By Craig Miles · June 2026 · 10 min read
Let me be careful with this title, because it is easy to misread. I am not arguing that AI is better than a teacher. It is not. A skilled, present, engaged teacher is the most powerful educational intervention we know of. Everything else — textbooks, video lessons, interactive software, AI tutors — is a more or less distant second.
But that is not the choice facing 273 million children who are currently out of school. Their choice is not between a teacher and an AI. It is between an AI and nothing. And when the question is framed that way — what can offline AI do that a teacher in a remote village cannot — the answer becomes both more interesting and more important.
Because there are things that an offline AI running on a smartphone can do reliably, at scale, and without preconditions, that a teacher in a remote and under-resourced setting structurally cannot. Not because teachers are inadequate. Because the conditions under which remote rural teachers operate make certain things nearly impossible, and those are exactly the things that AI handles without difficulty.
Understanding what those things are is the foundation of the ReachED model’s educational architecture — and, I think, the foundation of any honest conversation about what AI’s role in global education access should actually be.
What a remote rural teacher is up against
Before making any claims about what AI can do, it is worth being precise about the reality of teaching in the communities where the 273 million out-of-school children live.
In many sub-Saharan African and South Asian contexts, a primary school teacher in a remote rural area is likely to be: responsible for a class of 40 to 80 children spanning multiple grade levels; working without reliable access to teaching materials or updated curriculum; earning a salary that is frequently delayed or unpaid; operating in a building that may lack electricity, running water, and functional furniture; teaching subjects outside their area of training because no specialist teachers are available; and doing all of this while managing the personal precarity of living in a remote community with limited access to healthcare, transport, and the services available in urban areas.
This is not a description of teacher failure. It is a description of a system that is asking individuals to compensate, through personal effort and dedication, for structural failures that no individual can fully compensate for. The teachers who work under these conditions, and who produce good outcomes for their students, are doing something extraordinary. The system that puts them in this position is the problem, not the people within it.
But it does mean that what a remote rural teacher can reliably deliver — in terms of individualised attention, consistent pacing, immediate feedback, multilingual support, and availability outside school hours — is severely constrained. Not by ability, but by circumstances.
The things AI does without effort that teachers cannot
Infinite patience and zero fatigue
A teacher managing 60 children cannot give each child more than a few minutes of individual attention in a school day. When a child does not understand something, there is rarely time to explain it a second way, a third way, a fourth way. The lesson moves on because it has to.
An offline AI tutor does not move on until the child understands — or at least until it has exhausted its repertoire of explanations and flagged the gap for follow-up. It can explain the same concept twelve different ways without any trace of frustration, impatience, or the subtle signals that tell a child she is taking too long. For a child who learns slowly, or who has been absent and missed prior concepts, or who simply needs more time — this is not a marginal improvement. It is transformative.
This matters particularly for neurodivergent learners, for whom the social dynamics of a classroom — the pressure to keep pace, the fear of asking a question that seems obvious to others, the sensory environment of a crowded room — can be as much a barrier to learning as the content itself. I work with neurodivergent university students in England in my DSA mentoring practice, and the single most consistent thing I hear from them is that they needed more time, more repetition, and more permission to not understand immediately. An AI tutor provides all three, unconditionally.
Consistent availability
A school operates for a fixed number of hours on a fixed number of days. A child who misses school — because of illness, because of agricultural labour demands on the family, because of distance, because of the rainy season making roads impassable — falls behind, and catching up in a classroom context is difficult. The lesson is not repeated. The teacher does not have time.
An offline AI is available at any hour, on any day, at any pace the child chooses. A child who is kept home during harvest season can continue learning in the evenings. A child who was ill last week can cover the missed content at the weekend. The learning does not stop because the school is closed, the teacher is absent, or the road is flooded.
In communities where school attendance is irregular — which describes a significant proportion of the 273 million — this continuity of availability is not a convenience. It is the difference between a learning trajectory that compounds over time and one that perpetually resets.
Multilingual instruction without a multilingual teacher
In many of the regions where out-of-school populations are concentrated, children speak a local language or dialect that is different from the official national language of instruction. A child in rural Nigeria may speak Hausa, Yoruba, or Igbo at home, but the school curriculum — if she can access it — is delivered in English. A child in rural India may speak one of dozens of regional languages, but instruction is in Hindi or English.
The research on mother-tongue education is unambiguous: children learn faster, understand more deeply, and retain more when they are initially taught in the language they speak at home. Transitioning to a second language of instruction is a significant cognitive burden that compounds other learning challenges.
A teacher in a remote community typically speaks one language of instruction — the one they were trained in. An offline AI model can be fine-tuned to operate in multiple languages, including low-resource languages that have historically been excluded from digital education tools. The localisation challenge is real and requires deliberate investment, but it is an engineering problem with a tractable solution — not the structural impossibility it represents for human teacher deployment at scale.
Adaptive pacing without administrative burden
Genuinely adaptive teaching — adjusting the difficulty, pace, and sequence of content to match each individual learner’s current understanding — is the gold standard of educational practice. It is also extraordinarily difficult to deliver in a classroom of 60 children with one teacher and no teaching assistant.
In practice, classroom teaching converges on the middle of the attainment distribution. Children at the top are under-challenged and disengage. Children at the bottom fall further behind. The teacher knows this is happening and can do very little about it.
An offline AI delivers adaptive pacing as a default, not a premium feature. Every interaction is calibrated to the individual child’s demonstrated understanding. A child who masters a concept quickly moves to the next one immediately. A child who is struggling receives additional support before advancing. No child is held back by the pace of the class, and no child is left behind by it.
The OECD Digital Education Outlook 2026 identifies AI-powered Socratic tutoring that prompts reasoning and reflection as one of the most effective uses of AI in education — precisely because it replicates the kind of individualised, questioning-based dialogue that a skilled human tutor provides but that is structurally impossible at classroom scale.
Progress tracking without paperwork
Assessment is the foundation of effective teaching — you cannot adapt to a learner’s needs if you do not know what they know and what they do not. In under-resourced classroom settings, formal assessment is often infrequent, inconsistent, and used primarily for grading rather than for instructional adaptation.
An offline AI generates a continuous, granular record of every interaction — every question asked, every answer given, every concept where the learner hesitated or made errors. When the device syncs to the satellite connection during a scheduled LEO pass, this data is uploaded to provide teacher oversight, credentialling, and programme monitoring. The assessment is not a separate event that disrupts learning. It is embedded in every moment of learning.
What AI cannot do — and why that matters for design
I said at the start that I would be careful not to overstate the case, and I mean it. There are things that a teacher can do that no AI can replicate, and the ReachED model is designed around that reality, not in denial of it.
A teacher provides social belonging — the experience of being seen, known, and valued by a trusted adult. This is not decorative. It is educationally significant. Children who have a positive relationship with a teacher learn more, persist longer, and develop stronger intrinsic motivation. An AI cannot provide this.
A teacher provides moral and civic education — the transmission of values, norms, and ways of being in the world that are embedded in human relationships, not in curriculum content. An AI can deliver content about civics. It cannot model what it means to be a good person.
A teacher can see a child who is hungry, frightened, or unwell, and respond to that as a human being. An AI cannot.
The ReachED model does not ignore these things. It is designed as a bridging mechanism — a way to deliver structured learning to children who currently receive nothing, while the longer-term work of building human educational infrastructure continues. The AI is not the destination. It is the means of reaching children who would otherwise receive no education at all while the destination is being built.
For a child who is currently receiving nothing, a patient, consistent, adaptive, multilingual AI tutor available at any hour is not a compromise. It is the beginning of a learning life that would otherwise not exist.
The OECD verdict — and what it means for deployment
It is worth noting that the case for offline AI in education is not being made only by engineers and satellite entrepreneurs. The OECD Digital Education Outlook 2026 finds that even resource-constrained environments can benefit from AI in education through offline-capable small language models — a significant endorsement from the organisation that sets the international standard for education policy thinking.
The OECD’s framing is important, though. The key finding is that learning outcomes depend on how AI is embedded in pedagogical design — supporting thinking and agency rather than replacing them. This is the design principle that separates AI that genuinely improves learning from AI that creates the appearance of learning while actually producing passive consumption.
The ReachED AI tutoring layer is designed explicitly around this principle. The interaction model is Socratic — the AI asks questions rather than delivering answers, prompts the learner to reason rather than to recall, and treats every error as a diagnostic signal rather than a failure. The goal is not a child who can pass a test. It is a child who can think.
Small language models — the technical reality in 2026
A brief technical note for readers who want to understand the engineering basis of these claims.
Small language models — typically defined as models with between 500 million and 7 billion parameters — have matured significantly in the last two years. Models in this range can run on consumer smartphones without cloud connectivity, with response times that feel natural in a tutoring interaction. Microsoft’s Phi-3 Mini, Google’s Gemini Nano, Meta’s Llama 3.2, and Mistral’s Ministral-3B are all capable of running on mid-range Android devices with 4GB of RAM.
The educational use case is well suited to this class of model. It does not require the vast world knowledge of a frontier model. It requires reliable, structured, pedagogically sound interaction within a defined curriculum domain — exactly the kind of narrow, well-specified task that small models handle well, and that fine-tuning on educational content makes them handle even better.
Solar charging resolves the power constraint. Periodic satellite synchronisation resolves the content update constraint. The device generation required — Android smartphones from 2018 onwards — is already widely distributed in the communities ReachED aims to serve, and available as refurbished stock at near-zero cost in high-income countries.
The engineering is not the hard part. The hard part is the curriculum design, the community integration, and the will to deploy.
The question worth asking
If a technology exists that can provide a patient, consistent, adaptive, multilingual tutor to every child on Earth — available at any hour, on any day, requiring no building, no road, no grid connection — and we choose not to deploy it while 273 million children grow up without any education at all, that choice requires justification.
I have heard several versions of the justification. AI is not as good as a teacher. True, and irrelevant — the comparison is not with a teacher, it is with nothing. Technology-based education has failed before. True, and instructive — the failures tell us exactly what to do differently. We should invest in building proper schools instead. True, and insufficient — the proper schools are not reaching the children who need them on any timeline relevant to their lives.
The children who are out of school today will be adults in fifteen years. In fifteen years, will we still be building schools, or will we have found the will to send the signal?
The full architecture of the ReachED model — how the satellite, AI, and device layers fit together — is explained in detail in What Is the ReachED Model.