How generative AI tools impact IP rights of users

The recent release of GPT-4 is an example of how quickly AI development is taking place. But in such a fast-moving technology market, there has to be an appropriate balance between enabling media to be inputted into AI tools in support of innovation and enabling IP rightsholders in that media to exercise control over whether and how their works are used.

Benefits to be gained from the use of generative AI tools are at risk if the balance is weighted too heavily in favour of AI developers and users.

What is GPT-4?

GPT-4 is the latest version of ChatGPT, the generative AI tool that has shot to global prominence in recent months, released by AI lab OpenAI on 15 March 2023. According to the developers, GPT-4 is more reliable and able to handle nuanced queries more easily. It also scores better on exams – GPT-4 scored in the 90th percentile on the Uniform Bar Exam compared to ChatGPT, which only reached the 10th percentile.

However, perhaps the most significant upgrade is that GPT-4 is now multi-modal, meaning that it can accept image inputs as well as just text inputs. This multi-modal ability means it can generate detailed descriptions and answer questions based on the contents of an image – for example, providing recipe suggestions in response to photographs of ingredients.

With a range of new types of inputs comes a plethora of new IP rights that subsist in such inputs, of which AI developers and users alike need to be aware before taking advantage of them. However, opportunities are already arising where GPT-4 can be used for good.

Potential real-life uses of GPT-4

Be My Eyes, a Danish start-up specialising in connecting individuals who are blind or vision-impaired with volunteers for help with daily life tasks such as shopping or navigating unfamiliar places, has developed a tool called ‘Virtual Volunteer’, currently in beta testing and integrated into its existing app, that the developers claim can generate the same level of context and understanding as a human volunteer. Users retain the option to connect with a human volunteer in case GPT-4 cannot provide satisfactory assistance.

Now that GPT-4 can understand image prompts, it can be shown a photo of the contents of someone’s fridge, for example, and can respond with a wide range of answers, such as suggesting recipes incorporating the ingredients it sees, or suggesting that the user should stock up on a certain item.

Be My Eyes has already reported a person with visual impairment using its app to navigate the railway system, not only to get details about where they were located on a map but also to obtain point-by-point instructions on how to safely reach their destination. GPT-4, after hours of training, has also learned to identify the important parts of a webpage and can read these out to a vision-impaired person, as well as making links between different salient parts of that webpage and jumping between them. This represents progress on traditional screen reader software which reads through a webpage line by line, without discriminating between unimportant and important information.

This sort of development could be really valuable in the medtech and healthcare sectors. It could be used by healthcare providers to make the lives of people who are blind or visually impaired easier, allowing them a more fulfilling and independent life.

Other potential further uses include in driverless vehicles – GPT-4, or a future iteration of it, could be used for capturing and interpreting images of the outside world and using that data to control the car and feed prompts to the passengers in the vehicle.

Potential risks

Despite there being exciting potential uses, there are potential risks in using GPT-4 from an IP rights perspective. GPT-4 is a type of generative AI, which relies on training inputs, i.e. data the AI tool, and its algorithms, are trained on. With GPT-4 now being multi-modal, OpenAI has to manage copyright infringement risk not only in the context of text but also in relation to images.

OpenAI has amended GPT-4’s terms of use to specify that content provided by users is not used to develop or improve its services – in other words, to train GPT-4. However, similar generative AI tools which do use user inputs to train their AI systems will have the additional burden of ensuring that not only is the material they use to train their AI non-infringing, but also that inputs prompted and directed by users are non-infringing.

For example, it would be an act of copyright infringement for a user to substantially copy a literary work, such as by taking a description of a key character or setting in a book and feeding that into the AI system. Similar issues arise for images – if the user created the image themselves, for example, by taking a photograph, then they will usually own the copyright in it and if they themselves, or someone else with their consent, fed that into the AI tool, the image would not be infringing content. However, that would not be the case where the image was of a third party, unless they waived their rights in it, or if the image had been created by a third party and not the AI user.

Copyright owners will, therefore, face challenges around controlling use of their works where AI tools are trained from user input, such as content-sharing website DeviantArt.

Some AI tools are open-source and rightsholders can track back to find out where the data or images used by the AI developer themselves to train the tool were scraped from en masse, and whether their own copyright works were subject to that activity. However, beyond scraping, it is unlikely to be easy for rightsholders to establish if individual users have inputted their copyrighted works and then used the AI tool to generate something new based on those works.

Practically, technological protection measures can be put in place, such as preventing users from ‘right clicking’ to copy or download an image. However, this does not stop the copying of images in other ways, such as by printing, photographing, or screenshotting.

Consequences

If users of AI tools are planning to input more than just text descriptions, then they will need to manage the risk of infringing third party copyright, in respect of which both they and the AI developer could be liable.

The potential for copyright works to be used as training data by users may discourage rightsholders from sharing their works online if they know they cannot control where those works go and how they are used, unlike when such works are used by AI developers in formal training data since rightsholders may be able to trace this.

At the moment, GPT-4 is a paid-for service. This limits its availability. It seems likely that OpenAI has updated its terms of service after having assessed the risks of IP infringement arising from using user inputs to train its AI systems. These factors arguably reduce the risks presented by its predecessor ChatGPT. However, IP risks in the wider generative AI market in the context of other AI developers developing similar tools still remain.

Rightsholders will want to monitor whether developers take a similar approach and refuse to train their AI systems on user-inputted content, or whether they will instead place more liability on users of AI tools in their terms – leaving it up to the user to decide whether they want to upload text, images, or videos, but placing the burden on that user to assess whether they have the right to do this. Such an approach could spur an increase in copyright infringement claims being made against users, should rightsholders have an appetite to pursue such cases.

Co-written by Bella Phillips and Concetta Scrimshaw of Pinsent Masons.