Education Equity in the Age of AI

Artificial intelligence is transforming education. But the question that matters is not whether it will — it is for whom.

Every generation of learning technology has arrived with a promise of access and then sorted itself by the same criterion it was supposed to remove. Radio promised universal schooling and became a supplement for those already in classrooms. Television promised the same and became a tutor for children with parents who could supervise. The web promised the library at everyone's fingertips and then made the ability to read English the price of admission.

AI is at that fork now. If the tools are built for English first and localised later — or never — they will deepen the divide, handing English speakers a compounding advantage while everyone else gets a translated summary of it. If they are built for the reader's own language from the start, they could do something no previous technology managed: deliver the world's best explanations to people who never had to learn a second language to receive them.

The risk is structural, not malicious

It is worth being precise about how this happens, because almost nobody chooses it.

Model quality tracks training data. Training data tracks who publishes online. Publishing tracks who has the infrastructure, the audience and the commercial incentive. The result is that a question asked in Marathi may be answered by a model that has read a hundred thousand English pages on the topic and four hundred Marathi ones — and the gap is not in the model's intelligence, it is in what it has to work with.

That asymmetry compounds in a second, subtler way: evaluation. Benchmarks are written in English, by English-speaking researchers, about problems English-speaking users have. A model can score well while being materially worse for a Tamil speaker, because nobody measured the Tamil case. What is not measured does not get fixed.

The third mechanism is the one with the shortest fuse. When AI tutors get deployed into schools, they are procured by administrators, configured by vendors and priced by the seat. A tool that works only in English is not a neutral purchase — it converts every non-English-speaking student into an edge case the teacher has to work around manually.

The opportunity is real and it is specific

The counter-argument is not just optimism. There is a genuine structural difference between AI and every previous learning technology: the marginal cost of an explanation in the ninety-ninth language is close to zero.

Printing a textbook in Tamil and in Telugu required two production runs, two editorial teams, two distribution deals. Generating a well-grounded explanation in Telugu costs approximately the same as generating it in English — once the system is built to do it. That is a genuinely new economic fact, and it is the first time the incentive structure has pointed toward inclusion rather than away from it.

What that makes possible is not "translation". Translation gives a reader the same text in a different language. What becomes possible is re-explanation: taking the underlying idea and rebuilding the explanation out of material the reader already has — their own metaphors, their own schools, their own weather, their own market, their own references.

What the difference looks like

A concrete case makes the distinction clearer than a definition.

The failure mode of translation is subtle and easy to miss: it produces text that is grammatically correct, syntactically clean, and quietly meaningless, because the sentence's meaning depended on background the translation did not carry across. This is why "we support 100 languages" is not a meaningful claim on its own. The question is whether the depth survived.

Four choices that decide which future happens

If the outcome is a design decision, it is worth naming the decisions.

1. Localisation as architecture, not a feature. If the language a user thinks in is a setting you add in month eighteen, it will be a thin layer over an English product forever. If it is in the data model, the schema and the evaluation set from day one, everything downstream inherits it.

2. Evaluation in the languages you claim to support. A capability you have not measured is a capability you do not have. Benchmark quality in the actual target languages, with native speakers grading the output — not a translated English benchmark scored by a model.

3. Depth preservation as an explicit, tested requirement. Re-explanation has a failure mode that fluency hides: it can become simpler and shallower while sounding better. That has to be measured, or it will silently happen.

4. Access that does not require a card. A tool that removes a language barrier but adds an income barrier has not removed the barrier, it has moved it. Free tiers are not charity here; they are the difference between a product used by the people who need it and a product used by the people who already had options.

The uncomfortable part

There is a version of this argument that is comfortable for a technology company to make: AI will be inclusive, therefore we do not need to do anything special. That is not an argument, it is a hope.

The honest version is that inclusion is the harder path. Re-explaining in the reader's own context is more work than translating. Measuring quality in fifteen languages is more work than measuring it in one. Supporting a free tier is more work than not. The default outcome of least effort is an English-first product with translated surfaces, and that is precisely the outcome that widens the gap.

The future of education is either inclusive or it is not education at all. But that sentence only means something if it is followed by the specific decisions that make inclusivity true — and by someone actually measuring whether they worked.