Does society trust the co-creation of judicial decisions using generative AI?

Digital tools are part of everyday legal practice. Generative AI (Gen-AI) greatly expands their capabilities: it can rapidly process large volumes of text, autonomously structure relevant information, and generate tailored wording suggestions and even proposed judgments. It promises to reduce workloads, simplify procedures, free up time for complex cases and enable more impartial, ‘blind’ adjudication by reducing the influence of human error and bias. However, Gen-AI can also produce false information or fabricated sources and reproduce underlying societal biases in its judgments. A particular concern is that users often cannot understand how its outputs were generated, even though these outputs can appear highly convincing. The stakes are particularly high in judicial decision-making, given its far-reaching consequences.

The prospect of defendants being acquitted or convicted by Gen-AI alone remains an abstract future scenario. The EU AI Act classifies AI systems used in judicial decision-making as high-risk, subjecting their use to specific requirements. For example, Article 14 requires human oversight of high-risk AI systems, so that human judges retain responsibility for decisions.

While previous studies have shown a clear preference for human judges over AI, it remains largely unclear how the public perceives co-creation (i.e. close, iterative collaboration) with Gen-AI in judicial decision-making. Recent findings from other fields suggest that AI users are perceived as less competent, less diligent and lazier than non-users. This raises practical questions: Does society perceive judges who use Gen-AI in judicial decision-making as less trustworthy in terms of ability, benevolence and integrity? Does a judgment lose public acceptance simply because judges use Gen-AI?

Different paths to the same verdict

In five online experiments involving a total of 2,515 adults from Germany, the bidt-funded project ‘How “human” should justice be? Psychological determinants of trust in co-creation with generative AI in the context of the legal system (Gerechte KI)’ investigated how people evaluate co-creation with Gen-AI in judicial decision-making. Participants read different versions of fictional criminal cases, including descriptions of the court proceedings and the resulting judgment. The only aspect that varied between participants was the extent to which judges incorporated Gen-AI into the decision-making process.

In three experiments, the researchers distinguished between four decision-making variants. Participants were presented with different descriptions of court proceedings and asked to imagine the situations described. In the first, ‘purely human’ variant, a judge presided over the court hearing, reviewed all key information and issued the judgment. In a second variant, ‘co-creation with high human involvement’, a locally operated Gen-AI system that complied with data protection requirements produced a summary of the information relevant to the decision; the judge reviewed it and then reached a decision. In a third variant, ‘co-creation with low human involvement’, the AI additionally proposed a possible judgment. The judge could accept or reject this proposal and formally retained final decision-making authority. In the fourth variant, Gen-AI conducted the proceedings entirely on its own, without a human judge’s involvement, and also issued the judgment. This final variant is not currently realistic, but it allows for comparison with previous studies that distinguished between human judges and AI.

In two further experiments, the researchers compared the ‘purely human’ variant with ‘co-creation with low human involvement’, this time asking participants to adopt the victim’s perspective.

All participants subsequently rated the perceived ability of the decision-making entity, its perceived benevolence (towards society or towards themselves as victims) and its perceived integrity or impartiality. They also indicated the extent to which they would accept the judgment.

Does the public perceive co-creation with genetic AI as trustworthy and acceptable?

For the most part, yes – provided that human judges were described as being responsibly involved. Only in one experiment did the ratings of the judge’s perceived trustworthiness and the acceptance of the judgement decline as the involvement of Gen-AI increased. In the remaining four experiments, the purely human and co-creation variants received largely comparable ratings for perceived trustworthiness and acceptance of the resulting judgements.

However, the fully automated variant fared significantly worse. In other words: Gen-AI as the sole decision-making authority is rated as less capable, benevolent and principled than judges with and without AI support. The judgements handed down by Gen-AI were also rated as less acceptable.

Does the nature of the case make a difference?

Judicial decisions often require not only knowledge of the law but also empathy, moral judgement and sensitivity to context. The researchers therefore compared perceptions of criminal offences that were technical in nature and those that were emotionally charged. When Gen-AI decided alone, it was perceived as less competent in an emotionally charged fraud case than in a tax offence. This difference did not arise in decisions made by human judges or through co-creation.

One experiment varied the emotional complexity of the same assault case in a more targeted manner. Here, no additional negative effect on perceptions of Gen-AI was observed. Judgements in the more emotionally complex case were, on the whole, slightly less accepted – regardless of who made the decision. It cannot be conclusively stated whether emotional cases actually reinforce scepticism towards AI.

Human responsibility remains key

What can be concluded from these findings? A clear guideline emerges for the introduction of Gen-AI in courts: Gen-AI could serve as a support to judges without leading to a loss of trust in society, but it should not replace them. Meaningful co-creation combines AI’s ability to process vast amounts of information quickly with human contextual sensitivity, legal responsibility and the ability to critically reject suggestions.

However, the study examined hypothetical cases in which the defendants had pleaded guilty and, above all, the sentence had to be determined. Real-life proceedings and contested questions of guilt could be associated with more negative assessments of generative AI compared to judges.

In conclusion, it can be said that the public appears to regard judicial co-creation with generative AI in the sentencing process as trustworthy and to accept it. This applies, however, only when generative AI is used as a supportive tool within the judiciary – that is, as long as human actors are involved. The public reacts particularly critically when AI takes over the role of the judge. By contrast, a transparent, controlled co-creation process does not necessarily undermine trust and acceptance. The issue is therefore less a question of ‘AI or no AI?’ and more a question of who reviews, who justifies and who ultimately bears responsibility for the decision.

The blog posts published by bidt reflect the views of the authors; they do not reflect the position of the institute as a whole.

---