Metadata and provenance

How to protect your photographs from AI scraping: what works

A person stands at the open cabin door of a white high-wing aircraft.
Boarding a light aircraft.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Nothing a photographer can add to a published image stops it being copied. What the available measures do is state a position, add friction and leave evidence. Used together, that is worth having.

There is no setting, tag or mark that stops a published photograph being copied by a machine. If a browser can display the picture, a script can save it. Everything on offer to photographers and archives falls into one of three weaker categories: a signal that states your position, friction that makes copying less convenient, or a record that helps you later. Those are still worth having, provided nobody mistakes them for a lock.

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

What actually stops AI scraping, and what only asks?

Measure What it does What it cannot do
Not publishing, or publishing only small files Removes the usable copy from reach Help with copies already circulating
robots.txt Asks named crawlers not to fetch your pages or images Bind a crawler that ignores it; cover other sites
IPTC Data Mining metadata Puts a machine-readable reservation inside each file Survive a platform that strips metadata
Server or network bot blocking Refuses requests from identified bots Catch bots that disguise themselves
Visible watermark Deters reuse, identifies the owner Prevent download; resist cropping or retouching
Cloaking (Glaze, Nightshade) Adds pixel changes meant to mislead a model Promise lasting effect; its makers say so
Content Credentials (C2PA) Records how a file was made and edited Reserve rights or refuse training
Hashes and dated records Prove what you held and when Deter anyone

The first row is the only one that prevents anything. The rest are layers.

Does robots.txt stop AI crawlers taking images?

It stops the ones that agree to be stopped. The standard itself, RFC 9309, published in September 2022, is blunt: the rules "are not a form of access authorization", and the protocol "is not a substitute for valid content security measures".

In practice the large developers publish tokens for their training crawlers. OpenAI says that disallowing GPTBot "indicates a site's content should not be used in training". Anthropic describes ClaudeBot as the crawler that collects material which could contribute to training. Google-Extended and Applebot-Extended are not crawlers at all but control tokens: they tell Google and Apple how content already fetched by their ordinary crawlers may be used, and Google states that Google-Extended does not affect inclusion or ranking in Search. These are the companies' own descriptions of what their crawlers do. We have not tested whether any of them complies in practice.

Two limits matter for photographers. The file speaks for your domain only, so a picture you licensed to a magazine, or one reposted without permission, is governed by somebody else's robots.txt. And some fetchers sit outside it by design: OpenAI says that for requests triggered by a user, "robots.txt rules may not apply". The detail is in the plain guide to robots.txt for AI crawlers.

Do metadata opt-outs work?

The IPTC Photo Metadata Standard has carried a Data Mining property since version 2023.1, approved on 4 October 2023. It lets a file say "Prohibited for AI/ML training" in a fixed vocabulary that software can read. The free AI training opt-out metadata tool writes it into a JPEG in the browser, and the copyright metadata tool adds the creator, notice and licensing link that should sit beside it.

Its strength is that it travels with the file. Its weakness is that metadata is routinely discarded. In IPTC's 2019 tests of sixteen services, several, including Twitter and three free stock sites, delivered saved or downloaded files with the metadata stripped. A copy without the property says nothing, and few AI developers publish whether their pipelines read embedded XMP at all. More on this in what happens to metadata when you upload a photo and in the fuller account of the IPTC data mining property.

One point in the property's favour is legal. In the UK, section 296ZG of the Copyright, Designs and Patents Act 1988 gives a remedy against someone who knowingly and without authority removes electronic rights management information, where they know or have reason to believe this will enable or conceal infringement. It turns on knowledge, and it does not make routine automated stripping unlawful as such.

Do watermarks stop AI training?

No. A watermark is a deterrent and a label. It makes a picture less attractive to lift, tells a viewer whose it is, and makes unlicensed use easier to recognise. It can be cropped, painted out or simply ignored by a pipeline that does not care about quality.

Watermarks have had one notable legal moment. In Getty Images v Stability AI [2025] EWHC 2863 (Ch), decided on 4 November 2025, the High Court found limited, historic trade mark infringement where watermarks had appeared in outputs of some early versions of the model. That finding concerned trade marks in outputs. It was not a ruling that training on watermarked images infringes copyright, a question the court never decided.

If you watermark, do it on the web-size copy only and keep the clean master offline. The watermark tool applies a text mark in batches without uploading the files.

What about Glaze, Nightshade and other cloaking tools?

Glaze and Nightshade come from the SAND Lab at the University of Chicago. Glaze alters an image slightly so that a model fine-tuned on it learns a misleading style. It was built for artists facing style mimicry, though the project's own samples include photography.

The team's stated limits are the ones to remember. Changes are "more visible on art with flat colors and smooth backgrounds", which describes a good deal of studio photography. Glaze "is not a permanent solution against AI mimicry". And "it is always possible for techniques we use today to be overcome by a future algorithm". The project's page records published attacks on the method and an update made in response to one. Cloaking is friction with an unknown shelf life, applied at some cost to the picture.

Can a web host or CDN block AI bots?

Blocking at the server or network edge is stronger than asking, because the request is refused instead of politely declined. On 1 July 2025 Cloudflare announced that it was "changing the default to block AI crawlers unless they pay creators for their content". Many hosts and firewalls offer something similar.

The limit is identification. A bot that announces itself can be blocked. One that presents as an ordinary browser from an unlisted address is far harder to catch, and legitimate visitors can be caught by mistake. Login walls and low-resolution previews work on the same principle: reduce what an anonymous request can obtain.

What does the law add in the UK, EU and US?

Decided. The UK has no opt-out regime. The idea was the government's stated preference in a consultation that closed on 25 February 2025 and was dropped in the report published on 18 March 2026. The only UK text and data mining exception, section 29A of the 1988 Act, covers non-commercial research. In the EU, Article 4 of Directive 2019/790 allows mining of lawfully accessed works unless rights are expressly reserved "in an appropriate manner, such as machine-readable means". The Hanseatic Higher Regional Court held on 10 December 2025 in Kneschke v LAION that a reservation written in ordinary language in website terms was not machine-readable.

Pending. The German Federal Court of Justice heard the further appeal in Kneschke on 3 September 2026; PhotoWorkout reported that the court would announce its decision on 17 December 2026. The Court of Appeal heard Getty's appeal on secondary infringement on 11 and 12 November 2026, and no judgment could be confirmed on the date above. In the United States, we found no court ruling on fair use for training an image model when we checked. The cases tracker follows each of these.

So the same reservation has different weight in different places, and in each place part of the answer is still open.

A sensible layered set-up

  1. Publish web-size, watermarked copies. Keep masters and raw files offline.
  2. Embed creator, copyright notice, licensing URL and the Data Mining property in every published file, then confirm with the metadata viewer.
  3. Name the training crawlers in robots.txt, and block them at the network level if your host allows it.
  4. Put a short, human-readable licensing statement and a contact route on the site.
  5. Keep a dated record of what you published, with a file hash for each master, using the image hash and fingerprint tool.
  6. Check your settings twice a year. Crawler names change.

How the Robert Harper archive approaches it

Archive.London follows the layered approach described here. Only a small, watermarked selection of the Robert Harper Photograph Library is published on the site, its robots.txt asks AI training crawlers not to collect the photographs, and more than half of the archive, including the medium-format transparencies and other film, is still on film and has yet to be scanned. None of that is a guarantee about third-party copies of work that appeared in magazines. The archive is owned by 825 Ventures Ltd, the owner and rights holder, and licences for AI and data uses are negotiated directly with the single rights holder, subject to rights review. Developers who would prefer permission to scraping can start at the AI training data licensing page.

Frequently asked questions

Can I stop AI companies from scraping my photos?

Not with certainty once a usable copy is online. You can ask declared crawlers to stay away with robots.txt, embed a machine-readable rights reservation in each file, block known bots at the server or network level, and publish only small watermarked versions. Each reduces exposure or strengthens your position. None of them physically prevents a determined or undeclared scraper from copying a public image.

Do watermarks stop AI training?

No. A visible watermark deters casual reuse, tells a viewer who owns the picture and makes unlicensed use easier to spot. It does not stop a file being downloaded or used as training data, and it can be cropped or painted out. Its value is deterrence and identification, not prevention.

Does robots.txt protect images from AI crawlers?

Only against crawlers that choose to obey it. OpenAI, Anthropic, Google, Apple and Meta publish robots.txt tokens that they say control use for training, and naming those tokens is a reasonable step. The file is a request, not a barrier, it applies to your site only, and it does nothing about copies of your photographs on other people's websites.

Do Glaze and Nightshade work on photographs?

They were built at the University of Chicago to protect artists against style mimicry, by adding small pixel changes that mislead a model. The Glaze team says the changes are more visible on flat colours and smooth backgrounds, that the tool is not a permanent solution, and that future techniques may overcome it. Treat cloaking as experimental friction, not as protection you can rely on.

Is it illegal to train AI on my photos without permission?

It depends on where the copying happens and on questions courts have not finished answering. In the UK, commercial training is outside the only text and data mining exception. In the EU it turns on whether rights were validly reserved. In the US we found no court ruling on fair use for an image model when we checked. This is general information, not legal advice.

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