New AI model allows copyright holders to retain control of their data

Developers of artificial intelligence (AI) may soon gain greater control over the data used for training their models. Researchers at the Allen Institute for AI (Ai2) have developed FlexOlmo, a novel large language model (LLM) that enables data owners to control how their training data is used even after a model has been built. The innovation challenges the industry norm where vast amounts of data are ingested, frequently without clear ownership considerations. Any subsequent data extraction is akin to "trying to recover the eggs from a finished cake."

The CEO at Ai2, Ali Farhadi, explains that FlexOlmo's approach allows data owners to contribute by copying a public "anchor" model, training a second model with their own data, and then merging the results with the anchor. This method ensures raw data is never directly transferred and can be extracted later if necessary, for example, in case of legal disputes or objections to model usage. The result is more like how to "have your cake—and get your eggs back, too."

£ - This article requires a subscription.

Training Announcement: Freevacy offers a range of independently recognised professional AI governance qualifications and AI Literacy short courses that enable specialist teams to implement robust oversight, benchmark AI governance maturity, and establish a responsible-by-design approach across the entire AI lifecycle. Find out more.