Spawning built an opt-out registry, not an opt-out law

The system

Spawning is the company founded by musician-technologists Holly Herndon and Mat Dryhurst that built the Do-Not-Train registry and a related opt-out mechanism for AI training data. Spawning's own site currently displays a maintenance notice rather than full documentation, but that notice itself states its purpose plainly, directing 'AI trainers looking to respect the Do-Not-Train registry' to contact the company, confirming Spawning still operates the registry by that name. Spawning also maintains open-source tooling on GitHub under the organization Spawning-Inc, including a library called datadiligence, described in its own repository as software to 'respect generative AI opt-outs in your ML training pipeline.'

What the documents establish

The datadiligence repository's own README explains the underlying problem: training datasets are often collected without content owners' consent, and in the absence of one shared opt-out standard, different platforms have adopted their own methods for stating consent. The library documents that it currently checks three distinct opt-out methods: the Spawning API itself; DeviantArt's X-Robots-Tag HTTP headers, which the repository links to DeviantArt's own announcement that 'All Deviations Are Opted-Out of AI Datasets'; and C2PA and Content Authenticity Initiative metadata. This establishes Spawning's protocol as one of several opt-out signaling methods an ML pipeline can check, not a single universal standard, and DeviantArt as a concretely documented adopter of a Spawning-checked method. Spawning also maintains a verified organization on Hugging Face publishing public-domain datasets, indicating an active presence in the same ecosystem the opt-out tooling targets.

Craft and rights

For an artist or rights holder, the practical significance of Do-Not-Train is that it is opt-out and voluntary, not a legal requirement: it works only where a model trainer chooses to check it, using tooling like datadiligence, before scraping or training. Registering a work with Spawning does not itself prevent training on that work by a party that does not check the registry; it creates a documented signal of non-consent that a cooperating trainer can act on. That distinction matters for anyone relying on registration as protection: the registry's practical force depends entirely on which platforms and pipelines choose to query it, a fact Spawning's own materials do not overstate.

Outcomes and open questions

Because Spawning's main site was not fully reachable at the time of writing, this entry could not independently confirm from Spawning's own current materials the full current list of companies checking its opt-out signal beyond the DeviantArt example documented in its GitHub repository; that list should be rechecked against Spawning's site once it is available.

• Does the platform hosting a given work actually transmit an opt-out signal a trainer's pipeline would check?
• Is a company said to 'respect' Do-Not-Train confirmed by that company's or Spawning's own current documentation?
• Does opting out change how already-trained models behave, or only what future training pipelines that check the registry may collect?

An opt-out registry is only as strong as the number of trainers who choose to look at it, which is a market and norms question as much as a technical one.