Sean Flynn and Andres Izquierdo

On February 5, 2020, the U.S. Copyright Office and the World Intellectual Property Organization (WIPO) co-sponsored a well-attended event on Copyright in the Age of Artificial Intelligence (AI). The full-day event took an in-depth look at the development and operation of artificial intelligence (“AI”). It paid substantially less attention to the full range of intellectual property issues raised by this new field.

The single-day event featured visual artists, audiovisual producers, music composers and executives, software developers, guilds of diverse artistic interests, people developing artificial intelligence, and (mostly perhaps) copyright lawyers. 

AI Today

The event was strongest in its rich depiction of how far AI technology has developed. It featured AI created music (still not like Mozart. according to M. Harrington);  Rembrandt –like paintings, poetry, self-driving cars (including (AV) 4.0), illegal poaching detectors, and accessibility enhancing devices (e.g.  Wheelie control motorized wheelchair).

IP and AI

The event was somewhat less strong in its coverage of the complex intellectual property (IP) legal questions that AI involves. There are three main IP questions that AI raises:

(1) does the training of machines with data from copyrighted works require authorization under the operation of copyright law;

(2) once trained, when might we consider the outputs of AI technology in the form of new works to be copyright violations where they compete in the same market as the originals;

(3) who (if anyone) owns IP in the products of AI technology.

Most of the discussion of IP in the meeting focused on question (3), and to a lesser extent (2). But the most interesting and consequential questions for the technology’s development lies in question (1).

Copyright in AI Products

WIPO director General Frances Gurry opened the session with thoughts on the question of who should be considered to own AI-produced works. While emphasizing that WIPO is not preparing for norm-setting on the issue, and that this was a forum for raising questions, not answers, Director Gurry embraced a bold proposal. Perhaps, he opined, there should be a wholly new system to give hybrid exclusive rights to AI-created works. He proposed consideration of a form of copyright that might last for 10 or 20 years for machine-created works.

Many other speakers contributed on this point, canvassing existing doctrines requiring human (not animal or machine) authors and providing standards for identifying when that requirement is adequately met.

Copyright and Competing AI Works

There was a general consensus in the room that the outputs of AI in the form of new works would rarely violate copyright since copyright does not protect a particular artistic style.

Training AI with Copyrighted Works

There was unfortunately much less discussion of the question of whether and when authorization must be sought for the use of copyrighted works to train AI.

Here it may be useful to distinguish between AI and machine learning, on the one hand, and text and data mining on the other. Text and data mining refers to a broad range of applying computational processes to copyrighted works to derive data. Machine learning and AI involve applying programing techniques to data (often derived from text and data mining) to enable machines to dynamically “learn” from the data inputted. (Sag 2019).

Many text and data mining processes – such as Internet search or making a query of the Google Books database – does not involve machine learning or AI. Text and data mining are often necessary to train machines to do AI. And this is where the copyright issues arise that are common between the two. To do text and data mining, including to train machines, one needs a corpus of materials to mine. Creating this corpus requires the reproduction and storage of those materials, which may include copyrighted works.

And thus the interesting question presents itself – whether, and when, does copyright demand that authorization be granted to use copyrighted works in text and data mining – including to train machines for AI. This topic was addressed recently in Europe with the Digital Single Market (DSM) Directive, and in the United States with the HathiTrust (2014) and Google Books (2015) cases. Several other countries (Japan, German, UK, France) have recently adopted copyright exceptions for text and data mining. WIPO has asked for comments on a set of questions related to this issue. But largely it was avoided at this event.

The representative of the Authors Guild made the most interesting comments on this issue. The Authors Guild is traditionally quite protectionist in its view of copyright, including as lead plaintiff on two cases that (unsuccessfully) challenged the incorporation of copyright works in the Google Books and Hathi Trust corpuses. The Guild’s lawyer, Mary Rasenberger, opined that, in general, the output of AI in the form of new works could not be considered infringing because “they are just mashups of works that they have been fed.” She also opined that, in general, the copying of works to train the AI would be a fair use, similar to copying works in order to reverse engineer them. “But what about ingesting works to create competing works in the same market,” she asked? Here she intimated that the copyright question should be a closer call, referring to trends in and fears about declining artist revenue in the digital environment.

More Meetings Ahead: May 11 and 12

WIPO representatives announced that the next meeting on this topic will be May 11 and 12 in Geneva. It would be of great benefit to focus that meeting more on the IP issues and less on a show and tell of all the great technology out there. Primary among the questions at that meeting could be ones concerning WIPO’s own role in this space now that it has created a new division on IP and AI. For example:

  • What actions may WIPO take that may help balance the proper role of the copyright system in promoting creativity, disseminating knowledge, and fostering technological development in relation to machine learning, artificial intelligence, and text and data mining?
  • Should WIPO help explain the proper interpretation of existing law on the fact/expression dichotomy as permitting running queries and otherwise applying processes to a lawfully produced corpus of copyrighted materials, and
  • Should WIPO help facilitate the development of international norms and guidance on permitting the cross-border uses of materials and tools lawfully created in one member country to another?