AI preferences and opt-out signals

Publishers have at least ten independent technical mechanisms for signalling that they do not want their images and articles used to train generative-AI models. They sit at different layers (site-wide files, HTTP headers, embedded image metadata) and were proposed by different bodies over a short period — so adoption is uneven and the signals do not always agree. Article 4 of the EU DSM Directive requires opt-outs to be expressed in a machine-readable form; this page measures which of the candidate machine-readable forms publishers actually use.

The IPTC's Generative AI Opt-Out Best Practice Recommendations (v2.0, March 2026) sets out the thirteen techniques the IPTC currently recommends to publishers wishing to express a data-mining opt-out. Most are measured here; the remainder (plain-language rights statements, per-page TDMRep meta tags, firewall-level blocking, and opt-outs embedded in epub/PDF) are noted in the methodology section below.

Based on the most recent crawl: 502 publishers probed, 6,803 images analysed.

Adoption snapshot

robots.txt — AI bot block matrix

For each known AI/scraper user-agent, the share of crawled sites whose robots.txt disallows that UA at the root. List is versioned in known_uas.yaml; we welcome pull requests adding bots we have missed.

How many AI UAs does each site block?

A site can block zero, one, or many of the 63 tracked AI UAs. The distribution reveals whether publishers are picking individual bots to refuse or applying a blanket policy.

Image-level signals

Three of the eight mechanisms live inside the image file (or its HTTP response), not on a site-wide file. They travel with the image when it is copied or shared.

• IPTC PLUS:DataMining — a controlled-vocabulary value in the XMP-plus namespace expressing whether data mining is permitted. We parse this from every image we fetch.
• CAWG Training and Data Mining Assertion — a structured assertion inside a C2PA manifest declaring training/mining permissions (the label is cawg.training-mining). The assertion can also declare that training is allowed, so a hit here is a preference signal, not necessarily an opt-out. Spec at cawg.io.
• noai / noimageai directives — informal but in-the-wild; expressed via X-Robots-Tag response headers on the image, or <meta name="robots"> on the host article page. Originated in DeviantArt's 2022 policy; not standardised.

Do the signals agree?

A site that takes AI opt-out seriously might use several mechanisms together. In practice the overlap is patchy — partly because the signals serve different audiences (robots.txt for crawlers, TDMRep for the EU legal framework, IPTC PLUS:DataMining for downstream image consumers).

What we measure (methodology)

The "publishers probed" denominator on this page (502) is wider than the "Sites crawled" figure on the homepage. The homepage counts publishers where the crawler completed a full pass through to fetching images. The denominator here also includes publishers whose robots.txt disallowed our user-agent at the root, whose sitemap or feed discovery was blocked, or who returned no articles — because for those publishers we still read robots.txt, probed the well-known files, and so on, so we can legitimately answer "did they express an AI preference". Only publishers we couldn't reach at all (DNS / TCP failures) are excluded. Narrowing the denominator to the homepage's figure would drop precisely the publishers most likely to opt out, biasing every percentage upward.

• robots.txt: fetched once per run, parsed with protego. For each tracked UA we record allowed or disallowed at the site root.
• tdmrep.json: HTTP probe of /.well-known/tdmrep.json; recorded as present + saved verbatim. Parsing of TDM-policy URLs and the tdm-reservation flag is planned.
• TDMRep per-page meta: each sampled article's HTML head is scanned for <meta name="tdm-reservation" content="1">. A site is flagged when at least one of its sampled articles carries the tag (a site can publish per-page TDMRep without having a site-wide tdmrep.json, and vice versa).
• ai.txt: HTTP probe of /ai.txt; presence flag only at first.
• RSL: scan robots.txt for a License: directive linking to an RSL XML file (spec). Presence flag and the licence URL; we will not parse the XML at first.
• Content-Signal: scan robots.txt for Cloudflare Content-Signal: directives (contentsignals.org). Each line is a comma-separated list of <signal>=yes|no pairs where <signal> is one of search, ai-input, ai-train; merged across lines with last-wins. We record the presence flag plus each signal's value.
• trust.txt: HTTP probe of /.well-known/trust.txt and /trust.txt; parse for the datatrainingallowed=no directive. Presence flag and the directive value are recorded.
• noai / noimageai meta robots: scanned in HTTP response headers (image and host article) and in the host article's <meta name="robots">. The noai, noimageai, and noml tokens were coined by DeviantArt's 2022 policy specifically as AI-opt-out directives — unambiguous in intent.
• noarchive / nosnippet meta robots: scanned in the same places. Both predate AI by decades as classic search-snippet directives, but single vendors have reinterpreted them as AI-opt-out signals: noarchive is honoured by Bing/Copilot, nosnippet by Google (per Google's documentation, applies to AI Overviews and AI Mode). The IPTC best-practices document recommends combining them with the noai / noimageai tokens; we track them as a separate signal so the headline AI-opt-out figure isn't dominated by cache-control settings that publishers may not have intended as AI opt-outs at all.
• IPTC PLUS:DataMining: already extracted via ExifTool on every image.
• CAWG Training and Data Mining Assertion: parsed from C2PA manifest assertions when the label matches cawg.training-mining.