Generative AI: How to protect your business from copyright infringement

The sudden growth of Generative AI (Gen AI) has developed transformative tools that, when used in commercial business settings, have vastly improved efficiency and productivity levels. However, alongside the advantages are inevitable risks.

Companies must remain vigilant about the risks associated with integrating AI generated content into everyday operations. While these AI systems offer unprecedented efficiency and accessibility, they also enable the effortless creation of material that may infringe copyright, without the user’s knowledge.

As businesses embrace the benefits of Gen AI, it is crucial to understand the potential liabilities and implement the right strategies to mitigate them.

Are AI companies responsible for their generated outputs?

The majority of AI companies’ Terms of Service agreements appear to absolve themselves of any responsibility by purportedly asserting that the user is responsible for ensuring any output is used in a way that doesn’t infringe third party rights. Legally it’s not that simple.

Firstly, there is the issue relating to the training of the AI system. Most AI companies scrape the internet for data and materials to train their AI systems. While some AI companies have entered into licensing agreements to use data and materials (e.g. News Corp and Open AI, and Shutterstock and Apple), it is rare for AI companies to own or pay licences in relation to all the data and materials used. More common is that most, if not all, of the data and materials used to train the AI model are obtained and used without consent. This raises the issue of copyright infringement, i.e. the replication (even if temporary) of the original work by the AI model to train it.

AI companies relying on ‘fair use’ claim

In the USA, AI companies are seeking to rely on the defence of ‘fair use’, claiming that the use of the data and materials is ‘fair use’ so long as the work is being used for educational purposes, and in most instances, is not being copied for direct financial gain (i.e. the training of the AI model is not directly generating revenue). Opponents of this argument in the USA claim that while the training doesn’t result in direct commercial gain, the ensuing product (i.e. the trained AI system) is ultimately commercialised which results in a financial gain for the AI company. This gain occurs at the expense of the owners of the copyright, given that the protected works were used to train the AI system. On this basis, the opponent’s claim that the use isn’t ‘fair use’ and constitutes copyright infringement.

Lack of clarity in legal precedent

This lack of clarity is espoused in the mixed ruling of US District Court Judge William Alsup of San Francisco in relation to AI company Anthropic (and its chatbot Claude). Alsup J ruled that training AI systems using works in which copyright subsists is not an infringement of copyright, provided the source of the works is legitimate (i.e. not pirated copies). The AI system’s distilling from thousands of written works to be able to produce its own passages of text qualified as fair use under US copyright law because it was “quintessentially transformative”. In the words of Alsup J:

“Like any reader aspiring to be a writer, Anthropic’s (AI large language models) trained upon works not to race ahead and replicate or supplant them – but to turn a hard corner and create something different…”

However, “Anthropic had no entitlement to use pirated copies for its central library”.

‘Fair use’ defence not available in Australia copyright disputes

While the ‘fair use’ defence may assist companies established in the USA, this defence isn’t available in Australia where a much stricter defence of ‘fair dealing’ would need to be established. The limited exception permitted by ‘fair dealing’ includes research, study, criticism, parody and news. This narrower defence in Australia is likely to not extend to the utilisation of data or materials by AI companies to train their AI systems, and may result in an AI company based in Australia infringing copyright in works used to train its AI Systems. Until clarity is provided via legislation in Australia, or an action is commenced in the Federal Court of Australia for copyright infringement by an AI company for training its AI system on works in Australia, this question will remain unanswered.

This means, however,  that Australian businesses are likely to have a stronger foundation for successfully challenging the training of AI on their data and materials, provided the infringing conduct occurs in Australia.

The role of Terms of Service in AI copyright disputes

Turning then to the outputs of the AI systems, it’s evident that most AI companies have a Terms of Service agreement that attempts to remove their responsibility for any legal claims or copyright infringement issues that arise from the material generated by their AI system i.e. the outputs. These agreements place responsibility back onto the user, and the prompts used to generate the output. The risk with this process for businesses is that users may not identify an output as infringing copyright because they are unfamiliar with the original work or subject matter. For example,  an AI system may generate promotional materials which inadvertently reproduces a similar company’s marketing material (or image) due to how the AI system was trained or based on the input prompt used.

A more obvious example of where a user could identify potential copyright infringement is the groundbreaking Disney & NBCUniversal v Midjourney case, which concentrates on the similarity of the outputs, and alleges copyright infringement by Midjourney for enabling the outputs to be generated in such a fashion as they are substantially similar to the original works (being famous Disney characters e.g. Jasmine and Aladin). While this case is currently before the USA courts, the decision is likely to have repercussions in every jurisdiction around the world.

Why is AI generated copyright infringement important?

It is likely that whether copyright infringement can be established, will depend in each instance, on the control of the individual using the AI system over the output, prior to the individual or business disseminating that output to the public.

For instance, if the input used is “create a picture of a man and a woman on a flying carpet” and the AI system generates an output image of Aladin and Jasmine on a flying carpet in Agrabah, one may argue that the human providing the input has not been responsible for the images chosen, being the copying of the Disney characters in which copyright subsists. Rather it was the AI system (due to its training) that generated the image that potentially infringes copyright.

In contrast, if the input used is “create a picture of the Disney characters Jasmine and Aladin riding a flying carpet above the town of Agrabah”, one may argue that this input would directly lead to the creation of an image that potentially infringes the copyright of Disney.

In this situation, it would be apparent to most individuals that the unauthorised use of an image of Aladin and Jasmine on a flying carpet, would be likely to constitute copyright infringement, irrespective of the input provided. However, when the former example of the generation of promotional materials for a business is concerned, it becomes less obvious (unless using a famous or well-known image or text). This is where it will be crucial for businesses utilising AI systems to record the input prompts used for each output, in the event of copyright infringement allegations.

A business that can demonstrate how it created the image without reference to the original work, may be able to mount a case that the original work wasn’t copied, or at the very least is likely to be able to rely on the defence of innocent infringement.

Legal and commercial consequences of infringement

Possible actions that can be brought against a business for copyright infringement include:

• a notice to cease and desist
• an injunction requiring you to stop use of the infringing content, and
• damages or an account of profits that require you to pay the copyright holder all profits that arose from the infringing conduct.

In addition to the associated legal costs, there may be practical business costs for the removal and rebranding needed following an infringement action.

In recent months, copyright holders have been directing their resources towards acting against AI developers rather than users as seen in the groundbreaking Disney & NBCUniversal against Midjourney case. However, this does not mean you are safe to use AI tools without proper management. Smaller businesses are more likely to directly target infringers rather than developers due to the associated costs.

Furthermore, copyright doesn’t subsist in AI generated materials unless it can be demonstrated that there has been extensive input from a human during the creation process. Works in which copyright doesn’t subsist can be copied/replicated by anyone at any time. This means that if you utilise an AI system to create a new brand, teaching materials, policies, etc. it is likely your business will not own copyright in that work. Therefore, there is no way your business can prevent a competitor from freely copying your AI generated content or logo (unless you secure trade mark protection for it).

How to mitigate the risks of AI to your organisation, without discouraging use

1. Develop an AI policy that provides guiderails for the use of AI systems within the company.
   This will ensure the use of AI aligns with company values, expectations and policies (e.g. privacy policy, social media policy, HR policies and more).
2. Always have a human review AI generated material.
   This allows you to identify potential copyright issues and make changes to any outputs before use. Additionally, if every aspect of generated material has been edited or creatively built upon by a human it becomes more likely for it to attain copyright protections.
3. Build originality into your prompts to avoid the potential of reproducing existing works.
4. Keep records of what AI system was used, what prompts were inputted, and what edits were made to the output.
   This is especially important for any external works that will be published to the public, e.g. put on the website, sent to a customer, etc.
5. Review your contracts to ensure independent contractors assume responsibility for the creation of original works.
   This includes them warranting and indemnifying you against copyright infringement claims by third parties due to the use of the work they created for your business.

Generative AI: A legal perspective

Generative AI is an invaluable tool for efficiently improving company productivity and can be used in a variety of effective ways by business. However, any use of an AI system should be done responsibly with a human in the loop. This will minimise the real risk of copyright infringement, or the loss of valuable intellectual property from your business.

If you or your business is looking to use (or already using) AI systems, and you’d like assistance with developing AI policies, keeping track of your AI systems (including obtaining our excel spreadsheet to assist with this), please contact Mark Metzeling.

The information contained in this article is general in nature and cannot be relied on as legal advice nor does it create an engagement. Please contact one of our lawyers listed above for advice about your specific situation.