The students this is for
A grade-9 student in Mosul opens an AI tutor before her chemistry exam. She asks about an element. The model gives her a precise, modern answer. She writes it down. The next morning she copies it onto her exam. Her teacher marks it wrong. The textbook said something different, and the textbook is what gets graded.
This happens every day across Iraq and Syria, in classrooms where the textbook is the curriculum. Not a resource for the curriculum. The curriculum itself. Teachers move through it page by page. Exams test it word for word. A student who learns from a better source and answers more accurately is penalized, not rewarded.
Drop ChatGPT, Claude, Gemini, or any Khan-Academy-style assistant into that classroom. They were trained to be helpful and modern. They paraphrase. They modernize. They quietly correct outdated material. They sometimes invent page numbers when asked about a textbook they have never seen. For a child preparing for an end-of-year exam in Baghdad or Aleppo, these well-meaning tools are actively harmful. They give answers her teacher will mark wrong.
The textbooks themselves are out of date. Some chapters teach biology or geography from the 1990s. Some use mathematical notation that has fallen out of use elsewhere. None of this gets fixed quickly. Curriculum reform is slow, contested, and deferred. Meanwhile a generation of students is taking exams scored against pages that the rest of the internet contradicts.
Why this is hard, in plain terms
The technical problem is the inverse of what most AI labs are working on. Frontier labs are pulling their models toward what the modern internet knows. Iraqi and Syrian students need models pulled toward what their specific ministry printed in their specific year, even when the two disagree.
Three pieces have to come together.
First, the model has to be faithful to the textbook even when its instinct is to modernize. Prompting it to "answer only from the book" works for short factual questions and breaks down on anything harder. Fine-tuning helps but does not fix it. A real solution combines retrieval, training, and decoding constraints. Right now no one has measured which combination wins, because no one has built the benchmark to measure it on.
Second, the textbooks have to get into the system. Iraqi and Syrian textbooks circulate as scanned PDFs. Arabic OCR is meaningfully worse than English OCR. Layouts are crowded, diagrams matter, religious and political content is sensitive and has to be preserved structurally. Turning these stacks of scans into a clean, navigable curriculum is a research-grade job, not a weekend script.
Third, students do not speak Modern Standard Arabic. They speak Iraqi or Syrian dialect, mixed with English and French loanwords for technical terms. Almost no model is tuned on classroom dialect specifically. A student asking "shlon ahsubha?" deserves a dialect answer that still uses the exact MSA terminology her exam will demand.
A system that handles only one of these produces fluent confidently-wrong tutoring, or curriculum-faithful answers no twelve-year-old in Mosul will read.
What needs to happen
Three concrete pieces of work, each useful on its own. A team of three could deliver them in nine months.
A benchmark. Five textbooks each from Iraq and Syria, in the high-stakes grades and subjects. Five hundred carefully built question-answer pairs that capture where the textbook and the modern web diverge. Released openly. Without it, every claim about curriculum-faithful AI is unfalsifiable.
A textbook ingestion pipeline. Open-source. Tuned for the specific OCR and layout failure modes of Iraqi and Syrian printing. Useful to every Arabic-language education project that comes after, whether or not it competes with mine.
An honest comparison. Run retrieval, fine-tuning, and constrained-decoding methods head-to-head on the benchmark. Publish the results, hybrids included. The field will gain more from one clean comparison than from another product launch.
The dialect interface comes last and sits on top of whichever method wins.
Why me
I am building ManarWay, a product for Iraqi students learning English. We see this problem every day, in messages from parents and teachers. We have working relationships with Iraqi schools and with teachers who can review benchmark items in their own subjects. The textbook, the dialect, and the classroom are not abstractions for me.
ManarWay alone is the wrong vehicle for this work, because a product has to ship and a benchmark has to be open. Done well, this proposal is upstream of every Arabic-language education tool that comes after, including the next ten ManarWays, including the ones that compete with us. That is the right shape for this prize, and the right shape for the region.
The schools want better tools. The ministries are not blocking. The money is small. What is missing is the boring upstream work: the benchmark, the pipeline, the honest comparison. Nobody has had a reason to do that yet. That is what this proposal funds.