A learner finishes a lesson, selects the correct answer and moves on. The screen shows progress. What would happen if we changed the question, asked them to explain their reasoning, or revisited the idea a week later?
Those are the questions shaping how we build KaratuAI.
An AI Trend Hunter article about AI and children’s learning raised a concern that deserves a serious response: an assistant can make a task easier while leaving too little thinking for the learner. Vocabulary, the timing of help and the relationship a system encourages all deserve attention.
Our current work focuses on secondary-school learners, university students and people learning at work. We are translating those concerns into specific product decisions, with a clear distinction between features we have implemented and learning outcomes we still need to establish.
Consider a secondary-school learner practising algebra. In an illustrative exchange, they might ask for help with 3x + 6 = 18. The useful next step depends on what they already understand. Someone meeting equations for the first time needs an explanation and a worked example. Someone practising a familiar method needs an opportunity to attempt it.
KaratuAI’s tutoring guidance makes that distinction. For practice and homework, the intended sequence is an attempt, a targeted hint, a retry, then a clear explanation of the method and a different problem to try. A learner who remains stuck should receive teaching. Repeatedly asking them to guess is not a substitute for explaining.
We apply this principle more explicitly in structured homework and practice. In multiple-choice practice, selecting the correct option and explaining why it is correct are assessed separately. If the explanation cannot be reviewed reliably, it remains unassessed. That gives the learner a more useful signal than a confident mark built on uncertain evidence.
Learners select an answer and explain their reasoning. Both parts matter.
This approach also has a basis beyond AI. The Institute of Education Sciences’ practice guide on organising instruction and study recommends combining worked examples with problem solving, spacing learning over time and asking explanatory questions. These principles inform our design; they do not establish that KaratuAI itself improves learning outcomes.
Appropriate help also depends on context. A university student encountering a new topic may need smaller steps. An experienced employee may want a deeper account of why a procedure works. Educational level alone cannot tell us everything about either person.
KaratuAI uses educational context from the signed-in learner’s account or selected profile when preparing tutor guidance. On the practice screen, learners can choose Simpler, Standard or Deeper explanations. These are communication preferences. They do not diagnose ability, establish an exact age or replace attention to what the learner demonstrates in the conversation.
Learners can choose how much detail they want in subsequent tutor replies.
Simpler language must also preserve the concept. Explaining photosynthesis clearly, for example, still requires identifying carbon dioxide and explaining what it is. Removing an unfamiliar term can remove an essential part of the idea. Our tutoring guidance explicitly asks for essential terms to be explained in familiar words.
Learning should have somewhere to go after the lesson. For eligible saved secondary and university lessons, and eligible lessons assigned to staff, we have added practice that asks learners to apply the material to new situations. Newly generated question sets receive a separate AI review for correctness, alignment with the lesson and whether the tasks are sufficiently distinct. If a reliable set cannot be prepared, the system should not present repeated questions as fresh evidence.
For a university learner, an application could involve interpreting a different dataset. In workplace learning, it could involve deciding how to investigate a mismatch between a stock register and delivery records. These are illustrative situations: the purpose is to see whether the learner can use the principle when the details change.
After a successful fresh application without recorded in-app help, the practice flow can make a previously unseen check available about seven days later. Difficulty on that check leads back to recovery practice. Existing opted-in reminders can point eligible learners to an available check. The current question sets are finite, so this is a bounded practice sequence, not a promise of unlimited new questions.
We are also changing how we describe progress. Assisted answers, repeated questions and fresh attempts tell us different things. KaratuAI’s learning evidence keeps those conditions distinct and carries that context into subsequent tutor guidance. When a learner used a hint, that support should remain visible in the interpretation of their answer.
Missing evidence is labelled “Not yet assessed”; it is not treated as a judgement of the learner’s ability.
“Independent” has a specific limit here: it means no assistance was recorded inside the app for that attempt. We cannot establish whether someone received help elsewhere. One successful delayed check is encouraging evidence about one task. Establishing lasting understanding requires more observation.
Safety deserves its own treatment. In homework and practice, responses are checked for distress, abuse and other safety concerns before academic grading. Those checks are designed to recognise multilingual, mixed-language and imperfectly transcribed wording. A disclosure should prompt a supportive response and guidance towards safe human support.
Fixed support messages currently cover English, French, Hausa, Yoruba, Igbo and Swahili, with English used when a supported localised response is unavailable. Native-speaker and contextual review remain important work. KaratuAI’s tutor guidance also sets boundaries around romantic attachment, exclusivity and replacing trusted people. It does not contact a parent or emergency service on a learner’s behalf.
UNESCO’s guidance on generative AI in education calls for a human-centred approach and careful educational validation. We see teachers, parents, trainers and other trusted people as essential participants in that work. Our present scope does not establish suitability for young primary-school children.
Testing these decisions takes more than checking whether a feature opens. We use synthetic conversations to examine teaching behaviour across follow-up turns, languages and requests that try to bypass the intended approach. Separate AI evaluation helps identify failures, while human review remains necessary. Passing those tests measures behaviour under the tested conditions; it does not demonstrate better learning in everyday use.
We have also built an operator-run tool for reviewing actual slide images and narration. Automatic review of every generated lesson remains unfinished. A well-formed lesson script cannot establish that every diagram or spoken explanation is correct.
The next evidence we need comes from educator-supported studies with real learners: whether they can explain an idea, apply it without help and retain it later, across different subjects and circumstances. We have begun building the measurement foundation for those questions. We have not yet demonstrated a causal improvement in learning outcomes.
Our aim is to help learners become more capable of doing the work themselves. We want families and educators to examine the assistance, the opportunities to practise and the evidence that follows. If you would like to discuss a learning evaluation with KaratuAI, get in touch.