Industry
Artificial Intelligence
Pakistan has no AI statute. The legal work is mapping the laws that do exist — PECA, copyright, contract, and sectoral rules — onto products the drafters never imagined.
Advising an AI company in Pakistan means practicing law by mapping. There is no AI statute to comply with, no regulator to seek comfort from, and — as of mid-2026 — no Pakistani court decision that squarely addresses training data, model liability, or machine authorship. What exists instead is a lattice of general law that was written for other technologies and applies anyway: the Prevention of Electronic Crimes Act 2016 for data access and content harms, the Copyright Ordinance 1962 for what goes into and comes out of a model, the Contract Act 1872 for who bears the loss when the system is wrong, and sectoral rules that reach AI through the regulated industries that buy it. The legal work is translation, and the translation has consequences, because the first authoritative answers will be given in enforcement proceedings against whoever gets there first.
The training-data question
Every model is an argument about its inputs. Pakistani copyright law offers AI builders less shelter than they assume: the Copyright Ordinance 1962 protects software, text, images, and compilations, its fair-dealing exceptions are narrow and were drafted decades before machine learning, and there is no text-and-data-mining carve-out of the kind some jurisdictions have legislated. Scraping adds a second layer — sections 3 and 4 of PECA criminalize unauthorized access to information systems and unauthorized copying of data, and while their application to public-web scraping is untested, an aggrieved data owner does not need certainty to file a complaint. Our approach is provenance discipline: grade every data source by how it was obtained, prefer licensed, consented, and public-domain material where the product allows, paper the internal decisions, and keep the record an acquirer's diligence team will eventually demand. A company cannot retrain its way out of a tainted foundation cheaply, so the sourcing decision is the risk decision. Personal data in training sets deserves separate attention: even without a general privacy statute, commitments made in a product's own policies are contractually enforceable against it, and Pakistan's pending data-protection legislation would apply to datasets assembled today and used for years.
Liability when the model is wrong
Pakistani law has no strict-liability regime for defective software, no AI-specific duty of care, and provincial consumer statutes of uneven reach — which means that when an AI system causes loss, the contract is usually the whole battlefield. That makes the liability clause in an AI agreement the most important commercial term in it, and the most commonly botched. Vendor templates disclaim everything, which regulated and enterprise customers reject and courts may read down. Customer templates demand accuracy warranties no probabilistic system can honestly give. The workable middle is specific: describe what the system does in statistical rather than absolute terms, define the human-review obligations that condition permitted use, cap liability at a level the vendor's insurance and balance sheet can actually bear, and carve out the risks — data breach, IP infringement in outputs, regulatory penalties — that need their own treatment. Where the customer is a bank or insurer, the State Bank's or SECP's expectations of their licensee flow down into the vendor contract, and an AI company that has not seen those terms before will lose weeks discovering them mid-negotiation.
Content harms and the toolmaker
The sharpest near-term exposure for generative AI in Pakistan is not copyright — it is PECA. Manipulated images, cloned voices, and impersonating text engage the Act's offences against the dignity and modesty of natural persons, and the 2025 amendments moved enforcement to the National Cyber Crime Investigation Agency and sharpened the state's content machinery. Complaints in this space routinely sweep in the platform that generated or hosted the content, not just the person who prompted it. The defensible position is operational and documentary: abuse-prevention controls proportionate to the product, terms of use that prohibit the conduct in Pakistani-law terms, a takedown channel that demonstrably functions, and logs that show it. We build those files with clients before the first complaint, because the file that exists on the day the notice arrives is the one that counts.
How we serve AI companies
Our AI work is counseling at the edge of settled law, done honestly: we tell clients what is established, what is arguable, and what nobody knows yet, and we paper the unknowns so the client holds the better side of the ambiguity. That means data-sourcing frameworks, model and output ownership terms, liability architectures negotiated against real counterparties, and enforcement-readiness for the PECA complaint that successful products eventually attract. The position stated here is as of mid-2026; if Pakistan legislates on AI or data protection — both plausible within the horizon of a product roadmap — the mapping changes, and we re-run it with each client rather than assuming last year's answer.
The Five Recurring Problems
The problems this sector keeps producing.
- 01
Building under a statute vacuum
As of mid-2026 no Pakistani law regulates AI as such; a national AI policy exists at the federal level but a policy is not law [STATUS — TO BE VERIFIED BY REVIEWING LAWYER]. The risk is not the absence of rules — it is that existing rules written for other things apply in unpredictable ways, and the first authoritative readings will come from enforcement, not guidance.
- 02
Training data acquired by scraping
The Copyright Ordinance 1962 contains no text-and-data-mining exception, and its fair-dealing provisions are narrow. Scraping can separately engage the unauthorized-access and unauthorized-copying offences in sections 3 and 4 of the Prevention of Electronic Crimes Act 2016, and breach the source site's terms. A model trained on unexamined data carries that exposure permanently.
- 03
Nobody agreed who pays when the model is wrong
An AI product that misclassifies a transaction, drafts a defective document, or gives bad advice creates loss that Pakistani law will allocate by contract if the contract speaks — and unpredictably if it does not. Most AI contracts we review either say nothing about output reliance or disclaim everything in terms a court may not honor. Both are drafting failures.
- 04
Ownership of outputs and fine-tuned models
Pakistani copyright doctrine presumes a human author, and whether purely machine-generated output attracts protection is untested here as of mid-2026. Fine-tuned models sit on base models whose license terms — including open-weight licenses with use restrictions — govern more of the company's freedom than founders realize. What the customer owns, what the vendor keeps, and what the base-model license permits must be settled in writing.
- 05
Misuse of the tool becoming the toolmaker's problem
Generative products can be used to produce impersonations, manipulated images, and harassment — conduct that engages the dignity and modesty offences of PECA 2016, enforced as of mid-2026 by the National Cyber Crime Investigation Agency. Complainants and investigators do not always distinguish the user from the platform, and the toolmaker's abuse controls and takedown records become its defense.
The Regulators That Matter
Who you answer to — and for what.
- National Cyber Crime Investigation Agency
- Investigates PECA 2016 offences following the 2025 amendments — the forum where complaints about AI-generated content, scraping, and unauthorized data use will actually land.
- PTA
- Holds content-blocking and removal powers under section 37 of PECA 2016; AI products that host or generate public-facing content live within its reach.
- State Bank of Pakistan and SECP
- Neither regulates AI directly, but each regulates its licensees' use of it — an AI vendor selling into banks or capital-market firms inherits outsourcing, model-governance, and audit expectations through its customers.
- Ministry of IT & Telecom
- The policy owner — source of the national AI policy and the likeliest sponsor of eventual legislation; worth watching, not yet a licensing body [TO BE VERIFIED].
Mapped Services
The practices this industry draws on.
- Artificial Intelligence The dedicated practice — product counseling where no statute gives the answer directly.
- Technology Law PECA exposure, electronic contracting, and the sectoral rules that reach AI products sideways.
- IP Protection Protecting models, weights, and datasets through the copyright, trade-secret, and contract tools that actually exist.
- Data Privacy Lawful data sourcing and subject-facing commitments built by contract while the data statute remains pending.
- Commercial Contracts Liability allocation, output disclaimers, and IP terms — the only place AI risk can currently be priced.
- Risk Advisory Abuse-control frameworks and enforcement-readiness for products that will attract complaints before they attract rules.
Questions, Answered
What clients in this industry ask.
Not as such, as of mid-2026. There is no AI act, no licensing regime, and no AI regulator. What exists is a national policy document and a set of general laws — PECA 2016, the Copyright Ordinance 1962, the Contract Act 1872, sectoral financial rules — that apply to AI products the way they apply to everything else. The absence of a statute is not the absence of risk.
Treat it as a risk decision, not a settled entitlement. Pakistani copyright law has no training exception, sections 3 and 4 of PECA criminalize unauthorized access to and copying of data in some circumstances, and source-site terms usually prohibit scraping. We help clients grade sources, prefer licensed and public-domain data where feasible, and document the provenance decisions that diligence will later ask about.
Pakistani law has not answered whether machine-generated output attracts copyright at all — the statute presumes a human author. In practice, ownership of outputs is allocated by the contract between vendor and customer, which is why that clause matters more here than in most jurisdictions. We draft it deliberately rather than leaving it to an untested default.
Through architecture, not just a disclaimer: define the product as decision support, require human review in the use terms, cap liability with a structure the Contract Act 1872 will bear, and exclude specific reliance scenarios expressly. A blanket everything-disclaimed clause is weaker than a specific, negotiated allocation — and regulated customers will not accept it anyway.
Yes, but you inherit their rulebook. State Bank-regulated buyers will impose outsourcing, data-location, audit, and model-governance terms drawn from their own supervisory expectations, and your contract and infrastructure must absorb them. We prepare AI vendors for that procurement before it starts, because retrofitting mid-negotiation costs the deal its timeline.
Expect the complaint to name the platform as well as the user. PECA's offences against dignity and modesty cover manipulated intimate and impersonating content, and enforcement as of mid-2026 runs through the NCCIA. Your defense is built in advance: abuse controls, terms that prohibit the conduct, a working takedown channel, and records showing all three operated.
Related Insights
Prepared by The First Counsel · As of 2026-07-12 · Pending professional review — statements flagged in the text are being verified
This publication is provided for general information only. It is not legal advice, and neither reading it nor corresponding with the firm about it creates a lawyer–client relationship. The position stated must be verified against current law before it is relied upon.