Abstract
This paper examines the emerging threat to judicial cognitive integrity posed by the use of Large Language Models (LLMs) as preparatory tools in judicial decision-making. Drawing on established psychological and behavioural research, it argues that judges’ interaction with LLM outputs predictably generates automation bias and cognitive offloading, subtly shaping reasoning even where judges remain formally responsible for final decisions.
The paper challenges the prevailing assumption that judicial verification and transparency obligations are sufficient safeguards against such influence. It proposes an influence-focused disclosure and scrutiny framework, alongside a reorientation of judicial education away from tool-based training toward deeper engagement with the philosophy and sociology of technology.
Introduction
Judicial decision-making is a unique task done by judges regularly that requires the establishment of fairness in every stage. It comprises of several rigorous, technical, cognitive, emotional, expert processes to different degrees to be fair and neutral (Lopes, 2021). Nonetheless, research shows that several internal and external factors could influence judicial-decisions made by judges. These factors include intuitive bias, weight or merit of the case, public reactions, social media, news or public opinion, identity or social status of the parties, identity of the lawyers by whom litigants are represented before the court, automation bias, anchoring effect, statistical interference, personal favour of a judge, workload, time pressure, and insufficient resources etc. (Shahid et al., 2025, Kwame, 2024, Hanych, Smekal and Benák, 2023, Barry, 2021, Australian Law Reform Commission, 2021, Lopes, 2021, Fariña, Arce and Novo, 2003). In modern times, there are numerous attempts to explore the potentials of advanced technology such as Artificial Intelligence (AI) and particularly Large Language Models (LLMs) to assist the cognitive function of judges in making court decisions (Ayal, Elyoseph and Solomon, 2026, Grieman, 2021). At the same time, there are dangers of compromising cognitive integrity of judges through the integration of LLMs in judicial decision-making processes (Figueiredo Peluso Lopes, 2025, Pruss and Allen, 2025). As a result, the main focus of this paper is to assess how LLM tools affect the cognitive autonomy of judges.
Defining Cognitive Integrity for Judges
The cognitive integrity of judges denotes the ability of judges to deliver verdicts purely based on the applicable law and admissible evidence, totally free from any sort of cognitive contamination including personal biases, induced circumstances, or cognitive shortcuts (Cranston, 2025). This is uncompromisable and fundamental to ensure fairness and maintain public confidence in the justice system. The judges must perform their responsibilities consistently with the core judicial values of independence and impartiality, in conjunction with the moral and behavioural canons contained in conventions as well as written codes of conducts and guidelines. Maintaining integrity in decision-making and adjudicating is highly important for judges as it relates to competence, efficiency, and the good manners that the judge exhibits etc.
The Integration of Large Language Models as an Assistive Tool in the Bench
The integration of AI tools such as LLMs to assist judges to prepare or write judgments is growing globally (OECD, 2025, Javed and Li, 2024). This includes using AI tools to do legal research, document analysis, summarization of large documents, translation of legal instruments, precedent or source searching, opinion writing, and finding meaning of law lexicon etc. The proponents of these tools believe it brings efficiency by reducing cost, effort and time for judges and court staff on the one hand, and ensures consistency in judicial functions on the other hand (Ayal, Elyoseph and Solomon, 2026). Moreover, LLMs neither need incentives like wages, allowance, rest, training or leave, nor do LLMS lie or fear media comments (Sourav, 2025).
Nevertheless, while deciding to adopt new technology like LLMs, judicial authorities are required to be cautious and led by the solemn duty to deliver fair, just, accessible, independent, neutral justice for all, within a reasonable time. Technology must be integrated to implement and increase the delivery of the above principal justice values (Dias et al., 2025). The aim should not merely be to decrease the cost for the taxpayers of the country but rather to ensure a just system for them.
How the Application of LLMs Can Compromise Cognitive Integrity of Judges
There are numerous and prominent limitations in these technologies such as that LLM systems do not understand what it produces from a prompt or the core principles of justice, integrity, fairness, rule of law etc. (Lai et al., 2024, Stanford University IT, 2024). There is a perception that LLM tools generate the purity and accuracy of mathematical algorithms with a competence for decision-making free from human defects (Khatoon, Khan and Irshad, 2025). However, LLMs tools like ChatGPT generate an output based on vast training data that apply a probable next word or string of words in structuring a sentence. Consequently, it produces outputs like a ‘socratic parrot’ without having any true understanding of the meaning or, conceptual grounding.
Judicial bodies in various jurisdictions such as the UK, Ireland, Australia, New Zealand, Canada, Turks and Caicos Islands Hong Kong, Belize, Jamaica, etc., have issued guidelines as to how judges and judicial staff can use Generative AI (particularly LLMs) in their official functions to safeguard the fairness of justice (The Judicial Council, 2024). These guidelines usually allow judges to use these technology tools with greater caution to ensure the authenticity of the outputs, with the exception of core judicial functions like decision-making.
However, there is a widening gap between how the assistance of LLMs for judges works in the guidelines and how it actually serves in practice. In accordance with the aforementioned guidelines, modern LLM systems serve as a secondary tool that judges must cautiously oversee, verify, and rectify. The human judge still remains in the loop, employing critical judgment over each and every output. However, cognitive science demonstrates that the continuous use and application of scientific tools creates cognitive offloading (Gerlich, 2025, Chirayath, Premamalini and Joseph, 2025). It is not intentional that human judges evade crosschecking the veracity of outputs, rather it is a natural human tendency to offload cognitive labour to scientific and mathematical systems (Nogacki, 2025).
The growing practice of assistive reliance on AI-powered LLM tools for doing judicial functions raises psychological concerns of automation bias or the abdication of human judgment to a scientific system when humans are more inclined to simple rubber stamp decisions made by algorithms (Kazim and Tomlinson, 2023). Though judges are trained to be impartial, thoughtful, and unaffected by any influence, cognitive, neuro and behavioural science demonstrates that no decision-maker is immune to systematic psychological biases (Weinstein, 2002, Yalcin, 2025). Under the circumstance of judicial workloads, exhaustion, and, in some cases, institutional direction to deliver timely decisions with limited resources, this can lead to such psychological issues. As a result, it will compromise the cognitive integrity of judges and eat away the fundamental component in judicial decision-making. Consequently, there is a high risk that the initial support offered by LLMs may eventually substitute cognitive independence of judges.
Behavioural studies reveal that humans consistently interacting with computational systems primarily receive machine outputs with scepticism although over time, they build trust that ultimately motivates them to accept the outputs uncritically (Beck et al., 2025, Kabashkin, 2025). This mental process happens silently and slowly without deliberate delegation by human judges to the machine. Even judges being highly trained professionals and well qualified in critical thinking and rational judgment are not at all resistant to this natural human cognitive process (Nogacki, 2025).
For instance, advanced LLM tools like ChatGPT, Gemini, Claude and so on present a polished draft that looks sound from the surface and this can encourage humans to depend heavily on such tools without questioning. Although judges may refine or edit the language and references provided by an LLM, there is a risk that the LLM’s reasoning could be incorporated into the final verdict, potentially displacing human legal reasoning grounded in established legal principles.
This illustrates the delegation of judicial reasoning, a subtle allocation of the interpretive act from the human judge to the algorithmic systems (Linna and Linna, 2026). The following adjustments by judges may amount to mere cosmetic refinement. Because the inherent habit of the human mind is to preserve cognitive resources and to reduce mind labour, this works against this idea that judges will comprehensively examine the outputs generated by LLMs (Gera et al., 2023). This transformation happens not from inattentiveness, but as a predictable consequence of cognitive offloading when the mind interacts with systems perceived as competent and reliable (Gerlich, 2025).
Thus, the uncritical adoption of these technologies creates a deep tension with the fundamental and uncompromising judicial duty of independent reasoning. Without precise safeguards and a deep understanding of cognitive susceptibility, the integration of LLM tools, even as a secondary tool in courts, could unintentionally undermine the very deliberative process that the use of such tools seeks to augment.
The cognitive faculties of human judges that are essential for legal reasoning such as analogical thinking, doctrinal synthesis, and value-based judgment undergo a form of skill degradation or professional deskilling when their exercise is routinely outsourced to automated systems. Hence, the role of these assistive LLM tools ultimately question the cognitive integrity of judges while adjudicating cases in court.
The random use of earlier technological evolution such as the frequent use of the calculator to solve mathematical problems or the use of advanced navigational tools, including the assistance of digital maps while driving, support the above argument of subtle human deskilling through over dependence on technological tools. Similarly, the persistent application of LLM systems affect the analytical skill to construct a compelling alternative analysis other than an uncritical acceptance of the machine output.
Consequently, judicial over-reliance on LLM systems create a paradox where it can erode the cognitive autonomy that it was meant to assist. A judge habituated to rely on algorithmic support may lack the sharpened analytical skill to effectively challenge or replace its outputs. This is especially perilous for cases that fall outside predictable patterns, those requiring nuanced moral, equitable, or socially contextual judgments that lie beyond algorithmic capability. Accordingly, here the threat lies not in robot or AI replacing human judges but humans becoming mere verifiers of machine-generated content.
So, if LLMs assisting judicial work is allowed without necessary caution, then the bias towards computer generated outcomes pose a risk to alter human judges’ autonomous thinking and reasoning process into a mathematical process wherein there is a high chance that the summaries, search results, word meaning or legal interpretations that the LLMs present become anchors that influence judicial thought. That is why individual judges’ independence to adjudicate impartially cannot co-exist with the assistive presence of LLM tools.
The Way Forward to Keep the Cognitive Integrity Intact in the Judiciary
The solution lies not in refining human-AI interaction, but in externally reinforcing the non-delegable nature of adjudication through structural reform and by internally rebuilding the judiciary’s epistemic resilience through pedagogical transformation. To do this, further accountability tiers can be introduced to ensure transparency in judgments for both the judiciary and litigants as well.
Although conventionally judges are not required to disclose their preparatory works, there should be some changes to this established custom in the age of LLMs. For obvious reasons, if a judge is consulting tools like LLMs then a mere declaration of use is insufficient but rather an ‘Influence Disclosure and Scrutiny’ system could be introduced.
Judges should be required to file a brief adjunct memo with any judgment declaring where any AI or LLM was consulted, specifying the tool’s nature, role, its key outputs, and, crucially, how the judge’s final reasoning critically engaged with, or diverged from, those outputs. This practice may serve several purposes, i.e., it forces meta-cognitive reflection on the judge, it creates an auditable record for the appellate or superior court to potentially check the extent of delegation, and it demystifies how AI was involved in the judgment for the parties concerned. It shifts the burden from the aggrieved person proving the harm before a superior or appellate court to the judge demonstrating active intellectual command. Thus, no appellate or superior court should consult any such tools in order to be fully free from such influences to ensure fairness and transparency at the higher level.
In addition, there is a need to shift away from the current training that focuses on functionality and how to use LLMs to training that is underpinned by the philosophy and sociology of technology. Mandatory judicial curricula should include the epistemology of LLMs, critical algorithm studies, and the sociology of automation etc. This literacy and understanding will empower judges to deconstruct an LLM output’s persuasive authority and recognize its inherent limitations with regards to moral reasoning and social context. Ultimately, the goal is not to make judges better users of LLMs, but to ensure they remain its unquestioned masters.
Conclusion
It can be argued that the requirement of robust and strict human oversight may not be successful to tackle this significant problem. Instead, policy makers should think more cautiously as to the ways that the judiciary can utilize the benefits of technological tools for the bench. Thus, to prevent the silent delegation of legal reasoning and to keep judges’ cognitive sovereignty intact, a thorough revision of the present practice is urgently needed. The policy makers and leaders of the judiciary must safeguard judicial independence and the cognitive autonomy of all judges from any sort of influence and must amend the judicial set-up accordingly. Consequently, judiciaries need to rethink and restructure how judges deal with LLMs and create guidelines and necessary training outlining the ways that LLMs should (and should not) be used in the bench.
References
Australian Law Reform Commission (2021) Cognitive and Social Biases in Judicial Decision-Making (Background Paper JI6). Available at: https://www.alrc.gov.au/wp-content/uploads/2021/04/JI6-Cognitive-Biases-in-Judicial-Decision-Making.pdf (Accessed: 3 February 2026)
Barry, B. (2021) ‘How our judges judge – the science and consequences’, Law Society Gazette Ireland, 25 January. Available at: https://www.lawsociety.ie/gazette/top-stories2/how-judges-judge–the-science-and-its-consequences#:~:text=Subtle%20factors,decision%2Dmaking%20in%20different%20contexts. (Accessed: 3 February 2026)
Beck, J., Eckman, S., Kern, C. and Kreuter, F. (2025) Bias in the loop: how humans evaluate AI-generated suggestions. arXiv:2509.08514v1 [cs.HC]. Available at: https://arxiv.org/html/2509.08514v1 (Accessed: 3 February 2026).
Chirayath, G., Premamalini, K. and Joseph, J. (2025) ‘Cognitive offloading or cognitive overload? How AI alters the mental architecture of coping’, Frontiers in Psychology, 16, 1699320. Available at: https://doi.org/10.3389/fpsyg.2025.1699320 (Accessed: 3 February 2026).
Cranston, R. (2025) ‘Integrity’, in Judging. Oxford: Oxford University Press. Available at: https://doi.org/10.1093/9780191890994.003.0004 (Accessed: 3 February 2026).
Dias, J.P., Henriques, M., Fernando, P., Barros, D. and Cid Teles, M. (2025) Digitalising justice: a fundamental rights-based approach. Luxembourg: Publications Office of the European Union.
Fariña, F., Arce, R. and Novo, M. (2003) ‘Cognitive bias and judicial decisions’, in Vanderhallen, M., Vervaeke, G., Van Koppen, P.J. and Goethals, J. (eds.) Much ado about crime. Brussels: Uitgeverij Politeia NV, pp. 287-304.
Figueiredo Peluso Lopes, G. (2025) ‘Bias in adjudication and the promise of AI: challenges to procedural fairness’, Law, Technology and Humans, 7(1), pp. 47-67. Available at: https://doi.org/10.5204/lthj.3812 (Accessed: 3 February 2026).
Gazal Ayal, O., Elyoseph, Z. and Solomon, A. (2026) ‘Evaluating Large Language Models as Judicial Decision-Makers’, Justice Quarterly, pp. 1–36. doi: 10.1080/07418825.2026.2618254.
Gera, R., Bar Or, M., Tavor, I., Roll, D., Cockburn, J., Barak, S., Tricomi, E., O’Doherty, J.P. and Schonberg, T. (2023) ‘Characterizing habit learning in the human brain at the individual and group levels: a multi-modal MRI study’, NeuroImage, 272, 120002.
Gerlich, M. (2025) ‘AI tools in society: impacts on cognitive offloading and the future of critical thinking’, Societies, 15(1), p. 6. Available at: https://doi.org/10.3390/soc15010006 (Accessed: 2 February 2026).
Grieman, K. (2021) Judge dread: AI and judicial integrity [Online]. The United Nations Office on Drugs and Crime (UNODC). Available at: https://www.unodc.org/dohadeclaration/en/news/2021/30/judge-dread_ai-and-judicial-integrity.html (Accessed: 3 February 2026).
Hanych, M., Smekal, H. and Benák, J. (2023) ‘The influence of public opinion and media on judicial decision-making: elite judges’ perceptions and strategies’, International Journal for Court Administration, 14(3), pp. 1–19.
Javed, K. and Li, J. (2024) ‘Artificial intelligence in judicial adjudication: semantic biasness classification and identification in legal judgement (SBCILJ)’, Heliyon, 10(9), e30184. Available at: https://doi.org/10.1016/j.heliyon.2024.e30184 (Accessed: 3 February 2026).
Kabashkin, I. (2025) ‘Cognitive atrophy paradox of AI–human interaction: from cognitive growth and atrophy to balance’, Information, 16(11), 1009. Available at: https://doi.org/10.3390/info16111009 (Accessed: 3 February 2026).
Kazim, T. and Tomlinson, J. (2023) ‘Automation bias and the principles of judicial review’, Judicial Review. Available at: https://doi.org/10.1080/10854681.2023.2189405 (Accessed: 3 February 2026).
Khatoon, H., Khan, M.L. and Irshad, A. (2025) ‘Large Language Model (LLM) algorithms in reshaping decision-making and cognitive biases in the AI-leading world: an experimental study’, PsyArXiv [Preprint]. Available at: https://doi.org/10.31234/osf.io/xy82z (Accessed: 2 February 2026).
Kwame, D. (2024) ‘The influence of social media on judicial independence’, Journal of Modern Law and Policy, 4(1), pp. 49-61.
Lai, J., Gan, W., Wu, J., Qi, Z. and Yu, P.S. (2024) ‘Large language models in law: a survey’, AI Open, 5, pp. 181-196. Available at: https://doi.org/10.1016/j.aiopen.2024.09.002 (Accessed: 3 February 2026).
Linna, E. and Linna, T. (2026) Challenges for generative AI in legal reasoning. arXiv:2508.18880v2 [cs.AI]. Available at: https://arxiv.org/abs/2508.18880 (Accessed: 3 February 2026).
Lopes, G.F.P. (2021) ‘Cognitive enhancement in the judiciary’, Direito, Estado e Sociedade, (59), pp. 283-315.
Lopes, G. (2021) ‘Cognitive bias in the judiciary: the role of cognitive enhancement’, in Legal Knowledge and Information Systems: JURIX 2021: The Thirty-fourth Annual Conference, Vilnius, Lithuania, 8–10 December 2021.
Nogacki, R. (2025) The silent surrender: judicial cognitive atrophy in the age of AI [Online]. LinkedIn, 15 April. Available at: https://www.linkedin.com/pulse/silent-surrender-judicial-cognitive-atrophy-age-ai-robert-nogacki-xdpsf/ (Accessed: 3 February 2026).
Organisation for Economic Co-operation and Development (OECD) (2025) Governing with artificial intelligence: the state of play and way forward in core government functions. Paris: OECD Publishing. Available at: https://doi.org/10.1787/795de142-en (Accessed: 3 February 2026).
Pruss, D. and Allen, J. (2025) ‘Against AI jurisprudence: large language models and the false promises of empirical judging’, in *Proceedings of the Eighth AAAI/ACM Conference on AI, Ethics, and Society (AIES ’25)
Sourav, R. (2025) Relying on AI in judicial decision-making: justice or jeopardy? [Online]. UCD PublicPolicy.e, 4 March. Available at: https://publicpolicy.ie/papers/relying-on-ai-in-judicial-decision-making-justice-or-jeopardy/ (Accessed: 3 February 2026).
Stanford University IT (2024) AI Demystified: Introduction to large language models [Online]. Stanford University. Available at: https://uit.stanford.edu/service/techtraining/ai-demystified/llm (Accessed: 2 February 2026)
The Judicial Council (2024) Guidelines for judges – the use of generative artificial intelligence. Dublin: The Judicial Council.
Weinstein, I. (2002) ‘Don’t believe everything you think: cognitive bias in legal decision making’, Clinical Law Review, 8, pp. 783-822. Available at: https://ir.lawnet.fordham.edu/cgi/viewcontent.cgi?params=/context/faculty_scholarship/article/1421/&path_info=9ClinicalLRev783.pdf (Accessed: 3 February 2026).
Yalcin, N. (2025) ‘From bench to jury box: the impact of bias on judges and jurors’, Critical Debates in Humanities, Science and Global Justice, 5(2), pp. 1–12.
RaisuL Sourav is a Doctoral Researcher at the School of Law, University of Galway, Republic of Ireland; an Associate Professor of Law (on leave) at Dhaka International University, Dhaka, Bangladesh and an Advocate (Lawyer) at the Supreme Court of Bangladesh. He holds a Master’s Degree in International Energy Law and Policy from the University of Stirling in the UK. He was a recipient of the prestigious British Chevening Scholarship in 2017-18. RaisuL regularly writes in academic journals, blogs and newspapers, and presents his research in reputed academic conferences. He can be reached at M.SOURAV1@universityofgalway.ie His X Handle: @RaisuLSourav