Metadata and provenance

Signalling AI data-mining preferences in image metadata

People in high-visibility waistcoats stand among vehicles in a street lined with brick buildings.
Workers in a narrow street.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

One XMP property lets a photograph state its owner's position on AI training. It is a declaration, not a lock, and its legal weight depends heavily on where the reader of the file is standing.

A photograph can carry a sentence about itself that reads, in effect, "not for AI training". It sits in the file's XMP as a single property, plus:DataMining, and it was approved by IPTC members on 4 October 2023 as part of Photo Metadata Standard 2023.1. It is the only widely published rights signal designed to travel with an individual image rather than with a website.

This article is general information, not legal advice. Checked on 18 November 2026.

What follows sets out what the property says, how to write it, what it does in law in the EU, the UK and the US, and the several ways it can fail.

What is the IPTC Data Mining property?

The property was developed with the PLUS Coalition and uses a controlled vocabulary, so software reads a fixed identifier instead of interpreting prose. There are nine values:

Value Meaning in practice
Unspecified - no prohibition defined Nothing is stated either way
Allowed Data mining is permitted
Prohibited for AI/ML training No use for training any AI or machine-learning model
Prohibited for Generative AI/ML training No use for training generative models specifically
Prohibited except for search engine indexing Search engines may index the file's location; nothing else
Prohibited No data mining at all
Prohibited, see Other Constraints property Conditions are set out in the companion text field
Prohibited, see Embedded Encoded Rights Expression property Conditions are in an embedded machine-readable rights expression
Prohibited, see Linked Encoded Rights Expression property Conditions are in a linked machine-readable rights expression

Each has a URI of the form http://ns.useplus.org/ldf/vocab/DMI-PROHIBITED-AIMLTRAINING. A companion property, Other Constraints (plus:OtherConstraints), holds human-readable conditions.

Two statements in the specification deserve more attention than they get. Regional laws may prohibit, constrain or allow data mining for certain purposes regardless of the value chosen. And the absence of a prohibition does not mean the owner grants permission. "Unspecified" is not "Allowed".

How do you add an AI training opt-out to a photo's metadata?

ExifTool added support in version 12.67. For JPEGs, the free AI Training Opt-Out Metadata tool writes the property and optional Other Constraints text into the file's XMP in the browser, without uploading or recompressing the image. To confirm what a file carries, open it in the EXIF and metadata viewer, which highlights the Data Mining field when present.

The property is most useful next to the ordinary rights fields. A reservation with no named owner gives a careful licensee nobody to ask, so pair it with Creator, Copyright Notice and Web Statement of Rights, which the Add Copyright Metadata tool can write. The wider field set is covered in the guide to IPTC fields every archive should fill.

It depends on jurisdiction, and in every jurisdiction part of the answer is pending.

European Union. Article 4 of Directive (EU) 2019/790 permits text and data mining of lawfully accessed works by anyone, for any purpose, unless the rights holder has expressly reserved the right "in an appropriate manner, such as machine-readable means in the case of content made publicly available online". A valid reservation is therefore the thing that matters. What is decided: the Hanseatic Higher Regional Court in Hamburg held on 10 December 2025, in Kneschke v LAION, that a reservation written in ordinary language in a website's terms was not machine-readable. What is pending: the German Federal Court of Justice heard the further appeal on 3 September 2026, and PhotoWorkout reported that it would announce its decision on 17 December 2026. Separately, the European Commission consulted between December 2025 and January 2026 on which machine-readable protocols AI providers should respect, with the stated aim of publishing a list of generally agreed solutions. CEPIC reported strong disagreement among rights holders, and a final list could not be confirmed. robots.txt is the only protocol named in the general-purpose AI Code of Practice itself. So the IPTC property is a serious candidate for an Article 4(3) reservation, built for the purpose and machine-readable by design, but its formal recognition is not something to assert yet.

United Kingdom. There is no opt-out regime. An exception with rights reservation was the government's stated preference in the consultation that ran from 17 December 2024 to 25 February 2025, and it was dropped in the report published on 18 March 2026. The existing exception, section 29A of the Copyright, Designs and Patents Act 1988, covers copying for computational analysis only for non-commercial research by someone with lawful access. In the UK the property records the owner's position; it does not switch a statutory permission off, because for commercial training there is no such permission to switch.

United States. We know of no statute that gives a metadata reservation a defined effect, and we found no court ruling on fair use for image models when we checked.

One adjacent point applies in both the UK and US. Rights information embedded in a file is protected in its own right. Section 296ZG of the 1988 Act gives a remedy against a person who knowingly and without authority removes or alters electronic rights management information, where they know or have reason to believe that doing so induces, enables, facilitates or conceals infringement. 17 U.S.C. § 1202(b) contains a comparable prohibition on intentionally removing copyright management information, with its own knowledge requirements. Both turn on knowledge and intent, and neither makes routine automated stripping unlawful as such.

Where the signal fails

The limits are technical before they are legal.

  • Stripping. Many social platforms, content delivery networks and image-processing libraries discard XMP when they resize or re-encode. A copy without the property says nothing.
  • Not reading. A crawler has to parse the file's XMP to see the value. Few AI developers publish whether theirs do.
  • Copies you do not control. The property protects files you marked. Earlier unmarked copies, screenshots and third-party reposts are outside it.
  • No enforcement. Nothing stops a file with "Prohibited" in it from being copied. The property creates evidence of a stated position and, in the EU, possibly a legally relevant reservation. It is not access control.
  • Formats. The browser tool writes JPEG XMP. Other formats need other tooling.

IPTC, robots.txt, TDMRep, RSL and C2PA: which signal does which job?

Signal Where it lives Scope Status
IPTC Data Mining Inside the image file (XMP) One image, wherever it travels IPTC standard since October 2023
robots.txt Web server A site or path, per crawler Long-established convention; named in the EU Code of Practice; no statutory status in the UK or US
TDMRep HTTP header, a tdmrep.json file on the server, or embedded properties Site, path or publication W3C Community Group reports (latest version 9 August 2024); not a W3C Recommendation
RSL 1.0 Machine-readable licence terms that extend robots.txt Site-level licensing terms Published 10 December 2025; backers named in the announcement include the Associated Press, The Guardian, Vox Media, Cloudflare and Akamai; it names no AI developer
C2PA Content Credentials Signed manifest in or alongside the file Provenance of one asset Not a rights reservation

TDMRep is built around Article 4 and describes itself as a rights-reserved signal, not a "no" signal: its tdm-policy field points to where a licence can be obtained. RSL is aimed at stating licence terms for crawlers. C2PA does a different job entirely, explained in Content Credentials for photo archives.

A sensible arrangement for an image library is layered: the IPTC property in every delivered file, a site-level signal on the server, and a clear human-readable statement of terms. Each covers a gap the others leave.

A reservation is also an invitation

"Prohibited" can sound like a closed door. In practice a reservation is what makes a licence meaningful, since permission has value only where it is required. That is why the property should always sit beside a working contact route.

The Robert Harper Photograph Library is owned by 825 Ventures Ltd, the owner and rights holder, and its position is set out under rights and stewardship. More than half of the archive is still on film and has yet to be scanned, and only a small selection has been published online by Archive.London, including the studio work catalogued under jewellery and the other still-life categories. Licences for AI and data uses are negotiated directly with the single rights holder, subject to rights review. Teams assessing licensed image data can read the AI training data licensing page and make contact from there.

Frequently asked questions

How do I opt my photos out of AI training using metadata?

Write the IPTC Data Mining property into each file's XMP with a value such as 'Prohibited for AI/ML training'. ExifTool has supported the property since version 12.67, and browser tools can add it to JPEGs without recompressing. It records a rights reservation in machine-readable form. It does not technically prevent copying, and it vanishes if the metadata is stripped.

Is the IPTC Data Mining property legally binding?

Not by itself, and not uniformly. In the EU it is a candidate for the machine-readable reservation that Article 4(3) of the 2019 copyright directive requires, though the Commission had not confirmed a final list of recognised protocols when checked on 18 November 2026. In the UK and US no statute gives the property a defined effect. This is general information, not legal advice.

What is the difference between the IPTC Data Mining property and robots.txt?

robots.txt speaks for a website and tells crawlers which paths they may fetch. The IPTC Data Mining property speaks for one image and stays with the file when it is copied elsewhere. robots.txt cannot follow an image onto another site; embedded metadata cannot survive a pipeline that strips it. They cover different gaps and are best used together.

Do AI companies respect image metadata opt-outs?

There is no public, audited answer. Signatories to the EU's general-purpose AI Code of Practice commit to honouring robots.txt and other recognised machine-readable reservations, but robots.txt is the only protocol the Code names. Whether a given crawler reads embedded XMP at all is rarely disclosed, and many image pipelines discard metadata before any check could happen.

Does Content Credentials (C2PA) include an AI training opt-out?

No. On 22 January 2026 the C2PA stated that its specification contains no standard text and data mining assertion and that such signals are outside its scope. Content Credentials record how a file was made and edited. A rights reservation has to be expressed another way, for example with the IPTC Data Mining property.

Sources

Published by Archive.London. Corrections are welcome under our editorial policy.

For picture agencies, publishers and researchers

Fashion, beauty, jewellery, music, portraits and British life from the 1960s onwards, with 52,684 analogue originals still to be scanned. Tell us what you are researching.

See representation and researchWrite to the archive