Have I Been Trained

Have I Been Trained is the question-as-framing of the visual-artist consent-and-data-dignity register inside the 2022–2024 U.S. and European pushback on generative-AI training-data ingestion. The phrase, fixed by the tool of the same name launched in mid-September 2022 by the Spawning AI team of Jordan Meyer, Mat Dryhurst, and Holly Herndon, names the harm of unconsented training-data ingestion at the level of the individual artist's dignity rather than at the level of the bargaining-table demand: the question itself makes the harm legible — am I in there? — and the Do Not Train registry the tool surfaces instantiates the default the question presupposes, no, until I say so. The framing runs in parallel with, and is the visual-artist data-dignity sister of, the legal-pleading Consent, Credit, Compensation register that the Andersen v. Stability AI class-action complaint and Karla Ortiz's Senate Judiciary testimony anchor.

The tool and the registry

The Have I Been Trained search interface lets any user type a query, paste a URL, or upload an image, and search the LAION-5B image-text dataset — the open ~5.85-billion-image scrape that anchors Stable Diffusion and most other widely deployed open generative-image models — for matches against their own work or likeness. Matched images can then be added to a Do Not Train registry, which Spawning maintains as a portable opt-out list any downstream model-trainer can ingest before its next training run. The tool's first headline find, a patient discovering private medical photographs of herself inside LAION-5B within days of launch, fixed the dignity-harm argument behind the framing: that the question "have I been trained?" is not abstract, because the answer for many people is yes-without-knowledge-or-consent on artefacts ranging from copyrighted artwork to private medical records.

The Stability AI opt-out proof point — December 2022

The framing's working credibility as a campaign — and the registry's existence as a mechanism rather than a posture — was fixed by the December 2022 agreement under which Stability AI committed to honour the Do Not Train registry ahead of training the next iteration of Stable Diffusion. The VentureBeat write-up carried Stability AI's commitment alongside the working two-week opt-out window the partnership opened. By 7 March 2023 Spawning reported that the registry had collected opt-outs on roughly 78 million artworks ahead of the Stable Diffusion 3 training run — a figure independently carried by The Decoder and broken down as approximately 40,000 individual artworks opted out via Have I Been Trained itself plus tens of millions more via partner platforms (the ArtStation NoAI tag, Shutterstock's default opt-out for non-contributing content).

The artistic-practice register and the data-dignity argument

Spawning's working framing of the registry as "a creator consent layer for AI" anchors on an artistic-practice register that distinguishes the Have I Been Trained framing from the parallel U.S. legal-pleading register of Consent, Credit, Compensation. Holly Herndon and Mat Dryhurst — Berlin-based musicians whose prior practice included the Holly+ vocal-deepfake consent project — have framed the company's working posture as pro-AI-and-pro-consent: the position that the right response to the unconsented-ingestion question is not to halt model development but to insist that training is a relationship a creator can decline, accept, or shape. That stance puts the Have I Been Trained framing at a distance from both the SAG-AFTRA existential-threat register (which anchors on displacement of creative work) and the Concept Art Association–led Consent-Credit-Compensation legal register (which anchors on a copyright cause of action) — same wave, distinct argumentative shape.

The contested edge — opt-out as default

The framing's strongest published critique is structural rather than rhetorical. The 3 May 2023 TechCrunch seed-round write-up carried the working European-creator position that opt-out — even backed by a registry as well-built as Spawning's — fails the GDPR actively-given-consent standard and so cannot be the right legal baseline for an EU-applicable training-data regime. The European Guild for Artificial Intelligence Regulation (EGAIR), the organised European creators' coalition, has carried the structurally adjacent position that the underlying harm is the inability to deny consent at all and that opt-in by default — not opt-out by default — is the correct regulatory and ethical baseline. The two positions are working sisters of each other inside the same training-data-consent debate, with Have I Been Trained operating as the working artist-side instantiation of the opt-out posture and the EGAIR / EU AI Act push operating as the opt-in counterweight.

Carrying the framing into consensual training — Source.Plus and Public Diffusion

The framing's working continuation since 2024 has run through Spawning's own move from opt-out (defensive) to opt-in / consensual curation (constructive). The 11 June 2024 Source.Plus launch put roughly 40 million public-domain and Creative-Commons-zero images behind a curation tool that pays creators a tenth-of-a-penny per training download, and the parallel Public Diffusion / PD12M project trained an image generator only on public-domain and CC0 content as proof-of-concept that the data-dignity framing is buildable rather than only refusable. The argumentative arc those two projects close is the through-line of the Have I Been Trained framing as a whole: the question that opened the campaign in September 2022 — have I been trained? — is followed by the registry that lets a creator answer no, and then by the curation tool and the model that show what training looks like when the answer is yes-with-consent.