When legal chatbots confidently explain whether a bicycle counts as a vehicle or whether someone may drive a truck into a prohibited park zone, it can feel as if artificial intelligence has already mastered the craft of legal reasoning. Yet according to new research by Henrique Marcos of Maastricht University, published in AI and Society in the article Can large language models apply the law? This impression is misleading. The study argues that large language models may imitate legal analysis but cannot actually apply the law in the way humans do.
This finding arrives at a critical moment. Courts, law firms and public bodies are increasingly experimenting with generative AI. Policymakers are considering whether AI systems might eventually assist judges or even make legally binding decisions. As debates intensify, Marcos’s work offers a timely framework for understanding what AI can and cannot do in legal contexts.
Why mimicking legal reasoning does not make AI a legal actor
At first glance, large language models appear to apply the law with ease. When users ask whether a rule prohibits bicycles, dogs or drones in a particular park, the system responds with neatly structured explanations that resemble legal advice. This often leads people to believe that AI performs legal reasoning in the same way a lawyer or judge does.
However, Marcos’s research argues that this impression falls apart when examined more closely. The study distinguishes between two forms of law application. The first is called inferential application. This involves reasoning from a legal rule to a conclusion about a specific case. The second is a pragmatic application. This involves performing an authoritative action such as issuing a judgment or deciding rights and obligations.
At the inferential level, the study argues that AI systems lack semantic understanding. They manipulate patterns in text without understanding the actual meaning of legal concepts. This limitation is well known in fields focused on natural language processing and cognitive science. It draws on classic arguments such as John Searle’s Chinese Room thought experiment. The thought experiment suggests that a system may produce correct outputs in a language without understanding the meaning of its symbols.
Marcos applies this insight to legal reasoning. If an AI system says that a rule applies to a case, it does not understand the rule or the case. It matches patterns seen in the training data. Such pattern matching may produce useful suggestions, but it does not constitute genuine legal interpretation. The system does not grasp concepts such as obligation, permission, or rights. It processes symbols syntactically rather than semantically.
Law as a social practice that AI cannot join
A central argument of the study concerns the collective nature of applying legal rules. Drawing on the work of Ludwig Wittgenstein and subsequent developments in the philosophy of language, the study examines how rules acquire meaning through shared use within a linguistic community.
In this framework, applying a rule is not an individual mental activity. Instead, it is a practice that depends on how a community interprets, discusses, and challenges rules. This includes judges, lawyers, scholars, and citizens. Legal meaning is negotiated through what philosopher Robert Brandom calls the game of giving and asking for reasons. Participants justify their interpretations, challenge others and refine their understanding through dialogue.
Marcos argues that this collective dimension is fundamental to the application of pragmatic law. When a judge applies a legal rule, the action is recognised as authoritative because the judge is part of a community that shares standards for interpreting and applying rules. These standards are not created by any single person. They emerge from collective practice.
Artificial intelligence systems, however, do not belong to this community. They do not participate in social dialogue, challenge interpretations, or respond to criticism. Even if humans treat them as if they have certain capacities, AI systems are not agents capable of shaping shared legal standards. They do not participate in the social practices that shape the meaning of legal rules. Their outputs, therefore, cannot count as applications of the law in the pragmatic sense.
Marcos’s study emphasises that even perfect mimicry of legal language would not make AI part of the shared linguistic practice that grounds legal authority. The authority of legal interpretation does not arise solely from reasoning skills but from participation in a social institution.
Could future AI become part of the legal community
The study does not entirely rule out the possibility that future AI systems might gain some form of social standing. It notes that theories of intentionality, such as Daniel Dennett’s intentional stance, suggest that humans may treat AI as intentional agents if doing so is practically useful.
Likewise, social practices change. The study mentions ideas such as the dead internet theory, which suggests that AI may already be generating a large proportion of online content. If people begin to interact with AI systems as if they were human members of a community, perceptions of agency could evolve.
Yet the study remains cautious. Even if society begins to ascribe intentionality to AI, this would not automatically grant it the capacity to contribute to the collective practices that determine legal meaning. The standards for rule application rely on human moral judgment, accountability and social participation, none of which current AI systems possess.
The regulatory landscape and the future of AI in law
The study engages with ongoing policy developments such as the European Union AI Act, which aims to ensure transparency, safety and human oversight in AI systems. The Act classifies certain legal AI applications as high-risk and requires robust safeguards. It also imposes transparency requirements on general-purpose AI systems, including large language models.
Marcos suggests that such regulation could shape whether society treats AI as an agent capable of participating in legal practice. However, the study concludes that at present, the limitations of AI are too substantial for it to be considered a legal actor.
This regulatory context is critical for practitioners. As legal professionals increasingly rely on AI tools for research, drafting, or analysis, there is a temptation to imagine these systems as capable of independent judgment. The study’s findings caution against this. They suggest that even as AI becomes more advanced, its role must remain supportive rather than authoritative.
Conclusion
The research by Henrique Marcos offers a comprehensive and timely analysis of the capabilities and limitations of large language models in legal reasoning. The study concludes that while AI can mimic legal argumentation, it cannot apply the law. It lacks semantic understanding at the inferential level and is not part of the collective linguistic community required for pragmatic application.
These findings are essential for shaping responsible AI use in law. They remind readers that legal authority is rooted not in linguistic performance but in shared human practice. As AI continues to evolve, legal systems must ensure that the foundations of legal reasoning remain grounded in human understanding, accountability and social engagement.
Reference
Marcos, H. J. (2024). Can large language models apply the law. AI and Society, 40, 3605 to 3614. https://doi.org/10.1007/s00146-024-02105-9
