Imagine an Islamabad courtroom with marble floors, piled case files, and clerks bustling between benches in fluorescent light. Incorporate a backlog of 2.2 million unresolved cases into that scene, which puts pressure on a legal system that just does not have enough time in the day to handle them. In March 2025, Pakistan’s Supreme Court rendered a ruling in the case of Ishfaq Ahmed v. Mushtaq Ahmed, which has since gained recognition in legal circles outside of South Asia. Artificial intelligence was approved by the court as a means of handling its heavy caseload. However, it made a distinction: AI could help. It might not make a decision.
In courts elsewhere, the line that seemed to be clearly drawn in Islamabad is already becoming less clear. Finding the first judicial opinion written by AI in a lower court raises an awkward secondary question: how many more existed before anyone became aware of them? The identification itself suggests a detection—a moment when someone read a decision and sensed something wasn’t quite right. Not incorrect in the sense that a factual error is. There is something about the texture. the sentences’ cadence. the manner in which arguments were put together. The legal community may not be aware of how long it has been reading judicial language produced by artificial intelligence.
The risks listed are no longer theoretical. Documented cases from 2024 and 2025 in U.S. federal courts included AI-generated court documents citing cases that were just nonexistent. The phenomenon where a generative AI system generates confident, well-formatted output that is completely fake is known by researchers as hallucinations. A hallucinated citation is not a typo in a legal document. The precedent is erroneous. a fictitious case added to an actual proceeding. Lawyers were penalized by multiple courts for this, and by the first quarter of 2026 alone, financial penalties for AI-related filing errors had surpassed $145,000. It appears that courts have become impatient.
| Field | Details |
|---|---|
| Landmark Case | Ishfaq Ahmed vs. Mushtaq Ahmed |
| Court | Supreme Court of Pakistan |
| Date of Decision | March 2025 |
| Significance | First formally identified AI-assisted judicial opinion at high court level |
| AI’s Role | Endorsed to address case backlogs; human oversight required |
| Pending Cases (Pakistan) | Approximately 2.2 million |
| Key Risk: Hallucinations | AI-generated non-existent legal citations found in U.S. federal court filings (2024–2025) |
| EU Classification | AI used in judiciary classified as “high-risk” under the EU Artificial Intelligence Act |
| Key Principle Established | AI must support, not replace, human judicial judgment |
| U.S. Court Disclosure Examples | Florida 17th & 11th Circuits require mandatory AI disclosure in filings (2026) |
| Research Framework | Justice 5.0 — human-centred AI integration in legal systems |
| MIT Study Finding | Reliance on LLMs may accumulate “cognitive debt” in users (2025, preprint) |

The current state of affairs in the judiciary is similar to that of all professions that swiftly embraced AI before having to work backward to develop a framework for its responsible use. One of the strongest regulatory signals in recent years, the EU Artificial Intelligence Act designated judicial AI systems as high-risk and mandated fundamental rights impact assessments prior to deployment. Any attorney or self-represented litigant using generative AI to prepare a filing must state this on the document’s face and attest that all cited authorities were independently verified, according to mandatory disclosure requirements introduced by Florida’s courts. Illinois adopted a different strategy, allowing the use of AI without disclosure as long as ethical requirements were fulfilled. Even though the rules vary, the fundamental idea is the same for all of them: human responsibility persists. However, putting that idea into practice in a courtroom is proving to be much more difficult than just stating it.
A specific issue that legal scholars have begun to refer to as the “black box issue” is crucial for appeals. The reasoning behind an opinion written by a human judge is contained in the text. You can argue that the conclusion does not follow from the premises, follow the logical steps, and point out instances in which the judge balanced one factor against another. When an AI system produces that same text, the real reasoning, if it can be called that at all, takes place inside a neural network that no one can meaningfully question. The litigant who wishes to contest the decision has access to a document, but they are unable to review the procedure. Centuries of procedural law, which is based on the idea that justice must not only be done but also be seen to be done, are at odds with this opacity.
As these changes mount, there’s a sense that the legal system is changing at a rate that the institutions in charge of it weren’t built to handle. There is a problem with the judicial docket. The 2.2 million pending cases in Pakistan are not an abstract concept; they are the result of years of unresolved conflicts, postponed justice, and waiting. Early in 2026, South Korea’s AI attorney reportedly won its first actual case in court by analyzing the evidence, referencing obscure case law, and crafting arguments. That is a truly remarkable ability. Risks and efficiency gains are driven by the same factors, and it is impossible to distinguish between the two.
Researchers have begun to refer to the idea that the legal community appears to be coming to—sometimes reluctantly—as Justice 5.0. It is believed that AI in the legal system must continue to be truly human-centered, not merely ostensibly so. A judge signing the results of an algorithm is not a rubber stamp. It is not a scenario in which a system that has been trained on past case data subtly reinforces the same prejudices that have influenced criminal justice outcomes for many years, especially when it comes to bail and sentencing decisions where algorithmic recommendations are given undue weight. A judge who truly reads, challenges, and accepts responsibility for every word in an opinion—regardless of the source of the initial draft—is what is meant by human oversight.
Heavy reliance on large language models may result in what the researchers called cognitive debt, according to a 2025 MIT Media Lab study that is still in preprint but has already been frequently cited in legal AI discussions. The judgment that is being outsourced does not vanish. It simply moves to the machine, and the tendency to form it on its own may wane with time. That is a professional risk for attorneys. It’s more akin to a constitutional one for judges. The bench is in place because human society long ago determined that important choices pertaining to rights, property, and liberty should be made with human conscience. The idea is still the same. The ease with which it can now be concealed whether it is being honored is what has changed.
Disclaimer
Nothing published on Creative Learning Guild — including news articles, legal news, lawsuit summaries, settlement guides, legal analysis, financial commentary, expert opinion, educational content, or any other material — constitutes legal advice, financial advice, investment advice, or professional counsel of any kind. All content on this website is provided strictly for informational, educational, and news reporting purposes only. Consult your legal or financial advisor before taking any step.
