Journal · Islamabad Capital Territory

Proper Use of AI in Pakistani Courts: Rules, Limits, and Professional Duties

Data & Privacy 3 July 2026

This entry states the current standing position as of mid-2026. Law, guidelines, and institutional practice in this area are developing quickly, so treat anything below as accurate to that point in time and check for later changes before relying on it.

The question of how far artificial intelligence should reach into judicial work has moved quickly in Pakistan, from an isolated courtroom experiment in 2023 to a settled institutional position by 2026. In that span the higher judiciary has fixed a working boundary that most public commentary still describes only in pieces. AI can support the people who run the justice system, and it cannot take the place of the human reasoning the law requires. This entry sets out what that boundary means in practice, where it comes from in the Constitution and the procedural codes, and what duties now attach to judges, advocates, and court administration when these tools are used.

Quick Answer

In Pakistan, AI is currently treated as an auxiliary tool for the courts and not as a decision maker. The Supreme Court set the direction in Ishfaq Ahmed v Mushtaq Ahmed (2025), where it accepted AI for tasks such as legal research, drafting and language work, translation, and case management, while ruling it out for deciding cases, assessing evidence, or writing the reasoning of a judgment. The National Judicial (Policy Making) Committee then issued the National Guidelines for the Use of Artificial Intelligence in Judicial Institutions, which require a human in the loop, meaning a judge records independent reasoning at every point where AI has assisted. The proper use of AI in Pakistani courts rests on three points that run through all of the official material: human judgment stays with the human, any output must be independently verified, and confidential case material must be protected, which matters all the more while Pakistan has no fully enacted data protection statute.

From a Single Experiment to a National Position

The public conversation in Pakistan is often traced to one event in 2023, when a lower court referred to a generative chatbot while considering a bail matter. That episode was widely reported, and much of what still circulates online repeats it as if it were the whole story. It was not a framework. It was an early and informal use of a general consumer tool, and the tool itself flagged its own limits, noting that it could not access locally relevant case law and that its knowledge had a fixed cut off. The lasting value of that moment was not the outcome of the bail application but the debate it opened about whether a system that produces fluent text has any proper place near the exercise of judicial power.

That debate matured fast. Within two years the Supreme Court had given the subject a considered treatment, and within three years the judiciary had a formal set of guidelines aligned with a national policy. The arc changes the nature of the question. The issue is no longer whether AI will enter Pakistani courtrooms, since it already has, both informally and through official tooling. The issue is the discipline under which it operates, and that discipline is now written down rather than improvised.

What the Supreme Court Held in Ishfaq Ahmed v Mushtaq Ahmed

The leading authority is Ishfaq Ahmed v Mushtaq Ahmed, decided in 2025 and reported at 2025 SCP 112, also carried as PLD 2025 SC 582. The underlying dispute was an ordinary civil matter over property near Lahore that had moved through the lower forum and the Lahore High Court before reaching the Supreme Court. A two member bench heard it in March 2025 and resolved the property question in the usual way.

What made the judgment notable was its treatment of technology. The Court observed that the matter had taken years to reach finality, and it linked that kind of delay to the fair trial guarantee in Article 10A, treating timely justice as part of fairness rather than as a matter of convenience. On that basis it accepted AI as a tool that could help reduce delay, while stressing a measured and supervised approach rather than either rejection or open reliance.

The Court then separated what is acceptable from what is not. On the acceptable side it pointed to legal research, precision in language and drafting, and comparative study of case law, along with administrative uses such as assigning jurisdiction and allocating cases in ways that reduce arbitrary discretion. On the other side it was equally clear that AI is not to be used to record evidence, to write judgments, or to stand in for human adjudication. It set out the reasons that make the second list necessary, including fabricated output, bias, opacity in how results are produced, exposure of private data, and the risk to public confidence if people believe a machine decided their case. Recognising that a single judgment can set direction but cannot build a system, the Court asked the National Judicial (Policy Making) Committee and the Law and Justice Commission of Pakistan to prepare detailed guidelines on permissible uses.

Why the Constitution and the Codes Set a Hard Floor

The limit on AI decision making in Pakistan is not only a matter of preference. It rests on statutory and constitutional ground, which is what gives it force. Article 260 of the Constitution defines a Judge in terms of a person acting as a judge of a court, and Section 20 of the Pakistan Penal Code defines a Judge as a person empowered by law to give a definitive judgment. A court of justice, in the same scheme, means a judge or body of judges empowered to act judicially. These are human offices held by people who can be held to account.

The Code of Criminal Procedure completes the point on how decisions are recorded. Section 367 requires that a judgment be written by the presiding officer of the court, or from that officer's dictation, and be signed by that officer. Read together, these provisions mean that a valid judgment has to carry human reasoning attributable to an identifiable judicial officer. A system that produced a decision on its own would meet neither the definition of a judge nor the requirement of a signed and reasoned judgment. This is why the position on AI written reasoning works as a rule rather than as advice. It reflects what the existing law already demands.

There is a second dimension that shapes how far administrative uses can go. The Supreme Court and the High Courts hold their own power to make rules for practice and procedure, covering case allocation, formation of benches, service of notices, cause lists, the pronouncement of judgments, and record keeping. For AI to be built into these functions in a lasting way, the relevant court rules would need to be amended so that the practice has a formal basis rather than resting on individual habit. The judgment effectively invited the superior courts to use that rule making power to give any integration a proper footing.

The National Guidelines and the Human in the Loop

The guidelines the Supreme Court asked for were prepared through the National Judicial Automation Committee and finalised as the National Guidelines for the Use of Artificial Intelligence in Judicial Institutions of Pakistan. The National Judicial (Policy Making) Committee announced them in 2026, aligned with the National AI Policy 2025 and framed to preserve judicial independence. Their reach is broad. They are meant to be followed across the superior courts, the courts subordinate to them, tribunals and special courts, and other justice sector bodies, so that the standard is consistent from the apex court down to the district judiciary.

The defining feature is the human oversight requirement, expressed as a human in the loop. AI is to assist and support judicial functions and is not to replace the final decision making authority of judges. The part that summaries often miss is the mechanism attached to that principle. Every AI supported process has to include points of human validation at which the judge sets down independent reasoning, so that assisted work never quietly becomes the operative basis of a decision. The guidelines also deal with competence, providing that judges and staff receive training so they understand both what the tools can do and where they fail.

On permitted uses, the guidelines track the direction the Court set. Administrative and support functions sit at the centre, including case management, scheduling, routing of documents, translation, and transcription of proceedings and minutes. Research and drafting support are allowed within the same discipline of verification and human authorship. The line against AI acting as an adjudicator, weighing evidence, or fixing sentence stays firm.

The National AI Policy 2025 as Background

The judicial guidelines do not sit in isolation. The federal cabinet approved the National AI Policy 2025 in July 2025, and it serves as the wider national frame within which the judiciary's rules are meant to operate. The policy is built on six pillars covering an innovation ecosystem, awareness and readiness, a secure AI ecosystem, sectoral transformation, national infrastructure, and international collaboration. It sets up supporting structures such as a national AI fund, an AI council, centres of excellence, and regulatory sandboxes for supervised testing. For the courts the policy matters less for its economic aims and more for its insistence on responsible use, which the judicial guidelines then translate into the setting of adjudication.

Institutional Tooling: Judge-GPT and the Federal Judicial Academy

A part of the Pakistani picture that gets little attention is that judicial AI is not only about individual judges consulting public chatbots. The Federal Judicial Academy in Islamabad has deployed a purpose built tool, Judge-GPT, trained on Pakistan's statutes, procedure, and case law and used by more than 1,500 judges in the district judiciary to speed up research and improve the drafting of orders, with a standing instruction that every reference it suggests must be verified.

This detail sharpens the analysis. A tool built on Pakistani material answers part of the objection raised in 2023, which was that a general model has no reliable grasp of local law. It does not remove the duty to verify, and it does not move authorship away from the judge. It does show that the workable path in Pakistan is being built around bounded, jurisdiction aware systems with a person accountable for the output, rather than around open ended reliance on consumer tools.

The Confidentiality Problem Pakistan Has Not Fully Solved

Sound use of AI in litigation turns on what happens to the information placed into a tool, and here Pakistan carries a structural gap. The country has not enacted a comprehensive data protection statute. The Prevention of Electronic Crimes Act 2016, amended in 2025, addresses certain misuse of information, but it is not a general data protection law. The Personal Data Protection Bill, developed by the Ministry of Information Technology and Telecommunication, has not completed the legislative process, and the proposed National Commission for Personal Data Protection is not yet an operating regulator.

The practical effect is direct. When client material, case files, or draft pleadings are entered into a public AI service, that information can be processed and stored on external infrastructure, often outside Pakistan, without a domestic statutory regime governing how it is held, shared, or reused. For a lawyer this touches the duty of confidentiality and the protection of privileged material, both of which stand on their own regardless of whether a data protection law is in force. The absence of a general statute does not lower the obligation. It raises the value of choosing tools with care, limiting the sensitive detail ever exposed to a third party system, and favouring platforms that keep legal material within a controlled boundary.

Fabricated Output and the Duty to Verify

The most concrete risk for practitioners is not abstract. Generative systems can produce confident text that is simply wrong, including citations to cases that do not exist, quotations that were never written, and summaries that misstate what a real judgment held. Courts in several countries have already treated this as a professional failing rather than a technical quirk, and lawyers elsewhere have faced formal consequences after filing briefs that relied on invented, AI generated citations.

The point carries directly to Pakistan. A lawyer who cites an authority is representing to the court that it exists and says what it is claimed to say, and that representation is not softened by the fact that a machine supplied it. The duty of competence and the duty of candour to the court both require that any AI assisted research be checked against the actual text of the statute or judgment before it is relied on. The working method is simple to state: treat AI output as a starting point for research and not as the citation of record, read the underlying document in full, and confirm that it genuinely supports the point before it goes into a pleading or an argument.

Duties by Role

The official material distributes responsibility differently across the people involved in a matter. The table below sets out how the same principles apply to each role.

Role Permitted assistance Core duty
Judges Legal research, drafting language, translation, comparative reference, case and cause list management Record independent reasoning at each validation point, remain personally responsible for the decision, and keep adjudication, evidence assessment, and sentencing out of the tool
Advocates and litigants Research support, first draft language, summarising material, organising case documents Verify every citation and factual claim against the source, protect client confidentiality and privilege, and avoid presenting AI output to the court as confirmed authority
Court administration Scheduling, document routing, transcription, translation, case allocation to reduce arbitrary discretion Keep systems auditable and transparent, provide training on capabilities and limits, and ensure administrative automation does not quietly influence outcomes

Where This Sits Against International Practice

Pakistan's position lines up with the direction taken elsewhere, which is worth noting because it suggests the local framework is not an outlier likely to be reversed. UNESCO published guidelines for the use of AI systems in courts and tribunals in 2025 that treat AI as support for managerial and documentary tasks under human supervision and reject it as a replacement for judicial judgment. Judicial guidance in the United Kingdom places personal responsibility on the office holder for anything produced in their name, expects judges to read the underlying documents themselves, and cautions against entering private information into public chatbots. The common thread across these instruments and the Pakistani material is the same measured acceptance: the technology is welcomed as a genuine aid and refused as a substitute for human judgment.

Practical Takeaways for Pakistani Practice

For a lawyer working in Pakistan today, the proper use of AI comes down to a few habits that follow from the law and the guidelines. AI fits the preparatory parts of the work, such as first pass research, summarising long records, and tightening drafts, while the analysis stays with the person doing the work. Anything that will go before a court is checked against the primary source, since responsibility for a false citation rests with whoever files it. Confidential and privileged material is best kept out of public tools while Pakistan has no enforceable data protection regime. Tools bounded to the relevant jurisdiction and legal corpus tend to serve better than open systems drawing from an undifferentiated mix of sources, because a narrower and better grounded tool reduces fabricated output at the root. Platforms built for a defined jurisdiction, such as Legalise for Islamabad practice, follow this logic by keeping research and drafting tied to encoded legal material rather than open web guessing, which is the discipline the courts are pointing toward.

Frequently Asked Questions

Can a judge in Pakistan use AI?

Yes, within limits. In Ishfaq Ahmed v Mushtaq Ahmed the Supreme Court accepted AI as an auxiliary tool for work such as legal research, drafting, translation, and case management, and the National Guidelines confirm this. What is not allowed is using AI to decide a case, assess evidence, or write the reasoning of a judgment, because the Constitution and the procedural codes require a human judge to author and sign a reasoned decision.

Can a lawyer rely on AI generated case citations in a filing?

Not without independent verification. Generative tools can fabricate citations and misstate holdings. The duty of competence and the duty of candour to the court require that every authority be checked against the actual statute or judgment before it is cited.

What do the National Guidelines require in one line?

They require a human in the loop, meaning AI may assist judicial and administrative work while a judge records independent reasoning at set validation points, and users are trained to understand the tools' limits.

Does Pakistan have a data protection law that governs AI use of case data?

Not a comprehensive one yet. The Prevention of Electronic Crimes Act 2016, amended in 2025, covers some misuse of information, but the Personal Data Protection Bill has not completed the legislative process. Until it does, confidentiality of client and case material depends on professional duty and careful tool selection rather than a general statute.

Sources and Further Reading

← Back to The Digest