Truth roulette: the case for implementation of duties to disclose and minimize hallucinations in legal large language models

Abstract

The legal profession currently faces an access to justice problem. Both the language of the law as well as legal representation are inaccessible. The development of artificial intelligence (AI), particularly Large Language Models (LLMs) can address this problem because they democratise access to information, including access to legal information, and are cost friendly. However, at the current stage of LLM development, LLMs are bound to hallucinate which leads them to intentionally or unintentionally produce misinformation. Hallucinations are particularly harmful in the legal context due to an inherent information asymmetry between the user and the provider of legal services. Therefore, the user of the LLM may not know what is and is not accurate. Hallucinations therefore undermine the ability of LLMs to promote access to justice. There currently are no obligations on the part of developers to implement measures that minimise hallucinations or even inform their users about their existence. Technical measures to minimise hallucinations exist, meaning that developers have the ability to minimise hallucinations. This dissertation seeks to look into the kind of duties that should be imposed on the developers of LLMs that can give legal information. It seeks to do so by analysing the ability of LLMs to increase access to justice relative to the vulnerability of its users, as well as the availability of technical methods to minimise hallucinations. This study finds that countries should impose duties on developers to reduce the harmful effects of hallucinations. The duties include, at minimum, a duty to disclose the existence of hallucinations, and a duty to minimise hallucinations based on capacity.