Can large language models predict Court decisions?

Thursday 16 April 2026

The suggestion that large language models might one day predict judicial decisions with reliable accuracy sits at the intersection of two very different conceptions of law. On the one hand stands law as a system of patterns, precedents and observable behaviours, susceptible to statistical modelling. On the other stands law as a human institution of judgment, discretion and moral reasoning, shaped by experience, context and the often unpredictable dynamics of deliberation. It is within this tension that the recent reflections of Sonia Sotomayor acquire their significance.

Justice Sotomayor has, in public remarks, acknowledged that artificial intelligence will have a “huge” effect on the legal profession, altering how lawyers work even if it does not replace them entirely. That formulation is cautious rather than dismissive. It recognises both the transformative potential of machine reasoning and its limits. Implicit in her position is a distinction between assistance and substitution, between tools that augment legal reasoning and systems that purport to replicate judicial judgment itself.

The modern enthusiasm for predictive legal analytics rests upon a body of empirical research that is, at first glance, impressive. Statistical models trained on decades of judicial data have achieved accuracy rates approaching seventy per cent in forecasting the outcomes of United States Supreme Court cases. Similar exercises applied to the European Court of Human Rights have produced comparable figures, albeit with the sobering caveat that simple heuristics based on historical tendencies may outperform more sophisticated models. These findings suggest that courts, when viewed from a sufficient distance, exhibit patterns. Judges are influenced by precedent, institutional constraints, ideological leanings and the structure of legal argument. To that extent, the law is not wholly indeterminate.

Yet the very success of these models reveals their limitation. A seventy per cent accuracy rate is not a triumph of determinism; it is an admission of uncertainty. Three cases in ten remain mispredicted. In the context of law, where each case may carry profound consequences for individuals, institutions or even the constitutional order, such a margin of error is not trivial. It is the difference between a system that describes tendencies and one that genuinely understands judgment.

Justice Sotomayor’s own judicial career illustrates why this gap persists. Her record on the United States Court of Appeals for the Second Circuit defied simple ideological classification, showing neither consistent pro-business nor anti-business tendencies and adhering closely to precedent and fact-specific reasoning. Such variability is precisely what frustrates predictive models. It reflects the reality that judicial decision-making is not reducible to a stable set of inputs. It is instead an iterative process in which legal principles, factual nuance and institutional considerations are continually rebalanced.

Moreover judicial reasoning operates within a social and moral context that resists quantification. Sotomayor has repeatedly emphasised the importance of public confidence in the courts and the dangers posed when that confidence erodes, particularly when precedents are overturned too rapidly. These concerns are not merely rhetorical. They shape how judges approach cases at the margins, how cautiously they extend doctrine, and how they frame their reasoning in written opinions. A machine trained on past data may capture the outcomes of prior cases, but it cannot easily internalise the evolving relationship between a court and the society it serves.

There is also the problem of deliberation. Courts, especially apex courts, do not decide cases in isolation. They are collegial bodies in which arguments are tested, positions modified and compromises forged. The internal dynamics of such deliberation are largely opaque and often decisive. A justice may alter her vote in response to a colleague’s reasoning, or tailor her opinion to secure a majority. These processes are not visible in the data upon which predictive models rely. They belong to the realm of human interaction rather than statistical inference.

The recent experience of artificial intelligence in legal practice further underscores these limitations. Courts have already encountered instances in which AI-generated submissions contained fabricated authorities or erroneous reasoning, leading to sanctions and warnings about the need for human verification. If models struggle to reproduce the basic discipline of legal citation, their capacity to emulate the higher-order reasoning of judicial decision-making must be regarded with caution.

None of this is to deny the utility of predictive systems. On the contrary they may prove invaluable as tools for lawyers seeking to assess litigation risk, to identify patterns in judicial behaviour, or to refine their arguments. They may even contribute to a more transparent understanding of how courts operate, highlighting the influence of factors that might otherwise remain implicit. Artificial intelligence might serve as a mirror held up to the law, reflecting its structures and tendencies back to those who practise it.

But a mirror is not a mind. The aspiration to predict court decisions with machine precision rests upon a category error. It assumes that judging is fundamentally a problem of pattern recognition, when in truth it is also an exercise in interpretation, persuasion and, at times, moral choice. Justice Sotomayor’s cautious embrace of artificial intelligence reflects an awareness of this distinction. She acknowledges the transformative potential of these technologies while implicitly resisting the notion that they can displace the human element at the heart of adjudication.

The deeper question therefore is not whether large language models can predict court decisions, but what it would mean if they could. A legal system that became fully predictable would, in a sense, cease to be a system of judgment. It would become a system of calculation, in which outcomes are determined not by deliberation but by algorithmic inevitability. Such a transformation would challenge the very legitimacy of the judiciary, which rests upon the idea that cases are decided by independent minds applying the law to the facts before them.

For the present that prospect remains distant. The empirical achievements of predictive modelling, while notable, fall far short of displacing human judgment. The law retains its character as a domain of uncertainty, shaped by reasoning rather than mere regularity. In this domain artificial intelligence may assist, illuminate and even occasionally surprise. But it does not, and perhaps cannot, replace the act of judging itself.