What one Supreme Court case reveals about AI and legal reasoning
When a case raises questions about both which legal rules apply and how they should be interpreted, the reasoning behind the outcome becomes particularly important. That made a case concerning the Danish Credit Agreements Act particularly interesting for my research into AI and legal decision-making.
A consumer had used credit to pay in advance for dental treatment. The dentist chain subsequently went bankrupt. The question was whether the consumer still had to repay the third party lender under a "buy now - pay later" arrangement, where the consumer paid no interest and costs, but the seller paid a commission fee to the lender.
Before the Supreme Court delivered its judgment, I asked eight AI systems to decide the case and explain their reasoning. This meant that the Supreme Court judgment itself was not part of the material on which they could base their answers.
Testing the reasoning as well as the result
The experiment included both general-purpose AI systems and platforms developed for legal work. Each received the same basic task, tested with three versions of the supporting material: the facts and relevant legal framework; that material supplemented by reasoning from three earlier judgments; and a further version including later legislative preparatory materials whose relevance needed to be assessed.
I repeated the tests in six separate sessions, producing 146 responses. This allowed me to examine how answers varied across systems, repeated attempts and changes to the context of supporting material.
When the Supreme Court ruled in favour of the lender, I could compare the responses with both its outcome and its reasoning.
A consistent outcome can conceal different reasoning
The result of the experiment gives rise to many observations amongst others:
One system, found in favour of the lender in all 18 of its responses despite the opposite ruling in the High Court. Its reasoning nevertheless differed from the Supreme Court’s. Other systems changed their conclusions when the supporting material changed.
A change of conclusion is not necessarily a weakness. Additional legal material may justify a different assessment. The question is whether the system has correctly assessed the relevance and weight of that material and can provide a sound explanation for the change.
The AI-systems that did not hit the same result as the Supreme Court provided plausible outcomes in most instances; We should not be surprised that AI-systems do not consistently agree with themselves and among each other (this is also seen with human judges).
The distinction matters when evaluating AI. Agreement with a court’s outcome does not, by itself, establish the quality of the reasoning. Equally, a consistently repeated answer needs to be examined on its merits.
Keeping legal judgement in the process
For lawyers and businesses using AI in legal work, these findings raise a practical question: what does meaningful review of an AI-generated assessment require?
My view is that it must include examining the sources relied on, the weight given to them and the connection between the legal arguments and the conclusion. A response that reads convincingly still needs to withstand that scrutiny.
The experiment concerns one case only. It does not establish a general success rate for AI in legal decision-making, or provide a basis for ranking the systems across legal tasks. Several of the specialist platforms were not designed to predict court judgments.
As part of my PhD research, I am exploring how judges make decisions and what their approaches can teach us about the use of AI. The aim is to better understand the risks and the measures needed to use these tools responsibly.
For me, the next question is how we build that understanding into everyday legal work, so that lawyers remain able to challenge an answer and explain why its reasoning should be accepted.
By Niels Chr. Ellegaard, Attorney-at-Law, Partner, Plesner