How AI study tools actually work, and how to spot the fake ones
Three real algorithms sit behind good study tools. Most apps that say 'AI' are using none of them.
Key takeaways
- →Three distinct algorithms do the real work: a scheduler, an ability estimator, and a predictor of what you know now.
- →FSRS schedules review by tracking each item's difficulty, stability and retrievability.
- →Item Response Theory explains why two students with identical scores can have different measured ability.
- →A large language model that explains an answer is not the same as an engine that decides what you see next.
When a study app says 'AI', it is usually doing one of three quite different things, or claiming to do all three while doing none. The three are worth separating, because only two of them affect what you actually study.
1. The scheduler: when should I see this again?
Memory decays on a curve that is predictable enough to compute. The job of a scheduler is to show you a fact at the point where recalling it is still possible but no longer easy, because that is where a single successful recall strengthens memory most.
FSRS, the Free Spaced Repetition Scheduler, is the current open standard and is what modern Anki uses. It tracks three numbers per item: a difficulty, a stability measured in days, and a retrievability that falls as time passes since the last review. The interval to the next review comes out of those numbers rather than out of a fixed rule like 'double it each time'.
This is the easiest capability to verify. Ask which scheduler an app uses. Real answers sound like 'FSRS' or 'SM-2'. Non-answers sound like 'our AI learns your pace'.
2. The ability estimator: how good am I, really?
A percentage tells you how you did on the questions you happened to be asked. Ask an easier set and the number rises without you having improved. Item Response Theory fixes this by giving each question its own parameters, typically a difficulty and a discrimination, and then estimating your underlying ability from the pattern of what you got right.
The consequence is counterintuitive and important: two students with the same raw score can have different ability estimates. Answering a hard, sharply discriminating question correctly is stronger evidence than answering an easy one correctly, and the model weights it accordingly.
This is the method serious testing bodies use for exactly this reason. It is also, in a survey of 474 apps on the Indian App Store in September 2026, mentioned by none of them.
3. The predictor: what do I know right now?
Knowledge tracing predicts the probability you would answer a given concept correctly at this moment, based on the sequence of everything you have answered previously. Modern implementations use transformers over your concept and response history, which means the order and recency of your mistakes carry information instead of being flattened into an average.
This is what should decide the next question in a study session, as opposed to inside a timed adaptive exam, where the better rule is to pick whichever question would most reduce uncertainty about your ability.
Where language models fit, and where they do not
Large language models are genuinely useful for generating questions, explaining a wrong answer, and answering a doubt at one in the morning when no teacher is awake. Those are real benefits and worth having.
They are not ability estimators and they are not schedulers. A chatbot bolted onto a question bank is a useful feature, not an adaptive engine. When an app's entire AI story is a chat interface, it is doing the third thing on this list and none of the first two.
Four questions that separate real from decorative
- Which scheduling algorithm decides my review intervals?
- How is my ability estimated, and is it different from my percentage score?
- What decides the next question, and is that rule different inside a timed test?
- Where can I read about this? A method that cannot be looked up cannot be evaluated.
Frequently asked questions
How do AI study apps decide what to show me next?
Well-built ones use one of two mechanisms depending on context. In a study session, a knowledge-tracing model predicts which concepts you are most likely to get wrong right now and surfaces those. Inside a timed adaptive test, the better rule is to choose the question that would most reduce uncertainty about your ability, which in Item Response Theory means the item with the highest Fisher information at your current estimate.
What is FSRS?
FSRS, the Free Spaced Repetition Scheduler, is an open algorithm that decides when you should next review something. It tracks a difficulty, a stability in days, and a retrievability for each item, and schedules the review for the point where recall is still possible but no longer easy. It is what modern Anki uses and it is freely published, so any app using it can say so.
What is Item Response Theory?
A statistical method for estimating a person's ability from their answers, where every question carries its own difficulty and discrimination parameters. It is why two students with the same raw score can receive different ability estimates: getting a hard, highly discriminating question right is worth more evidence than getting an easy one right. Serious testing organisations use it for precisely this reason.
Is ChatGPT enough to prepare for NEET?
It is genuinely useful for explanation and for answering doubts on demand, and it will confidently produce wrong answers often enough that you must verify anything factual against NCERT. What it does not do is track what you have forgotten or schedule your review, which is the part of preparation that most affects your score. Use it as a tutor for doubts, not as a study plan.