This is a course that, as the title indicates, will study how Large Language Models (“LLMs”) will impact the legal profession – and, truth be told, the law itself.

The roots of this course can be found in my eagerness to teach about LLMs; originally, this was simply a 2h session in the context of my Legal Data Analysis course,¹ then a full-on, 4h module I delivered at several institutions. But I thought that was not nearly enough to cover everything that had to be covered; hence this 24h version, which currently reaches to 3040 printed words.

This course was first taught at Sciences Po Paris in Spring 2025, to a cohort of M1 students who had to bear with what was a very rough draft. I am thankful to them, and to future students as well.

What Are LLMs ?

LLMs are advanced artificial intelligence systems trained on massive textual datasets to generate language outputs that closely mimic human speech and writing. The fundamental principle underlying their operation is next-token prediction – given a sequence of words, an LLM predicts the most probable subsequent word. While this sounds simple enough, this apparent straightforwardness hides great complexity, both in what is behind the scenes (it’s a tad more complicated than that, and we’ll see how) and in what it can achieve (a lot, as we’ll also see).

At their core, LLMs are substantially data-driven. They are trained on immense corpora, billions and billions of words taken from digitized books, articles, websites, Wikipedia pages, and other textual content sourced across the globe. Basically, if it has been written down and digitized, there is a good chance an LLM has seen it.² This scale of data allows them to grasp and replicate patterns of human communication at levels previously unimagined.

Now, these models are remarkably general-purpose. While you might initially use them for trivialities like summarizing an article, translating a restaurant menu, or writing a full course, their applications in professional contexts – especially law – are where things get interesting. LLMs can assist in drafting complex legal arguments, in going through enormous datasets for precedents, or in doing compliance checks. The key is that their abilities arise from the scale of their training data and the sophistication of their underlying architecture, not from explicitly programmed rules.

As such, they’re transformative, or even epoch-defining. Never before in human history has the production of coherent, contextually relevant text become so effortless, scalable, and affordable. This represents a genuine revolution in knowledge work, with immediate implications for the legal profession. For lawyers, whose entire trade revolves around crafting and interpreting language, this is not just an incremental improvement, it has the potential to be a seismic shift.

Yet, despite their impressive capabilities, LLMs have notable limits and idiosyncrasies. While we will spend an entire course on the limits of LLMs – but also on their potential, let’s chart out a few now, at various stages of the process:

Input: The quality of the output depends heavily on the quality of the input. This is the classic “garbage in, garbage out” principle. If the LLM is fed biased, outdated, or incorrect data, its output will likely reflect those flaws. And it’s not just about the data the model was trained on; the specific prompt you give it also matters. Ambiguous or poorly worded prompts will lead to ambiguous or unhelpful results.

Processing: Even with perfect input, the internal workings of an LLM are not foolproof. This is where we get into issues like “hallucinations” – where the model confidently generates false or nonsensical information. There is also the problem of transparency: these models are incredibly complex, and it is often difficult to understand why they generated a particular output. Whether this “black box” nature is a deal-killer will be something we will have to review at some point.

Output: Even if the processing is flawless, the output itself can raise issues. Who owns the copyright to text generated by an LLM ? Who is liable if an LLM provides incorrect legal advice ? These are complex legal and ethical questions that we’ll be grappling with throughout this course.

Before continuing, let’s set down a few definitions:

Definitions

Model: A mathematical framework trained to perform specific tasks by analysing patterns in data. In LLMs, the “model” refers to the neural network architecture trained to understand and generate human-like language.

Open-Source: Software or models whose code and architecture are freely available to the public. Examples include Meta’s Llama or DeepSeek R1, promoting customization and transparency.

Prompt: The input given to an LLM to guide its response (e.g., a question, statement, or task description). Minor changes in wording can drastically affect the generated response.

Token: The smallest unit of text (e.g., words, characters, or sub-word units) processed by an LLM. Inputs are broken down into tokens for analysis and generation.

Training Data: The dataset used to teach an LLM how to predict text sequences, drawn from diverse sources such as books, articles, and online content.

Inference: Generating a response based on a given prompt using a trained model.

Fine-Tuning: Adapting a pre-trained LLM to specialize in specific tasks or domains, such as legal applications. Usually faster and less resource-intensive than initial training.

CoT: “Chain-of-thought,” a method that prompts the LLM to show its intermediate reasoning steps, improving transparency and accuracy.

Temperature: Controls randomness of output; lower values yield predictable responses, higher values increase creativity and variability.

Parameter: A learned weight in a neural network that influences the model’s predictions by adjusting the strength of connections between neurons.

Why ? LLMs Are Everywhere (Law)

So, we have established what LLMs are – powerful text-processing engines. Now, let’s talk about why they matter to the legal profession. Why? Because law, at its heart, is about language – reading it, writing it, interpreting it, arguing about it. And LLMs are, fundamentally, language processing machines. They’re starting to appear everywhere, in almost every aspect of legal work, and that presence is only going to grow. The ubiquity of text in law means that, for better or for worse, impact will happen.

Consider legislation. LLMs can assist lawmakers in drafting bills, analysing existing statutes, and even summarizing public feedback on proposed legislation. Not only speeding things up, but also (potentially) improving the quality and clarity of legislation. A bit of this is already being tested out.³ In a world where “legislative inflation” has been decried for decades,⁴ LLMs have the potential to be a remedy that allow legislators to make sense of that bramble (of course, they can also be a poison that will multiply the amount of text – a true pharmakos). It somehow took a full year after ChatGPT burst onto the scene for the first (reported) AI-generated norm, but chances are good that countless parliamentary clerks and government lawyers have been honing their AI skills for the past few years. Indeed, one of the charges levied at the (second) Trump administration was that some parts of their Executive Orders were AI-generated.⁵

For legal advice, LLMs are already being used for legal research, case preparation, and even drafting arguments. Even the once-sacred domain of courtroom advocacy and litigation has seen infiltration by AI-generated texts, raising new and intriguing questions about authenticity and trustworthiness. Imagine being able to quickly find relevant precedents, analyse thousands of documents, and generate a first draft of a legal brief in a fraction of the time it would take a human. For businesses, especially in heavily regulated industries, LLMs can help with regulatory compliance.

Finally, in legal education and training, LLMs can generate practice questions, create case studies, and even provide simulated training environments for law students and lawyers. In fact, chances are good that a substantial portion of this very course is LLM-generated (but I won’t tell).

We are only scratching the surface. The point here isn’t just that it can be done; it’s that all of this is becoming increasingly practical and affordable. This is a major shift: text, good text even, is now cheaper than ever.

Now, before we get too carried away with the utopian vision of AI-powered law, let’s take a step back and look at some, shall we say, less successful examples. Because, as it turns out, LLMs aren’t perfect. A few well-known examples:

Hallucinations: Cases of lawyers or litigants that have mistakenly cited hallucinated cases has now become a rather common trope – the number of cases keeps climbing. Consider the striking case of a misinformation expert who relied heavily on GPT-4o to draft his declaration, inadvertently including fake citations – an incident described memorably by one court as shattering his credibility (Kohls v. Ellison).⁶ This raises a fundamental epistemological challenge for legal practice: How do we ensure truthfulness and diligence when AI-generated text is involved ? [Readers might be interested in my database of AI Hallucinations in Court.]

Prompt Injection: If LLMs are sensitive to the data they are being trained on, and if that data stems from the Internet … what prevents anyone from touting his or her pet legal theory on enough fora, websites, comment pages, self-published articles, etc., to become a “fact” as far as LLMs are concerned ? Nothing, and I predict that was are only years away of this happening, deliberately or not. Consider the kerfuffle, in the USA, about the Equal Rights Amendment, which the departing Biden administration said was good law – even though most experts say it is not. While the innocent explanation here is to see it as a symbolic gesture, that might still enter the training data of an LLM as “good law”.

Disturbing the equilibrium: many systems function in a particular type of equilibrium, which may or may not proceed from its apparent assumptions. In particular, many systems are geared as being designed to provide a certain good, but would collapse if every potential recipient of that good actually sought to obtain it. This is what we could term the “gym membership” model of the law: it works only to the extent a minority of potential cases make it through the court system.⁷ Likewise, public comments on legislation make sense only to the extent expert and/or specially interested parties pitch in. But what happens when the levels of friction that used to prevent lawsuits/public comments are removed ? Could courts, for instance, cope with the workload of every possible tort case ? If legal actions become cheap, we might have to reconsider how these models work.

The “gym membership” model of the law

Famously,⁸ gyms are said to be profitable businesses only because a lot of people subscribe at the beginning of the civil/academic year, but never come up after a while. (Regular gymgoers know this and adjust their schedule.) In other words, If every member actually exercised regularly, gyms would become overcrowded, raising costs, lowering quality, or necessitating expansion – quickly destroying their profitability.

This equilibrium also reflects social norms – gym membership often symbolizes aspiration rather than actual commitment. The mere possession of membership to certain gyms can confer status or identity, creating value even without consistent use. Similarly, the symbolic accessibility of courts and justice systems can be valued even by those who never use them; many legal rights are universal and symbolic, but rarely exercised.

Taken from another angle, Humans regularly overestimate their commitment and underestimate future behavioural barriers (optimism bias). Many believe they will use the gym consistently, but fail to do so. This psychological bias helps maintain the profitable equilibrium. In legal terms, people often overestimate their willingness or ability to engage in litigation, and the frictions inherent in these systems (i.e., the need to hire a lawyer in certain jurisdictions, the typically antiquated systems and ceremonials) can be seen as ways to sink many a potential case.

Another way to put it is that many systems are only workable if they come with certain levels of friction.⁹

Why ? LLMs Are Everywhere (Else)

But this is not all; LLMs are also everywhere else you are looking. Here as well, several examples will suffice.

Slop: AI’s digital junk food

Let’s start with something even more pervasive than hallucinations, more insidious: slop. What is slop ? It’s low-quality, AI-generated content flooding the internet. It’s not necessarily wrong, like a hallucination, but it’s empty. Think of it as the digital equivalent of junk food – mass-produced, unsatisfying, and ultimately bad for you.

This connects to a rather unsettling idea called the “Dead Internet Theory.” The (maybe conspiratorial) theory suggests that a significant portion of the internet is no longer driven by genuine human activity. Instead, it is dominated by bots, AI-generated content, and automated interactions. A digital ghost town, populated by echoes and simulations.

Why is this happening? Economics. It is incredibly cheap and easy to use LLMs to generate vast quantities of text and images. Even if only a tiny fraction of people click on an ad or share a piece of AI-generated content, it can be profitable. This creates a perverse economic equilibrium where the cost of generating poor-quality content is lower than the collective cost borne by users forced to sift through digital garbage.

Figure : Shrimp Jesus was a nice example of slop

Geopolitics

Now let’s shift gears to something even bigger: geopolitics. LLMs are not just changing the way we write emails or research legal cases; they’re potentially reshaping the global balance of power. There are massive societal stakes: there will be economic impacts on the job market, of course, and the automation of knowledge work will transform work. There are also very important ethical concerns, such as the risks of misuse, biases in decision-making, and loss of human agency.

In particular, unless you live under a rock, you may have heard that there is a huge geopolitical competition playing out right now, primarily between the US and China, over who will control the future of AI. Technological dominance, economic power, military advantage, and ideological influence are all involved there. Think of it as a new kind of arms race, but instead of nuclear weapons, the weapons are algorithms and data.

One key aspect of this competition is the control of chips. The advanced semiconductors needed to train and run large LLMs are a scarce resource, and whoever controls the supply chain has a significant advantage. This is why you see governments investing heavily in domestic chip manufacturing and imposing export controls. Another critical concept is the idea of “sovereign AI”. This refers to a nation’s ability to develop and control its own AI capabilities, independent of other countries. Nations want to be able to train their own models, on their own data, and according to their own values.

Truth and Epistemology

Which lead us to a more important point: often, those who train LLMs get to decide what the truth is. The “reality” presented by an LLM is not necessarily objective or neutral; it is shaped by the data the model was trained on, the choices made by its developers, and the biases embedded in the system. Of course, most observers have (rightly) focused on the rather selective memory of certain Chinese LLMs (poor DeepSeek r1 must probably have been asked millions of times what happened at Tiananmen Square in 1989), but few realise that Western-based LLMs have their own sets of taboos and bias: some topics are hard to broach in Western societies, and LLM providers (understandably) err towards caution.

This isn’t entirely new. Whoever controls the flow of information has always had power. Le savoir n’est pas fait pour comprendre, il est fait pour trancher.¹⁰ But LLMs amplify this power to an unprecedented degree. They can generate seemingly authoritative, personalized “truths” at scale, making it increasingly difficult to distinguish between fact and fiction, between objective reality and curated narrative. Among other things, this has profound implications for democracy, for the legal system, and for society as a whole.¹¹

This isn’t just about “fake news” or misinformation. It’s about the epistemological foundations of our society, how we know what we know, and who we trust to tell us the truth. And in a world increasingly mediated by AI, those questions become more urgent and more complex than ever before.

Leibniz, the Principle of Sufficient Reason, and AI

Gottfried Wilhelm Leibniz (1646-1716) was a German polymath – a philosopher, mathematician, scientist, and logician, among other things. He’s one of the key figures of the Enlightenment, and his ideas have had a profound influence on Western thought. He is also, in some ways, a key inspiration behind today’s AI models.

Leibniz’s Principle of Sufficient Reason, at its core, states that everything has a reason or cause. There is no such thing as a truly random or inexplicable event. For every fact, for every event, for every truth, there’s a sufficient reason that explains why it is the way it is, and not otherwise. In other words, everything can be made commensurable, since everything can be tied to other things.

On this basis, Leibniz believed in a fundamentally rational and orderly universe, governed by logical principles. He envisioned a kind of “universal calculus” – a logical system that could, in theory, explain everything. He thought that if we had enough knowledge and the right logical tools, we could deduce the reasons for everything that happens.

How does this relate to AI ? Modern AI, particularly the development of LLMs, can be seen as a kind of (unintentional) realization of Leibniz’s vision. LLMs, in a sense, are attempting to create a “universal calculus” through language. They are learning the underlying patterns and relationships between words, concepts, and ideas, based on massive amounts of data. In a sense, the LLM is providing a “sufficient reason” (in the form of statistical probabilities) for its output.

But here’s the crucial point. The “sufficient reason” provided by an LLM is statistical, not necessarily logical or factual. The LLM can generate text that is grammatically correct and statistically plausible, but that doesn’t necessarily correspond to truth in the real world. This is where we get back to the problems of hallucinations, biases, and the lack of grounding. The challenge, then, is to develop AI systems that are not just statistically proficient, but also logically sound and factually grounded. We need AI that can not only generate plausible text but also provide genuine reasons for its output – reasons that are rooted in logic, evidence, and a connection to the real world.

What to Expect

We will be investigating all of these issues in the coming course.

Introduction