AI training from search indexing. Whether a given signal counts as a valid rights reservation under EU law is still being tested in courts, so ask counsel about your situation.

### Is an AI training opt-out the same as the CCPA right to opt out?

Not exactly. The CCPA right lets a consumer tell a business not to sell or share their personal information with third parties, and it is not specific to AI training. Copyright reservations and vendor settings are aimed at training more directly.

### Does opting out of AI training stop all text and data mining in the EU?

No. An Article 4(3) reservation applies to the general text and data mining exception, but the separate research exception in Article 3, which covers research organisations and cultural heritage institutions, cannot be overridden by an opt-out. Ask counsel how this applies to your content.

### What does "Do not train" mean?

It is shorthand for an AI training opt-out: a signal, setting or contract term saying content or data should not be used to train AI models. It has no single legal definition, and how binding it is depends on the law and the mechanism used.

## Sources

- [Research paper on text and data mining opt-outs under the EU DSM Directive (arXiv 2512.02047)](https://arxiv.org/pdf/2512.02047)
- [The New Copyright Directive: Text and Data Mining (Articles 3 and 4)](https://legalblogs.wolterskluwer.com/copyright-blog/the-new-copyright-directive-text-and-data-mining-articles-3-and-4/)
- [Machine-readable but still not actionable: the lack of progress on TDM opt-outs (Part 1)](https://legalblogs.wolterskluwer.com/copyright-blog/laion-round-2-machine-readable-but-still-not-actionable-the-lack-of-progress-on-tdm-opt-outs-part-1/)
- [RFC 9309: Robots Exclusion Protocol](https://datatracker.ietf.org/doc/html/rfc9309)
- [RFC 9309 information page](https://www.rfc-editor.org/info/rfc9309/)
- [Journal article on robots.txt and text and data mining opt-outs (IIC, 2025)](https://link.springer.com/article/10.1007/s40319-025-01664-8)
- [EU AI Act unpacked #26: The final General-Purpose AI Code of Practice – a short guide](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-26-the-final-general-purpose-ai-code-of-practice-a-short-g-102ksv0)
- [AI Preferences (AIPREF) working group charter](https://datatracker.ietf.org/doc/charter-ietf-aipref/00-01/)
- [AI Preferences (aipref) working group: About page](https://datatracker.ietf.org/wg/aipref/about/)
- [Internet-Draft draft-ietf-aipref-vocab-00 (AI preferences vocabulary)](https://www.ietf.org/archive/id/draft-ietf-aipref-vocab-00.html)
- [Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf)
- [California Consumer Privacy Act statute text (effective 1 January 2026)](https://cppa.ca.gov/regulations/pdf/ccpa_statute_eff_20260101.pdf)
- [California Consumer Privacy Act (CCPA)](https://oag.ca.gov/privacy/ccpa)
- [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive)
- [User Outcry as Slack Scrapes Customer Data for AI Model Training](https://www.securityweek.com/user-outcry-as-slack-scrapes-customer-data-for-ai-model-training/)
- [News coverage of Slack's May 2024 terms and AI training update (20 May 2024)](https://www.theregister.com/2024/05/20/slack_ts_and_cs_update/)