top of page

Keeping the Human Touch: Guardrails for Courts and Lawyers Adapting to AI in the UAE

Mihika Gautam
3 days ago
4 min read

-- By Mihika Gautam


At the 15th UN Congress on Crime Prevention and Criminal Justice, held in Abu Dhabi, the UAE Ministry of Justice made clear that “no decision that affects an individual can be made by AI alone.” 


This principle of human oversight runs through several of the UAE’s AI guidelines. And yet, at the same time, the UAE government is adopting AI faster than ever, leading to efficiency and productivity gains across public services, healthcare, transport, energy and education. For instance, in October 2017, the UAE established the UAE National Strategy for Artificial Intelligence 2031, aiming to make the UAE a global leader in AI by 2031. In April 2026, it went further, committing to move 50% of government sectors, services and operations to agentic AI within two years. Just earlier this month, OpenAI signed with the UAE for AI development across energy, investment and government sectors.


The UAE’s approach is iterative: it is deploying AI and refining the rules that govern it at the same time, balancing faster adoption against vigilance. But what does this iterative approach mean for justice? Does human oversight come at the cost of efficiency? Or does machine-led decision-making risk delivering hollow justice?


As AI moves into the courtroom, the real test is whether human judgement stays present at every stage of decision-making. The Ministry of Justice’s agentic AI systems, and the approaches taken by the courts of the UAE’s two financial free zones, the Dubai International Financial Centre (DIFC) and Abu Dhabi Global Market (ADGM) offer valuable insights.



// Federal Courts, Agentic AI and Human Judgement


The Ministry’s agentic AI system for legal proceedings shows this iterative approach in practice: the technology is being deployed while the rules governing it are still taking shape. It involves seven stages, five of which are supported by agentic AI models in sequence. The system breaks a case into seven stages, five of which are handled in sequence by agentic AI models.


A case goes through a file review and ends with a non-binding draft for the judge to review, reducing the process from several days to a few hours. The final two stages are exclusively for the judge’s input, where they review, amend and approve what is generated. AI provides most of the groundwork for the judge’s reasoning; thus, the test is not only about the accuracy of the AI-generated output, but also about how complete the whole picture is, because AI can carry biases and follow how a question is framed, just as humans do.


The Ministry’s new AI ethics code sets out the principles of human accountability and oversight, but there is no published internal policy telling judges what each stage involves or how to review AI-generated work. The next step is an internal policy that directs judges to verify both the accuracy and the completeness of that work, and to record that they have done so.


// Standard Operating Procedures and Flexible Guidelines


The federal courts can look closer to home for a model. The DIFC Courts were among the first in the region to publish AI guidance: Practical Guidance Note No. 2 of 2023 on the use of large language models and generative AI in proceedings. It sets out principles of transparency, accuracy, reliability and avoiding over-reliance, along with “best practice” such as verifying AI-generated content and telling the other side in advance if AI will be used. Because it is guidance rather than binding rules, it can adapt as the technology changes, which is the iterative approach in action.


//Learning from Mistakes


Guidance only works if it is followed, and rules only if they are enforced. The cost of not verifying is already visible in the region’s courts. In ADGM courts' Arabyads Holding Ltd v Gulrez Alam Marghoob Alam [2025], the law firm that represented the defendant, Mr Alam, filed his Defence and cited “numerous false legal authorities”, according to Arabyads. The court concluded that the firm’s errors had “the hallmarks of the hallucinatory results that can occur when AI is used as a research tool” and ordered the law firm to pay AED 282,508 in wasted costs to Arabyads Holding. Similar examples are showing up around the world, from New York to London.


There have also been cases of judgments with hallucinated content in the United States.


Together, these cases show that errors carry real costs in legal fees and court time. They tell lawyers that responsibility for every authority they cite cannot be delegated, whether the research came from a junior or a machine. As for parties to a dispute, they learn that they may end up paying for their counsel’s shortcuts, which introduces a new element of scrutiny into the lawyer–client relationship. And judges learn that the bench is no more immune than the bar: the scrutiny they apply to counsel’s submissions must apply equally to their own


From a policy perspective, they also show why the iterative approach is crucial. Each failure gives policmakers both a reason and the evidence to develop regulatory frameworks best suited to the problems use of AI is actually creating in courtrooms, not the ones anticipated on paper.


// The Underlying Principle


So does human oversight come at the cost of efficiency? The evidence suggests not. The Ministry’s system shows that AI can take on the groundwork, cutting case file review from days to hours, while the decision stays with the judge. And the costliest failures, in Abu Dhabi, London and New York, came from too little human checking, not too much. The real threat to both speed and justice is unverified output.


The UAE has built the institutional foundations. The new federal Artificial Intelligence and Data Authority now brings AI strategy and data governance under one roof. What comes next is turning principle into procedure. Courts in England, California and Canada have already issued internal guidance requiring judges to verify AI output and barring AI from core judicial functions. A system that goes further, drafting judgments, needs at least as much. That means binding protocols for judges and lawyers to check AI-generated work for accuracy and completeness and to record that they have done so, with the Authority coordinating standards alongside the Ministry of Justice and the DIFC and ADGM Courts. Technology and rules will keep evolving, but the Ministry’s principle should not: no decision that affects an individual can be made by AI alone.


Human judgement is not a brake on AI-assisted justice - it is what makes it worth trusting.

Recent Posts

See All
bottom of page