Adobe sued over alleged misuse of authors’ work in AI training

Adobe is facing a proposed class action lawsuit in the US over claims that it used copyrighted books without permission to train its AI models.

The suit was filed by author Elizabeth Lyon in California federal court on December 17, 2025.

Lyon alleges Adobe used pirated copies of her instructional books, along with others, to train its SlimLM small language models, which are used for document-related tasks on mobile devices.

She filed the lawsuit on behalf of other copyright owners whose works were allegedly used and is seeking unspecified monetary damages.

This marks the first case of its kind against Adobe, amid a series of similar lawsuits targeting tech firms such as OpenAI and Anthropic over alleged copyright violations in AI training.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Adobe’s training data lineage reveals broader dataset contamination issues

Adobe trained SlimLM on SlimPajama-627B. SlimPajama is an open-source text set said to be released by Cerebras (an AI hardware company). The complaint says SlimPajama came from copying and altering RedPajama (an open dataset that included Books3, a large books collection of roughly 191,000 titles that has been the subject of copyright disputes) 1.
That setup still exposes Adobe to copyright claims. The company has promoted its use of licensed, public-domain, and Adobe Stock content for safe commercial outputs 2.
Developers using “open-source” datasets can face legal risk when those sets include unauthorized copyrighted works. RedPajama has been cited in lawsuits against Apple and Salesforce 2.
Courts must decide how to assign liability along the data supply chain. One question is whether downstream users (companies that train models on third-party datasets) are responsible when the original compiler made the unauthorized copies 2.

Opportunity emerges for licensed dataset vendors and provenance tracking tools

Language-model builders will need fully licensed content to avoid copyright fights. That could lift providers like SyndiGate (a vendor that licenses multilingual publications, journals, books, and datasets with explicit rights for AI training) 3.
Investors can watch vendors building dataset provenance tools (the origin and chain of custody of data). They can also follow compliance tools that trace training data lineage and verify licensing 2.
Enterprise AI teams will look for vendors with clear data sourcing and author opt-out mechanisms (ways for writers as well as rights holders to exclude their works from training). Publishers are striking licenses with OpenAI and Google, though a standard price has not emerged 4.