Adobe has built part of its artificial intelligence strategy on a very recognizable banner: protecting creators in a time of profound change. While other technology companies accumulated criticism for the origin of their data, the company presented itself as a responsible alternative. That position is now facing a lawsuit which focuses on the training of one of its models and the use of copyrighted works. The case is not an anomaly, but rather a reflection of a question that the industry has not yet been able to clearly answer.
The lawsuit was filed Tuesday in the U.S. Court for the Northern District of California and takes the form of a proposed class action. An author named Elizabeth Lyon accuses Adobe of using copyrighted books, including her own, to train the company’s AI models, with SlimLM at the center of the case, without permission. According to judicial documentation, these works would have been part of the training process of systems designed to respond to human instructions. Lyon claims to be acting on behalf of other rights holders who would find themselves in a similar situation.
The great debate about data that trains AI
To understand why this type of litigation is repeated with increasing frequency, it is worth taking a moment to look at how current artificial intelligence works. Beyond the visible applications, from chatbots to image generators, there are underlying models that act as the core of the system and learn from huge volumes of data. Generally speaking, more data can improve performance, although it is not the only factor. The problem appears when the key question arises about the origin of that information and the conditions under which it has been used.
The model indicated in the lawsuit is not Firefly, Adobe’s best-known creative system, but SlimLMa family of smaller language models designed for specific tasks. These models are designed to assist users with document-related functions, especially on mobile devices. It is not an AI aimed at large-scale creative generation, but rather a system that operates in the background. That difference is relevant because it shows that the debate over training data is not limited to the most visible applications.
According to the lawsuit, the conflict would not be in SlimLM as a final product, but in the data used during its training phase. Adobe has explained that these models were pre-trained with SlimPajama-627Ba open source data set published by Cerebras in June 2023. The court brief maintains that SlimPajama derives from RedPajama, another dataset widely used in the industry, and which in turn incorporates Books3, a massive collection of copyrighted books. That chain is the one that, according to the plaintiff, would have allowed the inclusion of works without authorization.
Until now, Adobe’s public narrative on artificial intelligence has been primarily articulated around Fireflya product clearly identified with respect for creators and the use of licensed content. The company has defended that these models were trained with licensed content, such as Adobe Stock, and public domain material, and has accompanied that message with compensation programs for Adobe Stock contributors. The demand, however, is not directed at that visible front, but, as we say, at SlimLM, a more discreet model, integrated into assistance tasks and without a direct commercial presence. This separation is key to understanding the real scope of the case.
The proceedings against Adobe are framed in a broader context of litigation in the United States related to the training of AI models. In recent years, authors and other rights holders have taken to court technology companies like OpenAIor Anthropicwith lawsuits alleging the use of protected works without authorization. Some of these processes are still open and others have ended in million-dollar agreements. This scenario explains why each new case is interpreted as another step in the legal delimitation of the use of data in artificial intelligence.
For now, the case is in an initial phase and leaves many unknowns open. The plaintiff requests a unspecified financial compensation and raises the action on behalf of other potentially affected parties, while Adobe did not respond to Reuters’ request for comment. It will be the judicial process that determines whether the lawsuit is successful, is filed or results in an agreement. Beyond its specific outcome, the litigation once again puts the focus on an issue that remains unresolved: how to balance the advancement of AI with the rights of those who create the content from which it learns.
Images | Rubaitul Azad | Adobe
GIPHY App Key not set. Please check settings