What the research actually says about early years learning, teacher workload and artificial intelligence — and the design decisions we took because of it.
A kindergarten teacher with thirty children cannot see thirty children's mistakes. She sees the ones who put their hand up, the ones who cry, and the ones whose books she happens to mark that evening. The quiet child who has been dropping the same vowel for three weeks is invisible — not through any failing of the teacher, but because nobody has thirty pairs of eyes.
Marking is also where teachers' hours go. In the OECD's 2024 TALIS survey, teachers reported spending an average of 4.6 hours a week marking and correcting student work and a further 3 hours on administrative tasks.1 In Singapore — the closest surveyed system to ours — it is 6.4 hours of marking, and teachers named too much administrative work (53%) and too much marking (49%) as their two biggest sources of stress.2
Malaysia did not participate in the TALIS 2024 round, so these are the nearest comparable systems rather than local figures. We state that plainly rather than borrowing a number that does not exist.
Two findings in education research are unusually well established, and both are about practice and feedback rather than content or technology.
Black and Wiliam's review of formative assessment found typical effect sizes of 0.4 to 0.7 in favour of classrooms where teachers regularly checked understanding and adjusted — larger than most educational interventions ever measured.3 The catch has always been that doing it properly for every child, every day, is more work than a teacher has hours for.
Dunlosky and colleagues reviewed ten popular study techniques and rated only two as high utility: practice testing and distributed practice — being asked to retrieve an answer, and meeting it again spaced out over time.4 Rereading and highlighting, the two things children are most often told to do, did not make the cut.
Early literacy intervention research consistently finds that support given before and around school entry produces meaningful, lasting effects — the Danish SPELL trial, for example, tracked gains that were still measurable in second grade.5 A gap noticed at age five is a few weeks of practice. The same gap noticed at age nine is years of remedial work and a child who has already decided they are bad at reading.
The barrier has never been knowing this. It has been noticing in time.
The honest case for AI in a kindergarten is narrow, and it is strong precisely because it is narrow. AI is good at reading thousands of small events and finding the pattern. That is exactly the job no human can do for thirty children at once, and exactly the job that unlocks formative assessment at a scale teachers can actually use.
On the teacher's side. Every answer a child gives is analysed for the pattern behind the mistake — a dropped middle vowel, a swapped letter pair, a word that has been retried five times and never landed. The teacher opens a child and reads it in plain English. The class view shows which word the whole group is failing and who has gone quiet this week. AI does the noticing; the teacher does the teaching.
Children do not chat with a language model. UNESCO's 2023 global guidance on generative AI in education recommends a minimum age of 13 for independent use of generative AI tools, and notes that debates are under way to raise it to 16.6 Our children are four to six. So the child-facing side of KinderBuddy is scripted content and fixed word lists, with one exception: a short daily encouragement written from findings that already exist, reviewed against rules we wrote, and containing no free conversation at all.
The patterns are computed by deterministic rules over the school's own data — rules a person can read and check. The AI is given those findings and asked only to phrase them. It cannot invent a number, a word or a behaviour that is not in the data, and if the AI is unavailable the findings and the child's note still appear, written from the same rules. That ordering is what makes the system safe to trust in a school.
Malaysia's Digital Education Policy, launched on 28 November 2023 by the Minister of Education, sets out a 2023–2030 programme for digitally competent educators and digital pedagogy, including AI-based adaptive learning, AI-assisted tutoring and learning analytics that personalise teaching to individual learner profiles.7
A kindergarten that adopts this now is not experimenting ahead of the system. It is doing early, and at its own pace, what the national policy expects schools to be doing by 2030.
| The research says | What KinderBuddy does about it |
|---|---|
| Practice testing beats rereading | Every station asks the child to produce the answer — spell it, write it, say it — never to recognise it from a list. |
| Spaced practice beats cramming | A word answered wrongly returns the next day; a word mastered returns after a longer gap, stretching as the child gets it right. |
| Formative assessment has large effects | Teachers see today's answers today, with the pattern already identified — no marking required. |
| Early intervention is cheapest | Stuck words, hint dependence and quiet weeks are surfaced automatically, so a gap is noticed in days rather than terms. |
| Under-13s should not use generative AI freely | No child-facing chatbot. Scripted content only, plus one reviewed daily sentence of encouragement. |
| Children's data needs protection | Data stays in Kinderworld's own Google Cloud project in Singapore. No names are ever sent to the AI. Raw activity is deleted after 12 months. |
Effect sizes and survey figures are quoted as published by their authors. Where a figure for Malaysia does not exist, we say so rather than substituting another country's number without saying which.