I want to spend a few thousand words pulling the word "adaptive" apart, because it's been so thoroughly abused by edtech marketing that the category has become functionally meaningless. Every platform with a quiz engine and a question bank now claims to be adaptive. Some of them have a difficulty slider that nudges easier or harder based on recent correctness. A handful do a little Bayesian Knowledge Tracing under the hood. One or two have actually built something that meets the technical definition. The rest are autocomplete with confetti animations.
This isn't pedantry. The conflation matters because learners — and the institutions buying on their behalf — make purchase decisions assuming that "adaptive" means roughly the same thing across vendors. It doesn't. The gap between the best of the category and the median is roughly the gap between a search engine and a phone book.
The Technical Definition
For a system to be meaningfully adaptive, four properties have to hold:
- Per-learner state. The system maintains a model that is specific to this learner, distinct from every other learner, and that updates over time as the learner interacts. Not a class average. Not a cohort. Not a difficulty band. A model of this person.
- Real-time updates. The model updates between interactions, not overnight, not after the session ends. The state that exists when question N+1 is selected has fully incorporated question N's answer. If the system batches updates, it isn't adapting — it's recording.
- Decision impact. The next thing the system shows the learner is causally determined by the current state. If two learners with different states would see the same next question, the system isn't adapting; it's serving from a fixed playlist with smoothing.
- Diagnostic granularity. The state distinguishes meaningfully between why a learner got something right or wrong. A learner who gets a question wrong because they don't know the topic is a different problem than a learner who gets it wrong because they misread it, or because the question itself is ambiguous, or because they failed a prerequisite the question implicitly assumed. A system that treats those four cases identically is not diagnosing — it's counting.
A system that fails any of those four is not adaptive in the sense that matters. A system that fails three of them and still advertises adaptivity is committing fraud at a marketing level.
What Most Platforms Actually Do
I'll be specific, because vagueness here helps no one. Here's what the leading consumer-facing edtech platforms actually implement:
Static question banks with a difficulty slider. The learner takes a question. If they get it right, the next question comes from a slightly harder bucket. If they get it wrong, the next comes from a slightly easier bucket. The buckets are predefined. The state is "current difficulty level," which is a single integer. This is what Khan Academy does for the most part, and what most cert-prep apps do. It is not adaptive — it's a difficulty-tuning loop with a one-dimensional state.
Item Response Theory (IRT) with offline calibration. Each question is pre-tagged with a difficulty parameter. The learner's ability is estimated using maximum likelihood over recent answers. The next question is chosen to maximize information gain at the current ability estimate. This is a meaningfully better approach than a bucket slider, and several test-prep vendors (Kaplan, the Princeton Review, ETS) use variants of it. But the state is still one-dimensional (a single "ability" scalar), the question parameters are static (calibrated offline, on a different cohort), and the diagnostic granularity is essentially zero — you can't tell why the learner is failing, only how often.
Bayesian Knowledge Tracing (BKT) with per-skill mastery probabilities. This is the closest the legacy edtech category gets. Each "skill" has a separate mastery probability, and answers update those probabilities through a hidden-Markov-style transition. ASSISTments was built on this, and most adaptive-learning research papers use BKT variants as baselines. The diagnostic granularity is better — you can at least tell which skill is failing — but the model is still discrete, the updates are coarse, and the state doesn't capture confidence calibration, response time, or any structured signal beyond binary correctness.
Deep Knowledge Tracing (DKT) and successors. Neural-network approaches (DKT, SAKT, SAINT, MonaCoBERT) replace the hand-built BKT mechanics with learned representations. These are genuinely adaptive in the technical sense — they meet most of the four criteria — and they're the state of the art in academic research. But they're almost entirely missing from production deployment in commercial edtech, for two reasons: the inference cost is non-trivial, and the architectures don't easily extend to new domains without retraining.
Everything else — Coursera, Udemy, Pluralsight, Quizlet, Brilliant, MasterClass — has no adaptive engine at all. It's static content with progress tracking. The "personalization" is at best a recommender pulling from "people who watched this also watched that."
Why This Matters
Two reasons. First, the user-experience consequences. A learner studying for the AWS Solutions Architect Associate exam through a platform with a one-dimensional difficulty slider is not getting personalized study. They're getting questions of roughly-correct difficulty from a fixed pool, with no awareness of why they got the previous question wrong. If they failed a VPC peering question because they don't actually understand IP subnetting, the system has no idea — and the next question won't address the gap. The learner spends three months in a study loop that never converges, fails the exam, blames themselves, and tells everyone they know the material was just too hard.
Second, the market consequences. The category is being commoditized at the marketing layer faster than it's being built at the engineering layer. Vendors who haven't built a real adaptive engine are now competing for the same word — "adaptive" — that vendors who have built one rely on as a differentiator. This is the same dynamic that destroyed the term "AI" in the 1990s and the term "blockchain" in 2017. By the time the buyer is sophisticated enough to ask hard technical questions, the word has lost all signal value, and the only way to differentiate is on price.
The Test I'd Apply
If you're evaluating an adaptive learning platform — as a learner, an institution, or an investor — the test is to ask three questions:
- What happens to the learner's state model when I get a question wrong by clicking the wrong option in a panic? A real adaptive system should distinguish that from a wrong answer driven by knowledge gap. If the vendor's answer is "it just lowers your difficulty level," walk away.
- What's the per-interaction update latency, end-to-end, measured in production? If they can't tell you, or the number is above 50 milliseconds, the system is not updating in time to influence the next decision. It's batched.
- If I show you two learners with identical recent answer histories but different prior knowledge, will the system show them different next questions? If the answer is no, the state model is shallow enough that it doesn't matter. The "adaptation" is theatre.
Any vendor who can't answer those three questions cleanly is selling a static-content product with the word "adaptive" stapled on for marketing purposes. Most can't. A few can — and those are the platforms worth paying attention to.
Where the Category Goes
The trajectory is going to look like the search-engine consolidation of the late 1990s. There are dozens of vendors today claiming adaptive intelligence. Within five years there will be three or four platforms that actually deliver it at scale, and the rest will be either acquired or commoditized. The technical bar is non-trivial — sub-50-millisecond per-interaction updates, multi-dimensional state, diagnostic granularity at the prerequisite level — and the platforms that have it will eat the platforms that don't, because the user experience gap is unbridgeable.
I've written before about what 10× lower latency buys you in this category. The latency story is the foundation. Everything else — the diagnostic granularity, the cross-domain transfer, the cognitive-state responsiveness — depends on it. Vendors who built on slow batch architectures five years ago can't retrofit. Vendors building now on the right substrate get the whole category.
That's the real story of "adaptive learning" in 2025: a category in the middle of being redefined by the engineering choices vendors make in production, while the marketing layer pretends everything is converging on the same point. It isn't. The gap is widening every quarter, and the platforms still selling static content with adaptive labels are eventually going to find themselves on the wrong side of it.
Part of a series on production adaptive-learning systems. See also Renkara Media Group's patent portfolio.