Artists' lawsuit puts training data on the balance sheet

A class action against Stability AI, Midjourney and DeviantArt tests whether copying billions of web images to train a model needs a licence.
Stable Diffusion learned to draw from more than five billion images copied from the web without consent or payment, and on 13 January three artists took that fact to federal court. Sarah Andersen, Kelly McKernan and Karla Ortiz filed a class action in San Francisco against Stability AI, Midjourney and DeviantArt, arguing that the model holds compressed copies of their works and that its outputs are derivative works competing with them.
Generative AI has been financed on the assumption that public web data is free to train on. This case puts a price on that assumption until a court rules, and the exposure sits in the training set, invisible in revenue figures. It also reaches beyond the lab that built the model: DeviantArt is named because its DreamUp generator runs on Stable Diffusion.
The claims are broad. They cover copyright infringement, removal of copyright information under the DMCA, rights of publicity and unfair competition, and the artists seek damages and an injunction. The training set at issue is LAION-5B, an open collection of 5.85 billion image and caption pairs and the dataset behind the open image model released in August. Stability AI raised $101 million in October, and the same legal team sued GitHub, Microsoft and OpenAI over the Copilot coding tool in November. Images and code are both in court now.
Whoever wins decides who earns from data. If training counts as fair use, data stays close to free and value accrues to compute and distribution. If it does not, model makers will need licences, and owners of large, well-documented libraries of images, text, code and audio become suppliers with pricing power.
A clean data record is already worth money. Licensed, consented or self-generated data shortens diligence and opens enterprise customers, who will ask about indemnities before they sign, so founders need to know where their training data came from and keep the record. Building on an open model does not remove the risk, since a company that ships a model inherits questions about how it was trained. Products that give creators opt-out tools, attribution or a share of revenue gain a commercial argument as well as a legal one, and provenance tracking, dataset licensing and consent management are becoming businesses in their own right.
Courts can still shrink the case. The motions to dismiss come first and will show whether a judge accepts that a model contains copies of its training images; without that, the central claim weakens. Two market moves would settle the economics faster than any verdict: stock image libraries and publishers filing suits of their own or signing licensing deals with model makers, and model providers offering customers indemnities against copyright claims, which would move the legal risk to the suppliers with the deepest balance sheets.
Training data is turning from a free input into an asset with legal title and a price. Fanaura's thesis for AI-native investing favours teams that own or license their data and can prove it, because that proof will be worth more than a few points on a benchmark.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building data licensing, provenance or consent tools for generative AI? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.