Artificial intelligence and mental capacity legislation: Opening Pandora's modem

People with impaired decision-making capacity enjoy the same rights to access technology as people with full capacity. Our paper looks at realising this right in the specific contexts of artificial intelligence (AI) and mental capacity legislation. Ireland's Assisted Decision-Making (Capacity) Act, 2015 commenced in April 2023 and refers to 'assistive technology' within its 'communication' criterion for capacity. We explore the potential benefits and risks of AI in assisting communication under this legislation and seek to identify principles or lessons which might be applicable in other jurisdictions. We focus especially on Ireland's provisions for advance healthcare directives because previous research demonstrates that common barriers to advance care planning include (i) lack of knowledge and skills, (ii) fear of starting conversations about advance care planning, and (iii) lack of time. We hypothesise that these barriers might be overcome, at least in part, by using generative AI which is already freely available worldwide. Bodies such as the United Nations have produced guidance about ethical use of AI and these guide our analysis. One of the ethical risks in the current context is that AI would reach beyond communication and start to influence the content of decisions, especially among people with impaired decision-making capacity. For example, when we asked one AI model to 'Make me an advance healthcare directive', its initial response did not explicitly suggest content for the directive, but it did suggest topics that might be included, which could be seen as setting an agenda. One possibility for circumventing this and other shortcomings, such as concerns around accuracy of information, is to look to foundational models of AI. With their capabilities to be trained and fine-tuned to downstream tasks, purpose-designed AI models could be adapted to provide education about capacity legislation, facilitate patient and staff interaction, and allow interactive updates by healthcare professionals. These measures could optimise the benefits of AI and minimise risks. Similar efforts have been made to use AI more responsibly in healthcare by training large language models to answer healthcare questions more safely and accurately. We highlight the need for open discussion about optimising the potential of AI while minimising risks in this population.