Watermarking models prevent malicious actors from obtaining powerful models.

Commercial actors may not have sufficient incentives to protect their models, and their cyberdefense measures can often be insufficient. In early March 2023, Meta demonstrated that this is not a theoretical concern, when their model known as LLaMa was leaked to the internet.22 As of the date of this publication, Meta has been unable to determine who leaked the model. This lab leak allowed anyone to copy the model and represented the first time that a major tech firm’s restricted-access large language model was released to the public.

Watermarking of AI models provides effective protection against stealing, illegitimate redistribution and unauthorized application, because this practice enables legal action against identifiable leakers. Many digital media are already protected by watermarking – for example through the embedding of company logos in images or videos. A similar process23 can be applied to advanced AI models, either by inserting information directly into the model parameters or by training it on specific trigger data.

We recommend that governments mandate watermarking for AI models, which will make it easier for AI developers to take action against illegitimate distribution.