The Learners Everyone Forgot: Why Adult Education Is Where AI Tutoring Will Prove Itself

The debate over AI in education has two settings: panic about children and plagiarism on campus. Meanwhile, four evenings a week, adults file into Montreal classrooms this debate has never visited. I teach them. A man arriving straight from a warehouse shift, still in his work boots, opening a math workbook. A mother doing her own homework after her kids are asleep. A newcomer who spends mornings in francization and afternoons at a job, and still shows up. Nobody writes think pieces about these learners. They are the ones everyone forgot — and they are exactly where AI tutoring will prove itself or quietly fail.

The Room Nobody Writes About

Quebec runs a full parallel school system most people have never set foot in: the centres d'éducation des adultes. Inside them, adults finish interrupted schooling, earn the credits that unlock trades programs, learn French through francization, and rebuild foundations that were never poured properly the first time. This sector is large, it is public, and it is almost completely absent from the AI conversation [1]. When a ministry, a startup, or a newspaper says “AI in education,” it means children and undergraduates. It never means my students.

The scale deserves better than silence. The OECD's Survey of Adult Skills has found, cycle after cycle, that a large share of adults in Canada — by most readings, close to half — score below the literacy level associated with comfortably handling complex written text [2]. And Statistics Canada's data on adult learning tells the other half of the story: the adults who need training most are the least likely to get it, blocked by cost, schedules, and family responsibilities [3]. These are not edge cases. They are working, voting, tax-paying adults — and the system meant to catch them remains under-discussed and under-tooled.

The Hardest Classroom in the Country

Walk into my classroom and try to define “grade level.” In the same room, on the same evening, one student is rebuilding fractions, another is finishing grade-11 functions, and a third is somewhere in between, working through an individualized module — because Quebec's adult general education framework is built on individualized progression rather than a single lockstep curriculum [4]. The teacher does not lecture to the middle. There is no middle. You circulate, you triage, you become twenty different tutors an hour. It is the most demanding teaching I have ever done, and I say that as someone who loves it.

Now add the constraint that defines everything else: time. My students do not have study weeks. They have shifts, kids, appointments, and fatigue. A father misses Tuesday because the daycare called. A woman on night shifts sleeps through the morning session she is enrolled in. And the stakes are not abstract. A failed module delays a diploma; a delayed diploma delays a trade program; a delayed program delays the job the whole plan depends on. For an adult learner, a lost month is not a setback. It is rent.

Which produces the most honest user base in education: adults have zero tolerance for gimmicks. Nobody assigned them to be there. They chose it, at real cost, and they vote with their feet. If a tool wastes twenty minutes of the only hour they have, they never open it again. Confetti animations and streak badges do not move them. Progress moves them. Any AI tutoring product that survives contact with this audience has earned something no pilot study in a well-funded lab school can grant.

What a Tutor That Never Sleeps Changes

Here is what I cannot do: be available at 11 p.m., when the warehouse shift ends and the workbook comes out at the kitchen table. Or at 5 a.m., before the first bus. Or on Sunday afternoon, the only quiet stretch a single parent gets. The defining feature of adult learning is that it happens in the cracks of a working life — and the defining limit of adult education is that the humans who teach it are on a schedule. A tutor that is genuinely available around the clock is not a luxury feature for these learners. It is the difference between studying and not studying.

The second thing an AI tutor offers is harder to measure, but I watch it every week: review without judgment. Many of my students carry decades of shame about school. Asking the same question a third time in front of a class costs them something real. Asking it a fifth time to a machine costs nothing. We have known since Bloom's famous two-sigma paper that one-on-one tutoring dramatically outperforms group instruction [5]; what we undersell is that for adults, the tutor's patience matters as much as its pedagogy. A model does not sigh. It does not glance at the clock. For someone rebuilding fractions at 42, that is not a small thing.

And in Quebec, there is a third offer no imported product takes seriously: language. My classrooms run in English, in French, and often in both within the same explanation. A tutor that can explain a math concept in French, restate it in English, and switch back mid-problem — without making the learner feel the switch is a failure — matches how these classrooms actually work. Bilingual support is not a checkbox here. It is the terrain.

What It Must Not Do

Now the hard part, because the learners who benefit most from AI tutoring are also the easiest to hurt with it. Start with the obvious failure mode: handing out answers. An adult with one hour and an exam in three weeks will take the shortcut if you offer it — and the shortcut is a trap, because the exam is on paper, alone, with no chatbot in the room. This is why I built Mnemosyne around guided discovery: the tutor asks before it tells, probes what the learner already knows, and refuses to be a vending machine for solutions. A tutor that answers instantly produces dependent users. A tutor that questions produces graduates.

The second failure mode is pretending hallucinations do not exist. Every large language model sometimes produces confident nonsense; I have written about this before, and nothing has changed. For a learner with fragile foundations, a wrong answer delivered fluently is worse than no answer — they lack precisely the background knowledge that would let them catch it. So verification cannot be optional, and neither can the habit of doubt. Part of teaching adults with AI is teaching them to interrogate the AI. That is not a limitation of the product. In 2026, it is a life skill.

Law 25 Applies to Them Too

The third failure mode is legal, and it is not hypothetical. An adult learner's file is dense with personal information: the record of what they struggle with, sometimes their immigration journey, their work situation, their interrupted past. Quebec's Law 25 does not have a carve-out for edtech that means well [6]. Consent has to be real, data location has to be known, and retention has to be defensible. If a tutoring product cannot answer where a learner's transcript is stored and under whose jurisdiction, it has no business in a public adult education centre.

This is where building in Quebec is an advantage rather than a burden. Compliance designed in from the first commit costs a fraction of compliance bolted on after a procurement officer starts asking questions. The learners in these centres include some of the province's most vulnerable adults. They deserve tools that treat their data with the same seriousness as their learning — not products that discover Law 25 exists during contract negotiations.

The Real Benchmark

Every AI tutoring demo I have seen works beautifully under favorable conditions: a motivated user, fluent in the interface language, with time, a laptop, and no fear of school. Adult education strips those conditions away, one by one. Fragmented time. Second languages. Old wounds. Wildly mixed levels. High stakes. It is the adversarial test suite for the entire category — and almost nobody is running it.

I think about one composite student — every adult ed teacher knows her. She is 42. She works two jobs. She left school at sixteen for reasons that were never really hers. She is finishing grade-11 math because the diploma gates a health-sector program, and she studies in fragments: twenty minutes on a bus, an hour after bedtime, Sunday afternoons. No pilot program was designed for her. No demo video features her. She is the actual market, and the actual test.

So here is my benchmark, and I hold my own product to it before anyone else's. Forget the lab-school pilots and the university partnerships for a moment. If AI tutoring works for her — between two jobs, in her second language, at 11 p.m., with an exam that decides her next five years — it works. Full stop. And if it only works for the learners who were always going to be fine, we have built nothing new. We have just found a shinier way to forget the same people.