Saudi Arabia made AI a compulsory subject. What that asks of the technology underneath

Saudi Arabia's Ministry of Education has made artificial intelligence a compulsory school subject from the 2025-26 academic year, reaching more than six million pupils — one of the largest deployments of technology education attempted anywhere. For a school group, a publisher or a ministry team, the syllabus is the easy part. The hard part is that a compulsory subject has to be delivered in Arabic first rather than in English with a translation layer over it, assessed in a subject with no past papers and no item statistics, taught by teachers who learned the material weeks ahead of their pupils, and recorded in systems that Gulf regulators ask specific questions about.

None of those four is unusual on its own. What makes them hard together is the deadline. A pilot picks its schools; a mandate takes all of them, on the same morning.

A mandate is a different problem from a pilot

Pilots select for enthusiasm: a teacher who already reads about the subject at weekends, a working computer room, a head who wants the visit. A mandate also takes the school with one over-subscribed lab and a teacher covering the new subject alongside two others. Plan for the median school, not the flagship — shared devices, thin bandwidth at 8am when every class logs in at once, and sessions that survive a dropped connection without losing a pupil's work.

The second change is bureaucratic and easy to underestimate. A compulsory subject has to be evidenced upward — to a group office, a directorate, a ministry — not merely observed in the classroom. That obligation lands on whatever system holds the record, and it lands in the first term.

• Every school is in scope on the same date, including the ones that would never have volunteered.
• Coverage has to be reported upward, so "we taught it" needs to be a query someone can run, not a conversation.
• Pupils who move mid-year carry a record in the new subject with them, and the receiving school has to be able to read it.
• Guardians ask what the subject is and what their child is doing in it, often before the school has finished deciding.
• The first cohort's results become the baseline every later cohort is judged against, whether or not that baseline is any good.

Arabic first, not Arabic afterwards

This subject punishes a translation layer harder than most. AI teaching material is full of left-to-right fragments — Python snippets, library names, URLs, notation — embedded in Arabic prose. Bidirectional rendering is where a platform fails visibly: brackets that reverse, punctuation that jumps to the wrong end of a line, a code block that reads correctly in the editor and wrongly in the printed worksheet. Ask to see a lesson with a code sample, in Arabic, on a phone. It takes a minute and settles the question.

Then check the side nobody demos. The learner interface is usually the translated one; the coordinator's screens, the grading queue and the exported PDF are where mirrored layouts get missed. A coordinator who switches to English to build a coverage report is doing the work twice, and will end up keeping the real record in a spreadsheet.

Content sourcing deserves its own decision. There is far more AI teaching material in English than in Arabic, so most groups will translate or generate. Machine translation across a whole textbook produces terminology drift — the same concept named three ways in three chapters, which is fatal in a subject where the vocabulary is the learning. Fix a glossary in Modern Standard Arabic first, agreed with the teachers who will use it, then hold every unit to it and have a subject teacher review the result, not a translator.

Assessing a subject nobody has taught before

Everything a mature subject leans on is missing in year one. No past papers, no anchor scripts, no cut scores, no item difficulty statistics, no examiner with fifteen years of moderation behind them. So the first year's assessment carries two jobs at once: report honestly on this cohort, and build the evidence base that makes the second year defensible.

Write outcomes as observable competencies before writing any questions. "Can explain why a model trained on unbalanced data misclassifies the smaller group" can be moderated across forty schools; "understands machine learning" cannot. In a new subject the outcome statement is the only fixed point you have, which is why competency frameworks earn their keep here more than in a settled one.

Use rubrics deliberately rather than by habit. Analytic when you want a diagnostic profile of where a cohort is weak — and in year one you badly want that. Holistic when you need defensible marks quickly at scale. Single-point when the target is a published standard. Then keep the item statistics from the first sitting: difficulty and discrimination per item, so the question everybody answered correctly and the question nobody did can be repaired before they are reused. Ordinary psychometrics matter more in year one than in year ten. The machinery for all of it — rubrics, moderation, item statistics — is set out on the assessments page.

One problem is specific to this subject: pupils will use a model to do the homework about models. Design around it rather than policing it — in-class performance tasks, portfolios that keep the drafts as well as the final piece, short oral defences of a project. Using the tool well is part of what you are teaching, so assess the judgement rather than the typing.

Teacher capability is the constraint that binds

Do the arithmetic before buying anything: teachers to be trained, times hours each, divided by the weeks before term, against the people who can actually deliver the training. In most groups the number does not work, and the usual answer is cascade training — train the trainers, who train the trainers, who train the teachers. Fidelity drops at every hop. A fourth-hand lesson on neural networks is not a lesson on neural networks.

What survives the cascade better is material a teacher can hold: the same unit the pupil gets, worked answers, and somewhere to ask a question and get one back from someone who knows. Give teachers the same platform as the learners, so there is one record and one set of content rather than a training portal beside a learning portal with separate logins and no shared reporting.

Measure completion, not invitation. Attendance is not evidence that anybody can teach the material; a short assessment and a certificate naming what was demonstrated is closer. And build one channel back — which lessons collapse in a real classroom with thirty pupils and twelve working machines. That feedback is worth more than any inspection in year one, and it has a short shelf life: teachers stop reporting problems once they decide nobody reads the reports. A group running many campuses has more of this to coordinate than a single school does; the schools page covers the group-level version.

Data protection questions get asked differently in the Gulf

Saudi Arabia's Personal Data Protection Law, administered by SDAIA, sets conditions on how personal data is handled and on transferring it outside the Kingdom. Procurement here opens with residency, transfer and sub-processors, not with a feature list. Teaching the subject with AI tools adds a question a geography lesson never raised: what happens to a pupil's work when it is sent to a model, and whose model is it.

Get these answered in writing, before signature, from every vendor including us.

• Where is pupil data stored, in which country, and what is the notice period if that answer changes?
• Which sub-processors touch it — naming any model provider behind an AI feature — and where does each of them run?
• Is pupil work used to train anyone's model? "No" belongs in the contract, not in a blog post.
• How long is data kept after a pupil leaves, who can delete it, and is deletion demonstrable?
• What can a guardian see, and can that be scoped by category rather than all-or-nothing?
• Can you produce a record of who read or changed a pupil's data, and can that record be shown to be unaltered?
• Which of the vendor's own staff can see pupil data, under what approval, and is that approval logged?

"We are GDPR compliant" is not an answer to a PDPL question. It usually means the vendor has done the engineering that makes a PDPL answer possible — lawful-basis handling, consent records, subject access, erasure — but the mapping is your legal team's to check, not the sales engineer's to assert.

What a platform can and cannot take off you

No platform teaches this subject. What it removes is the duplicated work that eats a rollout: authoring the same unit once per school instead of once, chasing which sites have covered which outcome, exporting spreadsheets to answer one question from the directorate, and running one system for pupils and another for the teachers being trained to teach them.

That is the shape of platform we build, so read what follows as interested rather than neutral. Lurno ships in four product languages — English, Arabic with right-to-left layout throughout, French and Italian — and the Arabic covers the administrative screens, not only the learner ones. On the assessment side: 19 question types, 7 assessment types, rubric grading in analytic, holistic and single-point forms, peer review, portfolios, a unified grading queue, and psychometrics including adaptive IRT/CAT for when an item bank is mature enough to justify it. Competency frameworks carry evidence and automatic mastery, which is what turns "we taught it" into something a directorate can inspect. Tenant isolation is enforced in the database rather than the application — separation enforced underneath the platform — and sensitive changes are written to a hash-chained audit log.

And plainly, the parts that are not there. SCORM 1.2, SCORM 2004, xAPI and LTI 1.3 are in development — modelled in the product, runtime still being built — so if a tender requires standards conformance on day one, say so at the first meeting. SAML/OIDC SSO, MFA and passkeys are on the roadmap; what exists today is partner sign-on (silent SSO), a different thing that should not be sold as the same one. Payments, checkout and self-serve signup are not built: accounts are created by invitation. We hold no SOC 2 or ISO 27001 certificate; the practices are aligned to ISO 27001 and the security page says which.

The groups that will look competent in three years are the ones treating this first year as calibration: outcomes written before questions, item statistics kept from the first sitting, teacher capability measured rather than assumed, Arabic tested on the screens nobody demos, and the data questions answered in the contract while the contract is still open. The syllabus will be revised — new subjects always are. The record of what was taught, to whom, and how well it was learned is what you will still need when it is.