New Irish research has raised concerns about the impact of the use of Artificial Intelligence on access to justice and on fairness and reliability in justice systems worldwide.
Having examined whether an AI judge can “be talked into a verdict”, the study, published on Monday, found the answer to that question is “yes”, and “sometimes to a striking degree”.
Persuadability and LLMs as Legal Decision Tools, by Oisin Suttle of Maynooth University’s School of Law and Criminology and David Lillis of UCD’s School of Computer Science, is described as the “first systematic study” of how persuadable AI models are when confronted with hard legal questions from real cases.
It is published as the Irish judiciary is preparing new guidance for practitioners and parties on AI use.
RM Block
“It is important that a judge should be open to persuasion. Otherwise, there would be no point in letting the parties argue their case,” said Suttle.
“However, we don’t want a judge to be too persuadable. They should decide on the law and the facts of the case, rather than being swayed by an especially skilled advocate.”
The study results show that AI models in legal settings can be “very persuadable” and that raises worries about both reliability and fairness, he said.
It is “essential that society and decision-makers don’t simply follow hype when it comes to the capabilities of modern AI systems”, said Lillis, stressing the importance of scientific investigation of their capabilities and characteristics “particularly in sensitive areas such as legal decision-making”.
The research examined the extent to which LLMs (Large Language Models) in decision support roles are subject to persuasion in legal settings.
Frontier AI models, it found, are “significantly influenced” by the quality of legal arguments presented to them.
The researchers designed an experiment using real appellate court cases drawn from Ireland, England and Wales, and the United States. They pitted AI models against each other in a simulated legal contest with one AI acting as an “Advocate” arguing one side of the case, another arguing the other side, and a third playing the role of “Judge.”
Having varied which AI model argued each side and measuring how often each “won,” the team quantified how much argument quality shifted the outcome.
Every AI model tested was “measurably persuadable”, the study found.
Across 20 different model configurations, including leading systems from Anthropic, Google, OpenAI, Mistral and DeepSeek, stronger advocate models won between 58 per cent and 71 per cent of cases on average. In the “most extreme” matchups, the stronger advocate won more than 90 per cent of the time.
Noting that administrative agencies, courts, and tribunals in multiple jurisdictions are piloting AI systems to handle everything from research to case triage and judgment drafting, the researchers argue that persuadability is a dimension of these systems that has received almost no attention until now.
The implications are “particularly acute for access to justice”, they say.
If an AI decision-maker is heavily influenced by the quality of legal argumentation, parties with access to sophisticated legal representation or more powerful AI advocates “will have a systematic advantage over those without”.
The study found larger, more capable AI models tended to be “somewhat less persuadable” than smaller models but cautioned against interpreting that as reassurance.
Even the least persuadable models showed “significant sensitivity” to argument quality, the researchers said.
The reasons for lower persuadability in smaller models also appeared, “crucially”, to be different. Rather than exercising independent judgment, “these models may simply struggle to evaluate competing arguments at all”.
While not saying AI should not be used in legal settings, the study argues persuadability must be measured and disclosed as a standard part of evaluating any AI legal decision tool.
Persuadability and LLMs as Legal Decision Tools, available here, was presented on Monday at the 21st International Conference on Artificial Intelligence and Law in Singapore.
[ ‘More harmful than helpful’: Young people sour on AIOpens in new window ]
