AI-Adaptive Learning Games: How the Algorithm Decides What Your Kid Practices Next

Your kid has been on Prodigy Math for 45 minutes and hasn’t complained once. Either something is wrong or something is very right. What’s happening inside the app matters, because the question isn’t whether your child is engaged — it’s whether the engagement is producing learning. Adaptive learning games are built on a genuinely sophisticated idea: that software can continuously measure what a learner knows, identify what they’re ready to learn next, and adjust difficulty in real time. When it works, it’s powerful. When it’s poorly calibrated or used without teacher scaffolding, it produces 45 minutes of on-level busywork that feels like progress but isn’t.

Key Takeaways

• Adaptive learning algorithms use a concept called Knowledge Tracing — estimating what a student knows based on a pattern of right and wrong answers — to select the next question or activity.
• The most robust research on adaptive learning (VanLehn’s 2011 meta-analysis) found intelligent tutoring systems produce learning gains similar to one-on-one human tutoring — a finding that holds up in replications for structured domains like math.
• RAND’s 2016 study of personalized learning schools found meaningful academic gains when adaptive software was combined with teacher instruction, but limited gains when software was used as a primary replacement.
• Duolingo’s spaced repetition algorithm is well-validated for vocabulary retention; its effectiveness for grammar and conversational fluency is less supported.
• Parents can evaluate adaptive game quality by checking whether difficulty actually changes in response to repeated errors — not just session by session, but within a single session.

How Adaptive Learning Algorithms Work

The core concept in adaptive learning is Knowledge Tracing (KT) — a model of what a student knows at any given moment, updated after every response.

The original formulation, Bayesian Knowledge Estimation (BKT), treats each knowledge component (e.g., “adding fractions with unlike denominators”) as either known or not known, with a probability attached. After each question:

• A correct answer increases the probability that the student knows the concept (but accounts for the possibility they guessed).
• An incorrect answer decreases the probability (but accounts for the possibility they knew it but made a careless error).

More recently, Deep Knowledge Tracing (DKT), using recurrent neural networks, has replaced BKT in many platforms. DKT can model dependencies between knowledge components — recognizing, for example, that a student struggling with fraction addition is likely also struggling with equivalent fractions, and should practice those first.

After the system estimates what a student knows, it uses a second algorithm to decide what to present next. Most platforms use a version of item selection that targets concepts where the student is estimated to have ~70–80% mastery probability — the “zone of proximal development” (Vygotsky’s concept, operationalized computationally). Too easy: bored. Too hard: frustrated and disengaged. The 70-80% range produces the most learning per item.

Spaced repetition is a related technique used heavily in Duolingo, Anki, and vocabulary-focused apps. It schedules review of previously learned items at increasing intervals, exploiting the finding that distributed practice produces more durable retention than massed practice. The classic implementation is the SM-2 algorithm, developed by Piotr Woźniak in 1987 and still used (with modifications) in modern apps.

The Research on Whether Adaptive Learning Works

The research base is more specific — and more conditional — than most platform marketing implies.

VanLehn’s 2011 meta-analysis is the most-cited review of intelligent tutoring systems (ITS), the research category that adaptive learning games derive from. VanLehn analyzed 62 studies comparing ITS to classroom instruction and found effect sizes averaging 0.76 standard deviations — close to Bloom’s famous “2-sigma” finding about one-on-one human tutoring. This is a strong result, and it’s held up in subsequent research. The important caveat: most of these studies involved structured, well-defined domains (algebra, physics problem-solving) with high-quality ITS platforms developed in research settings, not commercial games.

RAND Corporation’s 2016 study of personalized learning schools (Pane et al.) followed 11,000 students in 62 schools using personalized learning approaches including adaptive software. Students in these schools outperformed a matched comparison group by 0.07 standard deviations in math in year one and 0.31 standard deviations by year two. But the analysis found significant heterogeneity: schools where teachers maintained a strong instructional role alongside software showed the most gains. Schools that used software as a primary delivery mechanism showed smaller effects.

Duolingo’s internal research (published on their research blog, with some peer-reviewed publications) shows strong effects for vocabulary retention — which is what spaced repetition should do well. Evidence for grammar learning and conversational skill development is thinner and more contested in the independent literature.

How Major Platforms Compare

Platform	Primary Approach	Strongest Evidence	Weakest Evidence	Best For
Khan Academy (Khanmigo)	Mastery-based, Bayesian KT	Math skill development	Writing, critical thinking	Structured math remediation
Duolingo	Spaced repetition, hearts system	Vocabulary retention	Conversational fluency, grammar	Language vocabulary building
Prodigy Math	Game-embedded adaptive math	Math engagement	Conceptual depth	Math practice, gamified reinforcement
IXL Learning	Mastery scoring, adaptive difficulty	Skill fluency	Open-ended reasoning	Skill drill with detailed diagnostics
DreamBox	Adaptive math, manipulative-based	Early math (K-8) concepts	Upper grade math	Elementary math concept building
Reading Eggs	Phonics-based adaptive reading	Early reading phonics	Reading comprehension depth	Pre-K to early elementary reading

The Zone of Proximal Development, Computationally

Vygotsky introduced the Zone of Proximal Development (ZPD) in the 1930s: the space between what a learner can do independently and what they can do with support. The best learning happens in the ZPD — challenging enough to require effort, supported enough to enable success.

Adaptive algorithms operationalize this by targeting items at the mastery-probability threshold where learning rate is maximized. Research by Lindsey et al. (2014) on memory strengthening algorithms found that items reviewed at ~85% expected recall probability produced better long-term retention than items at higher or lower probability — supporting the general principle.

The practical implication for parents: if your child is getting every question right, the algorithm may not be challenging them enough. If they’re getting fewer than half right, they may be outside the ZPD entirely and need human scaffolding. The sweet spot — noticeable effort, frequent (but not constant) success — is what you’re looking for.

How to Teach Your Kid About Adaptive Learning Algorithms

Ages 5–8: Notice When It Gets Harder

This age group can understand the basic concept: “The game is watching what you get right and wrong, and it changes the questions to be just right for you.” Play alongside your child for 10 minutes. When they get three in a row correct, ask: “Do you notice the questions getting harder?” When they miss two, ask: “Does it feel like it’s getting a little easier?”

This isn’t about understanding algorithms — it’s about developing the metacognitive habit of noticing their own learning challenge level. That habit is more durable than the app.

The question to ask: “How hard does this feel right now — too easy, just right, or too hard?”

Ages 9–12: Deliberately Fail and Watch the Response

Here’s an experiment: play a session of any adaptive learning app normally. Note the difficulty level. Then, in the next session, deliberately get 5–10 questions wrong in a row. Watch what happens. Does the difficulty drop? How quickly? What types of questions does the algorithm introduce?

Then return to normal play and observe the recovery. This teaches kids that the algorithm is tracking them and responding — it’s not random, and it’s not fixed. Understanding this changes how they engage with it (more honest effort, less strategic guessing).

The question to ask: “What did the algorithm ‘think’ about your knowledge when you got those wrong? How do you know?”

Ages 13+: Build a Flashcard System Using SM-2

The SM-2 spaced repetition algorithm is simple enough to implement in a spreadsheet or short Python script — and doing so teaches the core concept more deeply than any tutorial. Here’s the basic rule: after reviewing a card, rate your recall 0–5. Cards rated 0–2 are reset to review tomorrow. Cards rated 3–5 get an increasing interval: first review at 1 day, then 6 days, then multiplied by a factor based on your performance (typically 1.3–2.5).

Build this in Google Sheets: columns for card, last review date, next review date, interval, ease factor. Use a formula to calculate next review date based on rating. This is functionally what Anki’s algorithm does — and building it yourself demystifies the technology.

The question to ask: “What would happen to your review schedule if you always rated yourself a 5? What if you always rated yourself a 2? What’s the optimal honest strategy?”

What to Watch For Over 3 Months

Month 1: Establish a baseline. Track actual scores or progress metrics within the platform for the first two weeks before drawing conclusions about effectiveness. Most adaptive platforms provide some form of parent or teacher dashboard — learn to read it.

Month 2: Check whether the platform has moved your child to new content or whether they’ve been cycling through the same skill level. Cycling can indicate either mastery consolidation (fine) or a calibration problem where the algorithm isn’t advancing them appropriately.

Month 3: The real test: does performance on external assessments (school tests, state assessments) correlate with platform improvement? If your child’s Prodigy score is advancing but their classroom math grades aren’t improving, the platform may be measuring something different than what the classroom requires — or there’s a skill transfer gap that needs teacher support.

Red flag: a child who has learned to game the adaptive system rather than actually learn — strategically wrong answers to reduce difficulty, rapid clicking to get through items without thinking. This is a sign the gamification incentives (points, badges, game currency) are dominating over learning motivation. A brief reset of rewards, or a conversation about the purpose of the platform, usually addresses it.

Frequently Asked Questions

Is Prodigy Math actually good for learning, or is it just engagement?

Both, in different measure. Prodigy’s adaptive math engine targets grade-level standards, and the problems themselves are educationally sound. The engagement mechanism is effective — kids play longer than with traditional drill. The critique from many educators is that the game rewards currency system can lead to disengaged clicking rather than careful thinking. The platform works best when parents or teachers review the skill reports and target specific areas rather than using it as unsupervised entertainment.

How is adaptive learning different from just making a test harder if you do well?

Adaptive testing (like computer-adaptive standardized tests) adjusts difficulty to measure precisely where your performance level is. Adaptive learning goes further: it tracks which specific knowledge components are mastered and not mastered, sequences new material based on prerequisite knowledge, and schedules review for material that’s at risk of being forgotten. The goal is not just assessment but maximizing the rate of learning per unit of time.

Should my child use these platforms instead of homework?

Adaptive learning platforms are best used as supplements to classroom instruction, not replacements. Research consistently shows the largest effects when software is combined with teacher instruction and feedback. Using platforms instead of assigned homework is a decision that should involve the teacher — some platforms produce work that teachers want to see, and some schools have specific policies about AI or adaptive software use for assignments.

Can adaptive learning help a child who is significantly behind grade level?

Yes — this is actually where adaptive learning shows some of its strongest results. Because the algorithm can place a child at their actual mastery level (not their grade level) and build systematically, a struggling learner can work through prerequisite skills in sequence without the stigma of being placed in a lower-grade-level classroom. The research on remediation uses of ITS is particularly strong.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. VanLehn, K. (2011). “The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems.” Educational Psychologist, 46(4), 197–221. https://doi.org/10.1080/00461520.2011.611369
2. Pane, J. F., Steiner, E. D., Baird, M. D., & Hamilton, L. S. (2016). Informing Progress: Insights on Personalized Learning Implementation and Effects. RAND Corporation. https://www.rand.org/pubs/research_reports/RR1365.html
3. Corbett, A. T., & Anderson, J. R. (1994). “Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge.” User Modeling and User-Adapted Interaction, 4(4), 253–278.
4. Piech, C., Bassen, J., Huang, J., et al. (2015). “Deep Knowledge Tracing.” NeurIPS 2015. https://arxiv.org/abs/1506.05908
5. Lindsey, R. V., Shroyer, J. D., Pashler, H., & Mozer, M. C. (2014). “Improving Students’ Long-Term Knowledge Retention Through Personalized Review.” Psychological Science, 25(3), 639–647. https://doi.org/10.1177/0956797613504302
6. Duolingo Research Team. (2023). “Research Insights.” https://research.duolingo.com
7. Woźniak, P. A. (1990). “Optimization of Learning.” SuperMemo Library. https://www.supermemo.com/en/blog/history-of-spaced-repetition