Intellectual property and generative AI

As is often the case with new technologies, the risks of generative Artificial Intelligence (AI) are currently at the forefront of public discussions. The arguments are varied and complex, and explore the multitude of ways in which generative AI might, for example, upend job markets, saturate the internet with false information, create new security threats, accentuate existing societal biases, and more.

One risk that’s already leading to proposed changes in law is the thorny intersection between AI and intellectual property (IP).

Generative AI and IP 

In simple terms, IP refers to the intangible creations of humans, such as ideas, inventions, creative works, logos, brand names, and more. A related concept is Intellectual Property Rights (IPR), which are intended to protect creators and prevent others from using their creations without permission.

The rapid rise of generative AI has led to a whole range of new and complex IP challenges, and there are already several on-going legal cases against organisations that have released AI image generation, AI text generation and AI code generation tools, and more will undoubtedly follow.

These lawsuits have thus far focused mainly on the use of IPR-protected content in model training, but this is just the tip of the iceberg. There are many other IP risks relevant to generative AI. Here, we explain five of these risks in more detail.

Risk 1: Training data risks

This is one of the widest reported risks: the use of IP-protected data in training AI models. This risk is not unique to generative AI, but has become increasingly prominent as creators see their works replicated in the outputs of generative AI models, and feel their jobs threatened by these products.

Companies that develop generative AI models are facing backlash for using copyrighted material during model training without the permissions of the rights holders, and it is unclear whether existing exceptions, such as fair use, will cover these cases. Different countries and international bodies are actively working on relevant guidelines and legislation, and there is currently a spectrum of approaches, from tight restrictions and fully transparent disclosure of training material, to almost completely unrestricted use of copyrighted works.

Risk 2: Model output risks

Building on training data risks, generative AI models can produce outputs that incorporate or are otherwise derivative of protected works, often without informing model users. These users—which can be anyone from company employees to freelance workers—then face potential repercussions if these outputs are used in commercial work. It is currently unclear precisely how ‘derived’ a model output needs to be to present a risk to the user, and this will only likely be clarified through litigation or legislation. There are also outstanding questions about the IPRs that can be assigned to the outputs of generative AI, since they aren’t created entirely by humans.

Risk 3: Software licencing risks

Similar to model output risks, there are also software licencing risks that arise when developers use AI to generate useable code. If a generative AI model has been trained on code with attached licenses, and this model is then used to generate similar code that is incorporated in new software, then this software may be required to accurately reflect the licence(s) attached to the code in the training data. In extreme cases, for example, copyleft licenses may require commercial software to also be made open source.

Risk 4: Data leakage and trade secret risks

Depending on the specific product and model, input data can be used to train generative AI models on an on-going basis, and once trained, there are risks that this data is reproduced elsewhere for other users. This poses clear challenges in protecting sensitive and confidential data when using generative AI products and models, and there have already been reported instances of organisations banning the use of ChatGPT after employees used confidential company data within generative AI prompts, and this data later being found to have leaked beyond the organisation. 

Risk 5: Inventorship risks

In some fields—for instance drug design—AI has already been used to help create new inventions. However, most IP offices require a human inventor to be listed on a patent, and consequently, it’s currently unclear how—or even if—AI can be represented as having contributed to a patented invention. This may seem innocuous, but if a patent is reviewed by a court and found to not be principally the product of the named inventor, this can result in the patent being invalidated.

Conclusion

These are just some of the IP risks that exist within generative AI, and this list will undoubtedly grow as the technology continues to evolve. New regulations, new use cases, and new advances all bring the possibility of further IP risks that must be navigated.

Crucially, most of these risks apply not only to developers, but also to end-users in some form, meaning even relatively simple ‘off the shelf’ generative AI solutions carry inherent risks for organisations. This is one of the reasons why Deloitte has a bespoke IP practice: these risks can be complex and difficult to navigate, and it can be challenging for organisations to ensure that they are sufficiently protected whilst also allowing space for rapid and effective innovation.

If you’d like to find out more about how Deloitte can help your organisation manage the IP risks around generative AI, please get in touch.