“AI-powered learning” describes four different jobs
The phrase has been stretched over so many products that it no longer narrows anything down. It is more useful to ask which of these jobs the AI is actually doing, because they have very different value and very different risks.
Analysis
Reading a student's responses to infer what they understand. Hard, unglamorous, and the job that determines whether anything else helps.
Decision
Choosing what the student does next — which concept, at which level. Only possible if the analysis is good.
Explanation
Answering “I don't understand this step” in the moment. Highly visible, immediately useful, and the easiest to do badly.
Generation
Producing questions, worked examples and practice tests. Valuable for coverage, but it produces material, not understanding.
Most products that describe themselves as AI-powered are doing job 3, sometimes job 4. Jobs 1 and 2 are where learning outcomes actually move, and they are considerably harder — because they need a map of the concepts in a subject, not just a good language model.
A chatbot is not an adaptive engine
This is the single most useful distinction to hold on to when evaluating anything in this category.
| A study chatbot | An adaptive learning engine | |
|---|---|---|
| Scope | The question in front of it | Every response a student has given |
| Memory | The conversation, and then nothing | A continuously revised model of concept understanding |
| Output | An explanation | A decision about what to do next |
| Runs | When asked | Continuously, in the background |
| Fails by | Handing over the answer | Choosing the wrong concept to work on |
A student with only a chatbot gets unstuck faster and learns roughly as much as before — possibly less, if the assistant simply solves the problem. A student with only an engine gets the right work but no help when the work is hard. The two jobs are complementary, and a product that does one should not be described as if it does both.
What AI cannot do here — and shouldn't pretend to
An honest account of the limits is more useful than another list of capabilities.
- It cannot read a child's mind. An engine infers likely gaps from patterns in evidence. It does not know what a student was thinking, and any product that implies otherwise is overselling.
- It cannot be certain from one wrong answer. A single incorrect response is weak evidence — a slip, a misread, a distraction. Anything confident built on one data point is guesswork with a progress bar.
- It cannot see the room. A student who has stopped sleeping, or is being bullied, or simply had a bad week, produces the same pattern as a student with a concept gap. The teacher can tell the difference; the engine cannot.
- It cannot supply motivation. Knowing exactly which concept to fix does not make a discouraged student want to fix it.
- Generated material still needs checking. A model that can produce a thousand questions can also produce a thousand subtly wrong ones. Volume is not the hard part.
The practical consequence is that recommendations should be surfaced to a teacher as something worth looking at, rather than executed as a verdict. That is how Sublogy is built.
The questions worth asking about AI and children
If a product is going to sit between a child and their schoolwork, these are the things worth checking — of any vendor, including us.
- Is the assistant restricted, or open-ended?An assistant that will discuss anything is a general chatbot in a school app. Sublogy Buddy is restricted to the lesson or topic being studied.
- Are replies screened before the child sees them?Sublogy Buddy's replies are screened for material unsuitable for children before they are shown.
- What identifying information leaves the product?To answer a question, Sublogy sends the AI service the student's question, the relevant lesson material and the topic — not the student's name, username, school name or account identifier.
- Can the AI provider train on the data?Our agreement with the AI provider does not permit Sublogy data to be used to train its models.
- Are conversations stored?Sublogy Buddy conversations are held only in the app while it is open, and are not stored on our servers.
- Is there advertising or behavioural tracking?None anywhere in Sublogy — no ad software, no analytics following students across other apps or sites, and no selling or renting of personal data. Data is stored in India.
- Does it help the student understand, or finish?The difference between explaining the step a student is stuck on and handing over the answer is a design decision, and it is the one that decides whether the product teaches anything.
The full detail on all of this is in our privacy policy.
How Sublogy uses AI
Sublogy runs two distinct AI systems, deliberately kept separate because they do different jobs and carry different risks.
CUBALE — the engine
The Concept Understanding Based Adaptive Learning Engine, built in-house rather than assembled from off-the-shelf parts. It does jobs 1 and 2: analysing learning patterns across a student's responses, surfacing the concepts that need attention — including foundational ones a single question would never reveal — and choosing what the student practises next.
It improves using assessment information: whether an answer was correct, the number of attempts, how recently a concept was practised, which question was asked, and the concept being assessed. A student's name, username and school name are not used to train CUBALE. It is built to learn about concepts and learning patterns, not about individual children.
Sublogy Buddy — the assistant
A study assistant that does job 3. When a student is stuck mid-step, Buddy talks through the worked example on screen in short, simple sentences. It keeps to the lesson in front of it, and it is built to help a student understand rather than hand over answers. It is powered by a third-party AI service, under the constraints listed above.
Generated practice, checked
Sublogy also uses AI for job 4 — for example, generating Olympiad-style practice tests from the chapters a student has actually studied, then analysing the attempt and providing a worked solution for every incorrect answer.
What we don't claim. CUBALE identifies patterns and detects potential gaps. It surfaces concepts needing attention and recommends areas for reinforcement — adapting learning based on evidence rather than claiming to know every misconception perfectly. The teacher stays in the loop, because they can see what the data cannot.
Questions about AI in learning
What does AI-powered learning actually mean?
It covers several different jobs that are often described with the same phrase: analysing a student's responses to work out what they understand, choosing what they should do next, explaining something when they get stuck, and generating practice material. A product may do one of these or all four, so the useful question is always which job the AI is doing.
Is an AI chatbot the same as an adaptive learning engine?
No, and the difference matters. A chatbot answers the question in front of it and then forgets. An adaptive engine reads patterns across hundreds of responses to decide what a student should work on next. Sublogy uses both, for different jobs: CUBALE is the engine, Sublogy Buddy is the assistant.
Is AI-powered learning safe for children?
It depends entirely on how it is constrained. The risks worth asking about are: whether the assistant will answer anything at all or only the lesson on screen, whether replies are screened before a child sees them, what identifying information is sent to any third-party AI service, whether that provider may train on the data, and whether there is advertising or behavioural tracking in the product.
Does Sublogy's AI train on student data?
CUBALE improves using assessment information — correctness, attempts, recency, which question, which concept. A student's name, username and school name are not used to train it. Our agreement with the third-party AI provider behind Sublogy Buddy does not permit Sublogy data to be used to train its models, and nothing identifying the student is sent with a Buddy question.
Will AI replace teachers?
Not for the work that matters. AI is good at the part no person can do — holding thirty students, hundreds of weekly responses and dozens of interlinked concepts in view at once. Deciding what to do about a particular child is a teaching judgement, which is why Sublogy surfaces recommendations to a teacher rather than acting as the final word.
AI doing the analysis, not the teaching
Sublogy is provided through schools, for classes 6 to 10.
Sublogy for schools Read the privacy policy