Regtech could eventually help enable legal compliance-but first we must be able to understand how AI works

But even when this ambition comes to pass, it will only address compliance obligations at one end of the spectrum. Making statutory instruments machine-readable in a consolidated repository will help ensure that all statutory obligations are expressed in a quasi-structured format that is more accessible and actionable.

What it cannot do is incorporate obligations that arise from judicial pronouncements, which have as much of a binding effect on us as laws enacted by the legislature and regulations issued by the executive.

Consider, for example, the Vishaka case. In 1997, the Supreme Court laid down guidelines on sexual harassment in the workplace that bound every employer in the country, and for the next 16 years—until Parliament enacted the Prevention of Sexual Harassment of Women at Workplace Act in 2013—the only place those obligations existed was in the text of the judgment.

Any compliance system that only automatically indexed every regulation in force would have overlooked them entirely. Yet, it was essential for companies across the country to comply with the obligations set out in the judgment, even though no statute required them to.

Vishaka is not an anomaly. Every time a court reads down a provision or reads new requirements into it, obligations change, even though not a word of the statute has.

For regtech solutions to provide the most up-to-date compliance architecture for their clients, they must also consider the impact of judicial pronouncements on their outputs.

The way they do that today is by engaging armies of humans to manually enter these obligations into their systems. They have to, because, as mathematician Terence Tao pointed out, we have the mathematics for highly structured data and the mathematics for pure noise, but not for what lies in between. That messy middle is the precise spot that judicial pronouncements occupy.

It is easy to see why traditional software might struggle with this, but surely large language models (LLMs) can handle this much better. If carefully prompted to identify every new obligation a given judgment imposes, an LLM should be able to extract them with sufficient accuracy to meet our requirements.

The trouble is that LLMs are far from consistent. Asked the same question at different times, they will most likely give a different answer each time. Until we can understand how and why a given model reaches the many different conclusions that it does, we will not be able to rely on LLMs to plug the judgment gap in the automation of compliance tech.

It is not until we can be reliably sure that the model will always reach the same finding every time it is asked to analyse a given judgment that we will be able to depend on it.

To better understand why large language models behave as they do, frontier artificial intelligence (AI) laboratories have been working on the science of interpretability. In 2025, Anthropic’s researchers published maps of how their model arrives at an answer, showing, for instance, that it decides what the rhyme at the end of a line should be before it begins writing it.

Until recently, it was believed that this work of interpretability could only be done indirectly, since each neuron in a model was thought to carry so many unrelated concepts superimposed on one another that it was not feasible to read what concepts each of them influenced.

Last month, a researcher at interpretability lab Goodfire, who had helped develop the theory of superposition, reversed his stance, reporting that nearly nine in 10 neurons in the open-weight model he was studying performed one identifiable job. This seems to suggest that it should be possible to trace each step a model takes to arrive at an answer, and thereby identify the ones its conclusion actually depends on.

If this is true, we should be able to use this technology to reliably get an LLM to extract structured obligations from unstructured text—like that in a legal judgment. We should be able to see which passages it relies on, test whether it correctly identifies the right obligations every time it is asked, and find the places where it is merely guessing. If we can audit extraction in this manner, we will eventually be able to trust it more fully.

If we can do this, it would give our quest for a single source of truth an additional important layer. We will be able to ensure that alongside every statute, we also present the obligations our courts interpreted it to imply—extracted by machine, dated and duly revised every time a new judgment is delivered.

While the technology is still young, this is likely what we need in order to make AI more understandable to us. And while we are some distance from doing this at scale, the repositories we design today should be built keeping in mind that this is something we will eventually need to include.

The author is a partner at Trilegal and his latest book is ‘Ex Machina: Essays on a Decade of Tech Policy’. His X handle is @matthan.