It is no surprise that although lawyers use the same natural languages as other speakers, they look at language in a very different way and use it according to (even more) different rules. It is also no surprise that natural languages are data and that most legal systems comprise their rules within natural languages (with all due respect to traffic lights and road signs). This means that rules, coming out of words, are derived from very special data sets, in the sense that although they use natural languages that any speaker could use (notwithstanding some technical jargon and the occasional Latinism) they allow for very different meanings depending on whether it is a lawyer or a non-lawyer analysing the data (i.e., reading the words). Moreover, within the legal community there is often disagreement on the meaning of the data comprised in said sets and thus one must refer to specific criteria to determine the prevailing meaning, such as “communis opinio doctorum” or the case law of supreme courts.
This all means that data sets to train AI for the purpose of performing legal tasks such as interpreting normative sentences within statutes or judicial decisions are very peculiar. On the one hand both legal language and legal methodology tends to favour repeatability as a way to foster trust on the legal system – this is something wonderful for the purpose of training Large Language Models (LLM) – on the other hand, however, legal data sets are fugitive pieces – to train an LLM you will need i) the data set comprising as much normative texts as possible, on the lines of the Central Registry of Births, Marriages and Deaths in Saramago’s All the Names; and ii) the data set comprising judicial decisions that interpret and balance rules. These are only the two essential data sets that you’ll need to have to train any LLM into doing something similar, even if only ancillary, to what a lawyer (in whatever capacity) does. You could have more. For instance a data set comprising as much legal doctrine commenting judicial decisions.
This goes to show that legal data sets are not only language dependent, but very much legal system dependent, as data from any other legal system, even if using the same natural language, will feed the LLM with different texts of which different norms can be obtained or even similar texts of which similar norms can be obtained although applied differently by courts.
Hyperlocal approach must then be the appropriate way, as a recent piece has highlighted regarding a new legal AI being developed in Germany. This is actually a boon for a small country like Portugal, where the cost of obtaining said data sets is much lower than in big countries. In some cases the work is half done: think about the work being done over the years by the Diário da República Eletrónico, who now owns the data set of almost all the normative texts in force in Portugal. The same can be said about the old “Bases Jurídico-Documentais”, now at Instituto de Gestão Financeira e Equipamentos da Justiça, that includes a wide array of Portuguese case-law. This is why Portugal already offers some interesting choices in the field of legal AI.
Can it draft a legal opinion that you can trust blindly and send to your client? Can it write a legal paper that offers coherent interpretations to contentious normative texts? Maybe not, depending on the complexity of the matter, but it’s getting there and it will improve rapidly. Not only law firms but also legal research centres must take advantage of this new tool in the system. This will help address the elephant in the room: at a given time where is the line that separates what AI can do for you in a legal system and what is still for humans to perform. Complex balancing operations stemming from normative conflicts not solvable through norms of conflicts seems to be the last frontier. Don’t take it from me. Take it from AI. Here’s what Perplexity has to answer to my very simple question:
Is there a connection between the Mendonca-Moreso legal method of balancing and AI?
There is a conceptual connection between the Mendonca-Moreso legal method of balancing and artificial intelligence (AI), especially regarding how both approach complex decision-making involving competing principles or interests.
The Mendonca-Moreso model frames balancing as an intellectual operation for establishing normative preferences in situations of irresolvable normative conflicts, such as those involving fundamental rights and constitutional principles. This method relies on the articulation of reasons for action and the justification of choices in concrete cases, often requiring the weighing of conflicting values and the assignment of relative importance to each.
AI systems designed for legal reasoning, particularly those used for decision support or automated legal analysis, often implement similar balancing mechanisms. These systems can be programmed to evaluate multiple legal principles, assign weights or priorities based on the context, and justify outcomes by referencing the underlying reasons-mirroring the structure of the Mendonca-Moreso balancing method. The process of quantifying interests, comparing them, and providing reasoned justifications is central to both advanced AI legal tools and the balancing method.
However, while AI can operationalize aspects of balancing (such as aggregating and comparing quantifiable factors), the Mendonca-Moreso approach emphasizes the irreplaceable role of normative judgment and the need for reasoned justification rooted in fairness and context, which remains a challenge for current AI systems to fully replicate.
In summary, the Mendonca-Moreso method of balancing provides a theoretical foundation that aligns well with the logic underlying AI-based legal decision-making tools, although the human element of normative justification is not yet fully automatable.
This also goes to show the importance of good legal data sets: almost all the references that AI could list to support its answer were authored by Lisbon Public Law researchers (with two notable exceptions).
Ah! And there was still time to ask ChatGPT to illustrate said relation between the Mendonca-Moreso method and AI: it’s very humbling at least.
