The US looks at AI and copyright

The US Copyright Office has just published a notice of an ambitious inquiry into AI and copyright, covering key policy options, technical and economic considerations and practical measures.

The inquiry seeks views on fundamental and thorny issues, such as whether there should be copyright in AI outputs, what to do about training on copyright works (exceptions, compulsory licences, consent, opt-outs, remuneration, etc,) and how to apportion liability for infringement between developer and user. Reaching consensus on such issues will be very challenging.

Consensus and change may be more likely for some of the specific measures being explored, such as technical tools that could facilitate opt-outs, record keeping, labelling AI outputs and whether to prevent outputs in the style of specific artists. Many of these measures are already being considered by players on both sides of the broader debate. And, in the case of record keeping, AI developers are already facing potential obligations under the proposed EU AI Act (see my post on the draft here).

Even if no legislative change results from the inquiry, hopefully it will collect valuable information through the many technical questions. These cover, e.g., the storage and representation of “inferences” within models, whether an AI model can viably unlearn inferences based on specific works, whether the use of specific training examples can be identified and whether it is feasible to identify the contribution of specific training examples to an output.

Finally, there are questions about economic impact. This has been a neglected area in the debate and the last independent study I am aware of is from 2015: Nesta’s study “Creativity vs Robots”.

The notice is published here.