A registry lets artists flag training images, if trainers cooperate

The process wall

Spawning AI, founded by artists including Mathew Dryhurst and Holly Herndon, built Have I Been Trained to answer a question a maker could not otherwise check: has a public training dataset scraped my work. An archived version of Spawning's site describes the tool and its API as having helped artists opt out more than 1.4 billion images from public training datasets. The wall the project responds to is not enforcement but visibility: before a search tool like this existed, an artist had no way to look inside a dataset the size of LAION-5B, let alone register an objection to it.

What the documents show

Spawning's site describes two linked tools: Have I Been Trained, for searching whether specific images appear in a training dataset and registering an opt-out, and ai.txt, described as letting a site owner generate a file that selectively restricts or permits use of that site's content for commercial AI training. Spawning's own FAQ is explicit about the basis for this: there are, in its words, no guarantees that copyright will be sufficient to protect artists from AI training, so the company operates assuming it will not. The tools are framed as a consent layer built alongside copyright law, not a legal claim, aimed at what the FAQ calls the industrial-scale use of artist training data rather than individual experimentation.

Authorship and limitation

The mechanism is registration and voluntary compliance, not deletion. An opt-out recorded through Have I Been Trained only has effect where a model trainer has agreed to check the registry before or during training; it cannot reach into a model already trained on an image, since training is not a reversible process the way removing a file from a folder is. Nothing in the documents opened here states that an opt-out forces a non-participating company to comply, and the FAQ's own framing, that copyright cannot be relied on, is itself an acknowledgement that the tool's force comes from adoption, not statute.

What survives

What survives for a maker is a documented signal: a search record showing whether a work appeared in a public dataset, and a registered opt-out that cooperating trainers can check. Editorially, an artist should treat this as one input to a broader strategy, not a substitute for reading a specific model's own training documentation, since the opt-out's reach depends entirely on which companies chose to honour it.

• Search a specific image through Have I Been Trained and record what dataset result, if any, comes back.
• Generate a site's own ai.txt file and compare its terms against a model trainer's stated crawling policy.
• Ask a model provider directly whether it checks Spawning's opt-out registry before training, since the site itself does not name every trainer that does.

The registry gives an artist something to point to and something to check, but its guarantees end exactly where a trainer's willingness to look ends.