The vocabulary of modern learning technology has collapsed into mush, and the mush is commercially convenient: "adaptive", "personalised" and "AI-powered" now decorate products whose actual mechanisms range from a settings menu to a longitudinal learner model, with buyers given no way to tell which they are purchasing. This page is the canonical glossary this pillar and our product comparisons run on, replacing our earlier adaptive-learning overview, and it corrects that article's one overreach at the outset: dynamic systems are a powerful way to scale individualised practice, not the only way, and high-quality conventional question banks remain genuinely useful. Precision about mechanisms, not tribal claims about categories, is what the terminology below exists for.
The ladder of terms, from weakest to strongest
Personalised: the product reflects your selected exam, date, goals or chosen topics; a manually configured session is personalised without being adaptive in any sense, and most "personalisation" claims live here. Performance-informed: the product displays analytics or recommends content using your prior results, more meaningful, and still possibly static, a dashboard is performance-informed without changing anything. Adaptive selection: the system changes which question, topic or activity appears next in response to your performance, the first rung where "adaptive" is honest. Difficulty adaptation: the system serves easier or harder items in response to estimated ability, the mechanism of computer-adaptive testing, and conceptually different from revisiting weak knowledge, a distinction most marketing blurs. Spaced repetition: the system schedules retrieval attempts across time; merely filtering for previously incorrect questions is not, by itself, a spaced-repetition system, because a filter has no schedule. Mastery estimation: the product estimates your state for a knowledge component; a percentage-correct display is not necessarily a mastery model, and usually is not. Knowledge tracing: the system updates an estimate of what you know across a sequence of interactions, ideally accounting for difficulty, prior opportunity, forgetting and slip or guess probabilities, the genuinely model-based tier. Conversational explanation: you can ask follow-up questions about an item; useful, and not tutoring. Socratic tutoring: the system elicits your reasoning, identifies the misconception and scaffolds progressively before disclosing the answer, the tier the evidence on guardrails is about: /blog/chatgpt-is-not-an-ai-tutor-educational-guardrails. Study orchestration: the system combines performance, timing, coverage, deadlines and available activities to decide what you should do next across the whole preparation, the strongest claim, and the rarest genuine implementation. Score prediction: the system estimates a future external examination result, a claim strong enough to need its own validation standard: /blog/can-a-medical-qbank-predict-your-exam-score.
Worked examples: naming what products actually do
Five common implementations, named by the ladder. An incorrect-question filter: performance-informed review, not spaced repetition and not adaptive selection. A scheduler that resurfaces items at intervals adjusted by your success: spaced repetition, genuinely. A test that serves harder items as you succeed and stops when ability is estimated: computer-adaptive testing with difficulty adaptation and mastery estimation, an assessment function rather than a learning schedule. A bank that biases tomorrow's questions toward yesterday's weak topics: adaptive selection with performance-informed targeting, the commonest honest use of "adaptive". A dialogue attached to your wrong answer that asks what you were thinking before explaining: Socratic tutoring, if it genuinely asks first. iatroX, named by its own ladder for symmetry: adaptive selection with weak-area targeting, spaced repetition, Socratic tutoring and study planning, with mastery displays that are estimates, and a readiness indicator that is an internal composite, not a validated score prediction.
Which mechanism solves which problem, and what pages should disclose
The buyer's translation. Broad gaps and limited time: adaptive selection earns its keep. Forgetting what was once known: spaced repetition, nothing else on the ladder addresses it. Not knowing why you got it wrong: tutoring, conversational at minimum, Socratic ideally. Wanting to know if you are ready: honest mastery estimation plus full mocks, read with the calibration caveats at /blog/why-qbank-percentages-are-not-comparable. And the disclosure standard product pages owe buyers: which ladder tiers are implemented, what the system observes, whether the model persists across sessions, and which claims are vendor-reported versus validated, the standard our own comparisons apply to every provider, ourselves included.
Frequently asked questions
Is more adaptive always better?
No: mechanisms solve problems, and a learner with narrow, known gaps may be best served by manual selection plus spacing; the ladder is a menu, not a leaderboard.
Why does the terminology matter commercially?
Because identical words currently price wildly different machinery, and a buyer who can name the tiers can ask the one question that deflates marketing: which of these does your product actually implement?
Where does "AI-powered" fit on the ladder?
Nowhere: it describes implementation technology, not mechanism, and every tier above can exist with or without it; the question is always what the system does, not what it is made of.