General Large Language Models (LLMs) such as ChatGPT, Gemini, Claude, and DeepSeek, and even field-specialised models, are being widely used in almost every sector, including healthcare, business, engineering, education, law enforcement, and even various administrative decision-making processes. The judiciary is also no exception: across the world, judges and court staff are applying LLM systems in court proceedings. The application of LLMs and other Artificial Intelligence (AI) tools in the judicial decision-making process, even for non-core judicial functions, creates a tension in the procedural integrity of adjudication where these tools have potentials to improperly influence the final decision to the detriment of fair trial rights.

There are projects in some jurisdictions – including China, Brazil, and Argentina – as to how these technologies could be employed in the judicial decision-making process to reduce case backlogs and enhance the delivery of justice. Contrarily, some countries – such as the UK, Ireland, New Zealand, and Canada – have taken an opposing stance to prevent the adoption of LLMs in core judicial functions, including determination and adjudication. Instead, they permit LLMs as support tools for judges and court staff to perform routine work like summarising text, speech writing, searching for information, sources, and precedents, enhancing clarity and linguistic flourish, translating, and transcribing.

However, such ‘innocent’ use of LLM tools in routine or secondary court functions has the potential to influence judicial cognition and thereby affect the core judicial decision-making process. Institutional permission and repetitive use of these tools may lead to several cognitive effects, such as automation bias, cognitive offloading, anchoring, and the framing effect. Automation bias leads to over-reliance on automated recommendations, while cognitive offloading reduces mental effort through external tools. Anchoring causes over-dependence on initial information, and the framing effect shows how decision-making shifts based on information presentation.

Judging involves reading the case materials, identifying the issues, and developing the reasons by themselves, whereas a preliminary recommendation produced by an LLM system has the potential to serve as a starting point for such cognitive involvement of judges, anchoring the point of departure. On the other hand, when a judge outsources the meaning of legal lexicon, a summary of long texts, or the identification of precedents to an LLM, they ultimately engage in cognitive offloading. This cognitive offloading can lead to a superficial review process where judges make only ‘surface-level’ scrutiny or modifications to LLM-generated text, bypassing the deep cognitive work required to fully understand the legal reasoning and verdict.

Therefore, simply relying on a human judge to use LLM systems in preparatory stages, or to audit the veracity of the outcomes of such systems subsequently in the judicial decision-making process, is potentially problematic. Though judges are trained to be impartial, cognitive and behavioural science demonstrates that no decision-maker is immune to systematic cognitive biases. Decades of behavioural research demonstrate that professional expertise does not eliminate susceptibility to anchoring, framing, or automation bias. Judges, like other professionals, work within cognitive constraints.

These cognitive defects are significant because human judges are regarded as the principal safeguard to check and control any type of bias and are also regarded as neutral decision-makers to ensure justice. In the circumstances of heavy judicial workloads, exhaustion, and in some cases institutional pressure to deliver timely decisions with limited resources, cognitive issues can arise. As a result, it can undermine the fundamental principles of fair trial, including judicial independence, the judicial duty to provide well-reasoned decisions, and ultimately the rule of law.

The inherent persuasive features of LLMs that randomly generate text with human-like reasoning and confidence, in spite of factual or legal accuracy, may make human judges vulnerable to this uncritical adoption of LLMs even as a support tool for judicial decision-making. This creates a deep tension with the fundamental and uncompressible judicial duty of independent reasoning, signifying that without precise safeguards and a deep understanding of this cognitive susceptibility, the integration of LLMs in adjudication, even as secondary tools, could unintentionally undermine the deliberative process it seeks to augment. Hence, in an increasingly technology-driven judiciary, it is critical to examine whether judges’ cognitive limitations prevent them from effectively guarding against aggravating this tension.