Adobe hit with proposed class-action, accused of misusing authors’ work in AI training

Adobe, a leading player in the tech industry, has fully embraced artificial intelligence (AI) in recent years, launching various AI-driven services, including its media-generation suite, Firefly. However, the company is now embroiled in controversy following a proposed class-action lawsuit that accuses it of using unauthorized books for training one of its AI models. The lawsuit, filed by Oregon author Elizabeth Lyon, alleges that Adobe incorporated pirated versions of multiple books, including some of her own, into the training dataset for its SlimLM program. Adobe describes SlimLM as a small language model designed to enhance document-related tasks on mobile devices. According to the company, this model was initially trained on a dataset known as SlimPajama-627B, which was released by Cerebras in June 2023 and is characterized as a deduplicated, multi-corpora, open-source dataset. Lyon contends that her works were included in the pretraining dataset utilized by Adobe. The lawsuit, first reported by Reuters, asserts that the SlimPajama dataset was derived from manipulating the RedPajama dataset, which also includes Books3. This latter dataset is significant, as it contains around 191,000 books that have been utilized in training generative AI systems and has been at the center of various legal disputes within the tech sector. Legal challenges surrounding these datasets are becoming increasingly frequent. For instance, in September, a lawsuit against Apple claimed that the company used copyrighted materials to develop its Apple Intelligence model without permission. Similarly, in October, Salesforce faced a lawsuit for allegedly using RedPajama in its AI training. The trend of such lawsuits highlights a growing concern in the tech industry regarding the use of copyrighted material without proper authorization. In a notable case earlier this year, AI firm Anthropic agreed to a settlement of $1.5 billion with authors who accused it of incorporating pirated works into the training of its chatbot, Claude. This case could signal a pivotal moment in the ongoing legal landscape concerning the use of copyrighted content in AI training data.