My new open-access article was published in Springer Nature’s AI & Society journal, ‘Can large language models apply the law?’
In this piece, I argue that LLMs cannot apply the law.
Note that this is not a matter of whether LLMs should apply the law. Rather, I examine whether they can—I conclude that they can’t.
The paper distinguishes between two interpretations of the ‘can’ question. First, can LLMs apply the law as an ordinary individual would? And second, can they perform this function as a judge might? Using D’Almeida’s theory of law application, I divide this into inferential and pragmatic forms of law application.
My findings? LLMs fall short on both counts. They cannot perform inferential law application because they lack true semantic understanding—they engage only syntactically with legal text. Similarly, they cannot fulfil pragmatic law application, as applying the law is inherently a shared, public practice rooted in membership with the linguistic community. LLMs are not (yet) members of our linguistic community. That might change in the future.
I believe this research raises important questions about the future role of LLMs in legal contexts and invites a deeper discussion on the limitations of AI in law application.
Read the full article here: https://link.springer.com/article/10.1007/s00146-024-02105-9
Or download it below:
Henrique Marcos 