The strongest argument for technology in education is not efficiency. It is an old and well-replicated finding: students taught one-to-one substantially outperform students taught in a class of thirty. The effect is large and consistent across decades of research.
Nobody disputes it. The problem has always been that one tutor per child is economically impossible for a public system. If software can deliver even a fraction of that benefit at software prices, it is genuinely important.
That is the honest case. What follows is where it holds up and where it does not.
What tutoring actually does
Worth being precise, because "personalisation" is often taken to mean the wrong thing.
A good tutor does at least five things a classroom cannot:
- Adjusts pace. Moves on when you have it, stays when you do not.
- Diagnoses the specific misconception. Not "you got it wrong" but "you are subtracting the smaller digit from the larger regardless of position".
- Asks rather than tells. Prompts you toward the answer instead of supplying it.
- Maintains motivation. Notices frustration and adjusts.
- Requires you to produce. You cannot hide at the back.
Notice that only the first is about content delivery. Most educational software personalises pace and difficulty, which is the easiest of the five and the least valuable. The high-value items are diagnosis, questioning and motivation, and those are much harder.
Where the technology genuinely delivers
Practice with immediate, specific feedback
The clearest win. Retrieval practice with feedback is among the best-evidenced techniques in learning science, and it is limited in practice by marking capacity. A teacher with thirty students cannot mark thirty pieces of work within the minute in which feedback is most useful.
Software can, for anything with a checkable answer. Mathematics, grammar, vocabulary, chemical equations, code. This is not glamorous, it is well established, and it works.
Spaced repetition scheduled properly
Reviewing material at expanding intervals produces dramatically better retention than massed study. Doing it well requires tracking hundreds of items per student and scheduling each individually — pure bookkeeping, which is exactly what computers are for.
Explaining the same thing five different ways
A teacher has a limited number of explanations available in the moment. A language model has effectively unlimited variation: a different analogy, a worked example, a diagram, a simpler framing, the same idea from the other direction. For a student who did not get it the first time, being able to ask again differently is worth a great deal.
Reducing marking load
Not personalisation at all, but possibly the highest-value application. Teachers report administrative burden as a primary cause of exhaustion and attrition. First-pass marking, differentiated worksheet generation, and lesson resource drafting are real time savings, and time returned to a teacher is time available for the things only teachers do.
The best evidenced use is not replacing instruction. It is handling the parts of teaching that are mechanical — marking, scheduling, drilling, resource production — so that the human hours go to explanation, diagnosis and encouragement. Products sold as teacher replacements have consistently underdelivered. Products sold as teacher amplifiers have consistently done better.
Where it falls down
Answers instead of learning. A system that will give the answer teaches students to ask for answers. Struggle is not an obstacle to learning; it is substantially the mechanism. Tools that remove desirable difficulty feel effective and produce worse outcomes — and both the student and the dashboard will report satisfaction.
Confident wrong content. A tutoring system that states something incorrect with authority is worse than no system, because a student has no basis to doubt it. In subjects with clear right answers this is testable. In history, ethics and literary interpretation it is much harder to catch.
Motivation, which is most of the battle. Any teacher will tell you the hard part is not explaining fractions. It is getting a discouraged fourteen-year-old to attempt the problem at all. That is a relationship, and no current system does it convincingly. Engagement mechanics produce short-term compliance and then wear off.
Measuring the easy thing. What is easy to assess automatically is procedural fluency. What matters most is transfer — applying a concept in an unfamiliar situation. Systems optimise what they can measure, and a student can look excellent on a dashboard while having learned very little that generalises.
The equity inversion. Personalised learning tends to work best for students who are already organised, literate and motivated — and for households with reliable devices, quiet space and a parent who can help. Deployed naively, it can widen the gap it was supposed to close. This is the finding most consistently ignored by procurement.
Assessment is the unsolved problem
Take-home written work has become substantially less reliable as evidence of individual learning. Detection tools are not accurate enough to sanction a student on, and false accusations do real harm — particularly to students writing in a second language, who are disproportionately flagged.
The workable responses are structural rather than technological:
- Assess the process, not only the artefact: drafts, revisions, notes, a short viva.
- Move important assessment in-person where the stakes justify it.
- Set tasks that require personal or local specificity which generic output cannot supply.
- Permit the tools openly and raise the bar for what counts as good work.
- Stop treating fluent prose as proof of understanding. It was always a weak proxy; now it is a broken one.
Institutions that have adapted the assessment design are managing. Those relying on detection are not.
What good deployment looks like
If you are choosing tools for a classroom or a child:
- Ask what it replaces. If it adds work for teachers, it will not survive the term.
- Prefer questioning over answering. A tool that asks "what did you try?" beats one that produces the solution.
- Check the subject content is verified, not generated on demand, for anything with authoritative answers.
- Demand subgroup data. How does it perform for students below grade level, with additional needs, or learning in a second language? Aggregate improvement often hides a widening gap.
- Keep the teacher in the loop with real information. Not an engagement score — specific misconceptions, so the human can intervene usefully.
- Read the data terms carefully. Children's data warrants a higher standard than the sector generally applies.
- Confirm it works offline or on old hardware, if equity is a genuine concern rather than a slide in the pitch deck.
An honest expectation
The realistic outcome is not a tutor for every child. It is that the mechanical half of teaching gets substantially cheaper, teachers get some hours back, students get better practice with faster feedback, and the human parts of education — diagnosis, encouragement, the relationship that makes a discouraged student try again — remain stubbornly human.
That is a meaningful improvement, and considerably less than has been promised. The gap between those two is where most education technology money has historically been lost, and there is no particular reason to think this cycle will be different unless buyers ask harder questions than they usually do.