The real risks of AI in Pakistani schools, and how to manage each one

Most writing about AI in education is either a sales pitch or a scare story. Neither helps a principal deciding whether staff should use it on Monday. A more useful document is a risk register: a short list of what can actually go wrong, how likely each item is in a Pakistani school specifically, and what a careful school does about it. We build AI tools for teachers, so we have an obvious interest in this topic, and that is exactly why the honest version is worth writing down.

Risk one is factual error. AI models produce fluent text whether or not it's right, so a lesson plan can contain a wrong date, a bad definition, or an example that doesn't match your textbook. For Urdu, the evidence suggests extra care. Researchers behind the UrduFactBench and UrduFactQA benchmarks point out that existing fact-checking systems are mostly developed for English, and their evaluation of twelve language models found that open-source models face persistent challenges with factual reliability in Urdu. That doesn't mean Urdu output is unusable. It means Urdu output needs a reader who knows the subject, and a school should say so in writing.

Risk two is bias and cultural mismatch. UNESCO's guidance notes that current AI models are trained on data from online users that reflect the values and dominant social norms of the Global North. In a classroom that can show up as examples about snow, supermarkets, or holidays your students have never experienced, or assumptions about family and school life that don't fit. The fix is not to avoid the tool. It's to prefer tools scoped to your board and context, and to swap in local examples when the draft doesn't fit.

Risk three is the digital divide. DataReportal's Digital 2026 report for Pakistan puts internet penetration at 45.6 percent at the end of 2025, with roughly 139 million people still offline. If a school's AI strategy assumes every teacher and family has reliable connectivity and a decent device, it will widen gaps between well-resourced and under-resourced classrooms. A responsible plan starts with teacher-side tools that produce printable output, since a worksheet generated once can serve a class without any student needing a device.

Risk four is over-reliance. The danger isn't that a teacher uses AI to draft a plan. It's that, over a term, the draft becomes the plan, and nobody checks it against the learning outcomes or the class in front of them. Deskilling is quiet and cumulative. The remedy is a norm rather than a technology: every AI draft gets reviewed and edited by the teacher, and the teacher can say why each activity is there. A useful test is to ask whether the teacher could explain the plan without the screen.

Risk five is privacy. A teacher who pastes a student's name, marks, or family circumstances into a general tool has moved a child's information outside the school without anyone deciding to. Pakistan doesn't yet have an enacted general data protection law, so the school's own rules carry more weight than usual. We cover the vendor questions and the staff rule in detail in our post on AI data privacy.

Risk six is academic integrity, meaning students using AI to do work they should do themselves. It is a real issue, and schools often respond by buying detection software. The evidence for detectors is weaker than the marketing suggests, which we look at in a separate post on preventing misuse. For a teacher-facing tool like a planner or worksheet builder, this risk sits mostly with how assessments are designed, not with the generator.

Where our own product stands on this list, plainly: Muallim can produce a wrong fact, a stiff Urdu sentence, or an example that doesn't fit, like any AI tool can. That's why the Lesson Planner, the Worksheet Builder, and Quiz & Assessment open everything in an editor before export and never print, send, or publish anything for you. Scoping generation to a board reduces mismatch but doesn't eliminate it. Muallim is teacher- and admin-facing, so students don't have accounts, which removes one category of risk but not the need for a school policy. We're built by DIGIT Pakistan, we're a beta, and we'd rather say that than imply the risks are solved.

A practical way to use this register: put it on one page, add a column for how likely each risk is in your school and who owns the response, and review it when you adopt any new tool. Most of the mitigations are habits, not purchases: review before use, no student data in general tools, local examples where drafts don't fit, and a named person to report problems to. Schools that write this down tend to find AI less frightening and more manageable, which is the goal. Revisit the page after one full term of real use, cross out the risks that never showed up, and add the ones that did, because your own experience is better evidence than any general list.