What AI does to law, and the courts too slow to notice

A book of prompts that sold itself

The episode opens on something small that got big quickly. Adam Jabbar, a technology lawyer in Islamabad, spent one evening compiling the ChatGPT prompts he already used in his own practice — “review this contract for me”, “summarize this law for me” — into a list of 101 and published it on Amazon as a Kindle book, ChatGPT for Lawyers. He had self-published two books before, on navigating international privacy laws and the legal implications of AI, and a handful of people had bought each. This time he woke up the next morning to roughly 151 copies sold.

Then US legal-tech blogs picked it up. UK buyers followed. The news travelled back to Pakistan, the Tribune ran a story, and bar councils approached him — he gave them free copies on the principle that education should be free. By the time of the recording it was still moving thirty to forty copies a day. His framing for what the book is stays deliberately plain: a prompt is just language, the way telling a waiter “bring me a glass of water” is a prompt, and a better prompt gets a better output. The book exists to make a lawyer’s day shorter.

How AI eats the unglamorous half of legal work first

Muzamil sets up the larger question with an observation that recurs across the conversation: before generative AI, the consensus was that creative industries would be disrupted last. Instead they were disrupted first. The same surprise, Adam argues, is now reaching law — and it follows a pattern the profession has seen before.

He walks it back a decade. Lawyers used to do case-law research by hand, hunting precedents through libraries that might hold ten thousand books, with junior staff whose whole job was the search. Then a website called Pakistan Law Site put 90% of judicial decisions into a searchable database for a few thousand rupees a month, and that research could be done from a laptop at home. The library staffer’s job quietly disappeared. AI is the next step in the same line. Feed a model enough law — journals, case law, precedent — and it reasons over that material the way the previous tools indexed it, only faster.

Adam is candid about the limits. OpenAI trained heavily on US legal data and lightly on Pakistani material, so accuracy is uneven: he puts legal research at around 80% reliable, with roughly 40% of outputs carrying errors. But the direction is clear. The work that goes first is the work nobody romanticised — research, e-discovery, and the drafts that four associates used to produce. A petition or a legal notice that took a junior most of a day now takes minutes.

Whether a backlog should be solved by removing the humans

Muzamil reframes the problem from a computer scientist’s chair. Pakistan’s justice system has a bottleneck: the world’s fifth-largest population, cases in the thousands, too few judges, outdated civil and criminal procedure codes, and registrar offices with rooms full of files piled to the ceiling. When humans are the constraint, human problems follow — strikes, adjournments, “tareekh pe tareekh”, date after date. If a lawyer takes a simple recovery suit, Adam admits, the honest answer is that there is no mechanism in Pakistan to resolve it in a week.

So Muzamil pushes the model to its edge. Break a trial into inputs, processing and output. A plaintiff files, a defendant answers, a judge writes a decision from inputs and past history. Why not file online, let both sides submit their arguments, and have a system that has “heard” a hundred thousand cases return a judgment in a minute? A hundred years out, he asks, are session-court judges even required?

Adam’s answer is careful. The first move is not AI — it is digitisation. He credits the Peshawar High Court’s stance that filings should be submitted as soft copies, so a digital database of Pakistani cases can exist at all. Only once that data is in can a model learn what cases run, what decisions look like, and help judges triage and expedite. But he draws a firm line on what a judge does: weigh witnesses, facts, the arguments of both parties, and binding Supreme Court precedent into a neutral human decision. AI cannot replace that today, and a unique case it was never trained on is precisely where it would stumble. On the hundred-year question he refuses false certainty — given how fast the last few years moved, it could happen, or it could not. Use it as a tool, he says, and do not over-rely on it.

Bias, ownership, and the question of who is responsible

The conversation turns to the harder problem: regulation. The EU, Adam notes, is the body working aggressively to regulate AI, and the reasons cluster. Bias first. His example is uncomfortable and accurate — a hiring model taught by a prejudiced employer to filter out CVs of a particular race will produce discrimination at scale. Muzamil’s point sharpens it: AI is biased because it is a representation of biased humans, so it will stop being biased only when people do. How you regulate that, Adam admits, is a grey area nobody has solved.

Then liability. They argue it out properly. Adam starts from vicarious liability — Nestlé is responsible if its bottled water is bad, an employer is responsible for an employee’s workplace conduct — so the developer, OpenAI, should be responsible. Muzamil counters with the difference between a product and an intellectual service: a university isn’t liable for what its graduate posts online, you sue the New York Times but not Wikipedia, and Facebook escapes liability by calling its content user-generated. AI models, framed as teachers offering knowledge the user is free to misuse, disclaim everything — and they have. ChatGPT tells you to check legal advice with a licensed lawyer; OpenAI’s terms shed responsibility for what it produces. Push that logic to its end and the user is solely responsible, which Adam concedes is largely where the law lands, helped along by limitation-of-liability clauses nobody reads.

Ownership splits three ways: whoever’s work the model learned from, the platform, and whoever wrote the prompt. The platforms say the prompt-writer owns the output, with a partial license retained for themselves. Adam’s moral instinct points the other way — at the person being replicated. AI music in Atif Aslam’s voice, he argues, is like a fake Louis Vuitton bag: ethically the artist should still own it, even if the terms and conditions say otherwise. Muzamil presses on the boundary. A student inspired by Sadequain doesn’t owe Sadequain royalties; inspiration is how culture works. The two converge on a line: training on publicly available work and being broadly influenced is one thing, but the moment a product uses your name to imitate you, intellectual-property rights are torn up. Adam’s modest ask is that AI-generated content at least credit the source.

A data-privacy law written by copy-paste

The third strand is data privacy, and here Muzamil is honest about being in the grey. He believes regulation is necessary — without a Pakistani entity to hold responsible, crimes go unaddressed, the FIA can’t act, platforms sit in Singapore. He has limited patience for the Asian Internet Coalition’s objection, which he reads as a lobbying stand for very large tech firms threatening to leave rather than build data centres or comply. Build the regulation and they’ll come, he argues.

But the criticism he does agree with is about how the law was made. The Personal Data Protection Bill sat in the Ministry of IT for three or four years while Adam — who advocated for it and sent the ministry notices — watched data breaches, hacking, financial fraud and ATM skimming go unaddressed by any specific law. When the bill finally moved through the federal cabinet, he read the draft and concluded it had been written without thought: no industry input, no input from professionals, what he describes as a control-C, control-V of GDPR and CCPA.

His worry is concrete. Penalties in the millions land on companies like Rozee, PakWheels and Zameen, who are only now learning the rules apply. And the founder he keeps returning to is the graduate launching an e-commerce store with no awareness of how data protection works — who could lose years and a business to a penalty, and whose response on the way out is simple: I don’t want to live here anymore. In Muzamil’s framing, the country tends to get the theory right and the implementation wrong, handing a hard-line law to a bureaucrat who can arbitrarily shut one platform and not another, with data privacy as the cover for much else.

The limits of regulating an open world

The conversation widens to where the internet itself is going. Muzamil tells the story of the dot-com bubble — pumped through the nineties, peaked in early 2001, collapsed, then a roughly twenty-year cycle in which the internet matured into something a whole generation was born into rather than handed. That maturity is why governments worldwide — GDPR, Canada, Australia, and now India and Pakistan — are finally regulating the powerful platforms, with the top companies in the world almost all being tech. As web2 gets regulated and freedom of expression dips, he sees web3 emerging as the alternative: decentralized social media like Mastodon, decentralized video, data owned by users, no central entity to hold liable.

Adam is upfront that he hasn’t studied web3 deeply and is learning like everyone else. On blockchain he sees genuine good — smart contracts that document ownership transfers and transactions far better than the conventional, poorly documented world, which is exactly why traditional banks will resist it loudest. On the harder question — who does a state regulate when liability is distributed across every user — his answer is that states will have to innovate, think out of the box, and write new laws for a global, connected marketplace.

But his sharper point is that not everything can be regulated, and he proves it from recent history. Pakistani freelancers have moved their dollars through Payoneer for years; when the state and FIA made noise about regulating and taxing them, the debate died and the money — hundreds of millions — kept flowing, unregulated. Crypto sits in the same grey, with the State Bank still “working on it” and the FIA freezing accounts on instinct. The real fix, he argues, is fresh, young input: an FIA cyber-crime wing staffed by people who actually understand NFTs, crypto, web3 and the dark web, because the country has barely five or six technology lawyers and almost no one — investigating officers, judges, lawyers — who understands the field. His prescription for every tech and cyber law is the same: make it balanced, not very hard and not very soft, and let an open world live openly, because there is a real demand for freedom right now and you cannot run a democracy like a jail.

Pakistan in 2050, without the ifs and buts

Muzamil closes with the question he puts to most of his Pakistani guests — Pakistan in 2050, twenty-seven years out, the year Adam will reach retirement age — and asks him to skip the standard line about the youth bulge. Adam answers with positive hope and refuses the ifs and buts. He quotes the Quran’s promise that ease follows every hardship, and argues that no country’s hardship lasts a lifetime.

He breaks his own rule about the young population only to make a different point with it: this generation has technology, awareness and exposure their parents and ancestors never had, and the gap in thinking is enormous. Culture has changed, doing business has changed; some things are getting better and some worse, and it will take time. His closing ask is the one that runs through the whole conversation — take fresh input, bring all the stakeholders in, don’t work alone, and let the lawmakers soften and stay flexible. The most useful thing the technology lawyer says about the future is not about the technology at all. It’s that the people writing the rules need to start by listening.