Education and the literacy gap

The adult literacy gap

Roughly 30% of adult Nepalis have weak literacy by working definitions; the number is higher among women, rural residents, and in specific districts of the Far West and Madhesh. For most of these adults the gap is permanent — adult literacy programmes have run for decades with modest success.

This is the kind of social problem where AI can do something genuinely new. Three approaches are now plausible at low cost:

Speech-first interfaces that let a low-literacy adult interact with public services, banking, or commerce by voice in Nepali — as covered in Chapter 2 — remove the literacy bottleneck from a wide range of daily tasks. The technology is ready; the products are not yet built.

Speech-to-text assisted reading practice. A learner reads a short passage aloud; the model checks pronunciation and accuracy and gives gentle correction. This is well-developed in English-language education products and can be ported to Nepali with comparatively little effort.

Personalised written-Nepali tutoring that takes a learner from “I can recognise letters” to “I can read a newspaper” by adapting to their pace, common errors, and vocabulary. The same LLMs that struggle with cultural fluency are quite good at patient, repetitive, level-adapted explanation.

None of these replaces a teacher. They extend reach into populations where teachers are scarce — and they fit naturally into NGO-led adult literacy work that has been running for decades.

The school question

The other question is harder, because the technology is moving faster than schools can respond. A grade-9 student today can produce a competent Nepali essay in under thirty seconds. Most teachers can spot LLM writing — for now. Most assignment policies pretend the technology does not exist — for now. Neither of these is sustainable.

Three honest positions on what schools should do.

The “ban it” position. Pretend the technology does not exist; require all work to be produced in front of the teacher; police LLM use. This is being tried in many schools. The pattern from elsewhere is that it fails — students use the technology anyway, more secretively, with less guidance from adults. The ban model imports adversarial relationships between students and teachers that the country does not need.

The “embrace it uncritically” position. Allow LLM use freely; treat it as another tool; trust students to learn what skills are still worth developing. This has its proponents. It fails in a specific way: students who use LLMs as a shortcut never develop the writing-and-thinking muscle that comes from struggling through a hard paragraph. By university they are fluent in editing AI output and incapable of producing their own.

The “structured integration” position — and the one most likely to work — accepts that LLMs are now part of the writing environment and asks what skills students should still develop, and how. It tends to look like: in-person writing exercises with no devices remain core; LLM-assisted research and drafting is taught explicitly; revising, critiquing, and verifying AI output becomes a major part of literacy in itself. This is more work for teachers, not less, and requires textbooks and curricula that do not yet exist for Nepali schools.

What a working education-AI product looks like

Three traits, from the small set of education-AI products that have shown durable usage in lower- and middle-income countries.

1. It does one specific learning task well. A reading-fluency app that helps a learner read 100 graded Nepali passages with feedback is more useful than a “learning assistant” that tries to be everything.

2. It works with the teacher, not around her. Apps that turn the teacher into a viewer of a student dashboard succeed. Apps that try to replace the teacher with an avatar fail. Nepali parents and teachers have strong instincts about this — they are not wrong.

3. It is honest about what AI can and cannot do. A product that quietly inserts hallucinated Nepali history into a tutoring conversation will, eventually, be caught — and the embarrassment will damage adoption for the next generation of products. The honest approach is to keep AI on tasks where its strengths (patience, consistency, language fluency) shine, and keep it off tasks where its weaknesses (cultural accuracy, factual reliability) bite.

Check your understanding

Quick check

—

A school in Pokhara is deciding how to handle student use of LLMs in writing assignments. Which approach is most likely to produce students who can actually write — and think — well by the time they finish school?

1. APretend LLMs do not exist; police all assignments; punish students caught using them.
2. BAllow unrestricted LLM use; trust students to use the tool well.
3. CA structured-integration approach: in-person writing exercises with no devices remain core; LLM-assisted research and drafting is taught explicitly; revising, critiquing, and verifying AI output is itself part of the curriculum.
4. DMove all instruction online and let an LLM tutor each student individually.

Check answer→

What comes next

This closes Chapter 5 and most of the descriptive part of the course. We have toured the substrates — language, money, land, people — and where AI sits across them. Chapter 6 is the opinion side: what Nepal should build, who should build it, and what a credible five-year agenda actually looks like.