Generative artificial intelligence

By the Copyright Advisory Committee, with the help of dan brown, Professor, Cheriton School of Computer Science

Generative artificial intelligence (GenAI) and related technologies are subject to existing laws and regulations in Canada, such as intellectual property, copyright and privacy laws, among others. The legal status (e.g., copyright legislation and case law) of GenAI services is currently unsettled in Canada. This guide is designed to help you make an informed decision about using GenAI-created content. For further teaching and learning considerations, see the AVPA FAQ on ChatGPT and GenAI.

Concerns in brief

Source material used to train the system must be legally copied. GenAI services rarely share information about their source material, so its legality is often difficult to verify and
Impossible to attribute accurately.
Copyright ownership of GenAI content is unclear in Canada. Because AI-generated content may not be protected by copyright, creators may not benefit financially or professionally. Moreover, the rights of creators whose copyright material was used to train AI systems, and generate content, remains unsettled.
There has not been any AI-specific legislation or case law in Canada. The lack of copyright legislation would not preclude GenAI services from making a fair dealing argument or using other Copyright Act exceptions, though depending on the use case it may be a challenging argument to make.¹ Several copyright-infringement lawsuits have been filed in the United States against GenAI services and new lawsuits and threats of litigation continue to emerge on a regular basis.

What follows will discuss these four points in more detail. We recommend that you Exercise Caution when using AI services.

Legality of source material hard to verify

To use content in your teaching, you must have a legal copy. For example, even though fair dealing would allow you to upload one chapter of a book to LEARN, you should not use one chapter from a pirated copy of the book you found on LibGen. For more information on how to tell if materials online are legal copies, see FAQ 17. In the case of GenAI services, the content used to train the model (often called a corpus, the basis on which it produces output), is either unknown (so the content source cannot be verified) or known to be infringing. For example:

Stable Diffusion, an image generation service, is built off a wide variety of images scraped from the web, including those licensed by website owners from Shutterstock, Getty Images, and other commercial stock photo websites.² Stable Diffusion has been the subject of lawsuits alleging copyright infringement, which are mentioned below in more detail.³
ChatGPT’s current corpus is unknown.^4,5 Previous versions were claimed to have been scraped from openly licensed sources online. Other large language model training data sets have included claims that their data sources do not infringe US copyright law.⁶ OpenAI’s FAQ⁷ indicates that the current model was “developed using three primary sources of information: (1) information that is publicly available on the internet, (2) information that we license from third parties, and (3) information that our users or our human trainers provide.” No further information is provided, so we have no way of knowing if the publicly available information from the internet has been cleared.

Attribution of source material near impossible

GenAI services rely on the content in their training models to generate new content based on user prompts. The models cannot produce a list of the specific source material used to generate any given output, and so it is impossible for a user of a GenAI service to attribute the source material of a creation. This means that if a particular output relied on a certain source in a substantial way (and that source was still protected by copyright), not citing that source would infringe the rights of its creator. Given the way AI services provide outputs to users, there would be no way for a user to know if the output generated based on their prompt was a substantial use of another creator’s work. Furthermore, certain prompts can be almost guaranteed to infringe; for example, the prompt: “A cartoon of a family, in the style of Alison Bechdel” is more likely to use substantial portions of Alison Bechdel’s work, and therefore more likely to be infringing.

Copyright ownership of GenAI content is unclear in Canada

In Canada, the question of copyright protection and ownership of AI-created content is not clear. Neither legislation nor case law provides clarity. In contrast, AI-created content is not granted copyright protection in the United States. That said, the US Copyright Office is engaged in an initiative to examine how copyright should interact with generative AI, so its stance on copyright may change.⁸ Because AI-generated content may not be protected by copyright, creators may not be able to benefit financially or professionally from it. Works that are not protected by copyright may be used without permission, payment, and attribution. Moreover, the rights of creators whose copyright material was used to train AI systems, and generate content, remains unsettled.

Legislation and/or case law pending?

Given how new publicly available GenAI systems are, there has not been legislative change specific to AI.⁹ There have also been no legal challenges in Canada. The lack of copyright legislation would not preclude AI services from making a fair dealing argument or using other Copyright Act exceptions, though depending on the use case it may be a challenging argument to make.¹

There are a number of copyright-related legal challenges to GenAI services ongoing in the US: Getty Images v Stability AI¹⁰ (reuse of stock images), GitHub Users v. GitHub, Microsoft, and OpenAI¹¹ (reuse of open-source code without attribution), and Anderson, McKernan, and Ortiz v. Stability AI, Midjourney, and DeviantArt¹² (reuse of images scraped from the web). A more in-depth overview of the GenAI legal landscape in the United States is available through the Congressional Research Service. The outcome of the above cases or any legislative change in the US would not have a binding legal effect in Canada but they serve as a barometer for possible developments in Canada.

Exercise caution

We understand that use of GenAI services is widespread, not only at Waterloo, but across the higher education sector and in the general public. In certain areas, particularly in classes where you are teaching about AI, avoiding its use would be a detriment to your students. The following are recommendations for principled, risk-informed use of AI services.

Ensure the use of GenAI content directly pertains to the learning objectives of your course/module. A direct connection to an educational purpose strengthens any argument that you might make under the fair dealing exception of the Copyright Act.
Read the terms and conditions of the service and take note of your rights and responsibilities under the terms. Pay attention to requirements for attribution and responsibilities for use of the content. For example:
Keep records. Make notes of the prompt you used and from where you downloaded the content. This will enable you to remove or alter content should you need to.
Attribute properly (i.e., as it is not possible to attribute the sources accurately, at least attribute according to the style recommended by your discipline). The Library has a guide on AI-generated content and citation.
Do not use GenAI content for decorative purposes (e.g., to make slides more aesthetically pleasing).
Avoid using prompts that rely on a single source or a style that isn’t in the public domain (e.g., “in the style of Takashi Murakami”).
Do not upload/input significant portions of work you do not own to AI services, unless you have permission to do so. For example:
1. do not upload an image you don’t own to Midjourney’s image-based prompts service
2. do not input the entire text of an article you don’t own to ChatGPT to ask for a summary
3. do not input the entire text of your student’s assignment into an AI service to evaluate the quality of the content/whether it was written by an AI service¹³

GenAI services are rapidly changing as is the context (moral, ethical, and legal) surrounding their use. As more information becomes available about these services and/or as the legal status of them changes, this document will be updated.

1.Brown, Byl & Grossman. (2021). Are machine learning corpora “fair dealing” under Canadian law?
2. Baio, A. (2022). Exploring 12 million of the 2.3 billion images used to train stable diffusion’s image generator. 
3. Butterick, M. (2023). Stable Diffusion litigation. 
4. Cooper, K. (2021). OpenAI GPT-3: Everything you need to know. 
5. Boykis, V. (2023). Everything I understand about chatgpt. 
6. Gao, L. et al. (2020). The Pile: an 800GB Dataset of Diverse Text for Language Modeling. 
7. Schade, M. (2023). How ChatGPT and our language models are developed. 
8. U.S. Copyright Office. (2023). Copyright and Artificial Intelligence. 
9. In July 2021, ISED released a report “A consultation on a modern copyright framework for artificial intelligence and the internet of things” a signal that the government may be considering legislative change. 
10. Getty Images (US), Inc. v. Stability AI, Inc., 1:23-cv-00135, (D. Del.). 
11. DOE 1 v. GitHub, Inc., 4:22-cv-06823, (N.D. Cal.). 
12. Andersen v. Stability AI Ltd., 3:23-cv-00201, (N.D. Cal.). Back to note.
13. Turnitin’s GenAI detection tool has been enabled. More information about using this service is available on the Academic Integrity page about Turnitin - see "Artificial Intelligence Writing Detection".