Section 7.5 Copyright, Intellectual Property, and the Creative Economy

The source text data used for training is an issue. Typically, foundation models use self-supervised (i.e., unsupervised and unlabeled) pre-training from a good portion of the Internet. A proportion of that data will be intellectual property, perhaps even carrying copyright notices. Other parts may be private—names, addresses, etc.—and an LLM will have the ability to collate such information. Furthermore, a nefarious user, using the right sequence of prompts, may be able to get the LLM to collate information across disparate sources.

Beyond the ethical concerns of data collection, the creative labor of human artists is also under threat. The image generators DALL-E and Stable Diffusion can produce images in the style of well-known artists or video game illustrators. The images are not copies of images, but the style can be a copy of a style. This might be a problem. Illustrators may have an identity in part established by their style (and sometimes make a good living through their identity). LLMs seemingly can allow bad actors to steal a professional identity from others.

For example, there are now hundreds of thousands of images on the web that look as though they had been created by Greg Rutkowski (many produced by the LLM Stable Diffusion). Then the question arises: why would any company or creator pay Greg Rutkowski for an image, for use in a video game, when they can produce one for themselves essentially free?

Moreover, LLMs can give completely wrong or inappropriate responses. They will have learned enough grammar and vocabulary to ensure that what they generate will be entirely well-formed and well-written language. That may be good enough in one sense for fiction or poetry (although those literary forms may have additional desirable qualities which an LLM may or may not be able to provide). But non-fiction is another realm. The LLMs are working with probabilities.

In addition, LLMs can create fake content and deepfakes—not by accident, as it were, but because the user was trying to do exactly that. A deepfake is an image, a video, or a voice recording intended to simulate or portray an individual.

Turning to the legal dimension, as librarians license access to content from vendors, we need to ensure that contracts do not preclude our users from conducting text and data-mining research, algorithmically based research, and machine learning.

A foundational principle here is that copyright concerns the expression of ideas, not the ideas themselves. So, when Einstein wrote the theory of relativity, his actual words might have had some copyright protections but the theory of relativity itself did not. As a result, when NLP abstracts, or paraphrases, or summarizes, or text-data-mines documents or collections in its own words, it might be that there would be no copyright concerns. In contrast, extractive abstracts or summaries, or quotation of passages verbatim, etc., might be problematic.

Additionally, many copyright laws have exceptions for “fair use” which might include use for research, teaching, and non-commercial uses. When the authors of the resources license their work with Creative Commons licenses, the resources become Open Educational Resources and the teacher can reuse, transform, remix, and redistribute them freely. The only obligation, usually, is to quote the original author correctly. Therefore, it is important to identify when a resource is open or not.