Ask a general-purpose chatbot a NEET PG question and you will usually get a fluent answer. Sometimes it is right. Occasionally it is confidently wrong in a way that is hard to detect, because the prose is just as smooth either way. For an examination where a wrong answer costs a mark and a blank costs nothing, fluency you cannot audit is a liability rather than an asset.
That does not make AI useless for preparation. It means the useful version is not the one being advertised. The gains come from a narrow, unglamorous set of jobs that sit around a curated question bank rather than replacing it.
What genuinely helps
Four things move marks, and none of them require the model to know medicine on its own.
- Reading the answer you actually gave. A standard explanation tells you why the key is correct. It rarely tells you why the specific distractor you chose was attractive, which is the thing you got wrong.
- Naming the missing idea. Two candidates can miss the same question for completely different reasons. Being told which of those two you are is worth more than another paragraph of correct explanation.
- Deciding when you see it again. Knowing a fact today says little about knowing it in March. Scheduling is arithmetic, and it is exactly the sort of bookkeeping a person is bad at and software is good at.
- Noticing patterns you cannot see. Over a few thousand attempts, the topics you quietly avoid are visible in the data long before they are obvious to you.
What does not help
Generating new questions is the obvious temptation and the clearest failure. A NEET PG question is a piece of editorial work: a stem that is unambiguous, four options where the distractors are plausible for real clinical reasons, a key that survives scrutiny, and a reference to a current edition. Generated items look like that from a distance and fall apart on inspection — two defensible answers, or a distractor no examiner would write.
Free-text summarising is the second trap. A model asked to summarise a topic produces something that reads well and quietly omits the exception that the examiner is actually testing. The value in a good explanation is often precisely the caveat a summary drops.
The question is not whether the model is clever. It is whether anything it says can be traced back to a source a person checked.
Grounding is the whole design
The distinction that matters is whether the AI is answering from its own knowledge or from an approved source. A tutor grounded in a verified question, its key and its editorial explanation is doing something modest: rephrasing, re-angling and questioning material that a human already checked. It cannot invent a new answer, because it was never asked to produce one.
That constraint is what makes the feature safe enough to put in front of someone revising for a licensing examination. It also makes it less impressive in a demonstration, which is probably why it is not the version usually sold.
The three questions to ask of any AI study tool
Marketing in this space is uniformly confident, so the useful filter is not what a product claims but what it can answer about itself.
- Where did this answer come from? If the tool cannot point at a specific verified question and explanation, it is generating, and you are auditing medicine you do not know.
- What happens when it does not know? A grounded system should refuse. A generative one will produce something fluent, which is worse than refusing.
- Who checked the underlying content? An AI layer inherits the quality of what sits beneath it. A tutor on a badly written bank is a well-spoken guide to wrong material.
The cost nobody mentions: time
Every AI interaction takes time you were going to spend on questions. A three-exchange Socratic dialogue might take four minutes; four minutes is three more questions attempted. Neither is obviously better, but the trade is real and it is rarely made deliberately.
The rule that survives contact with a real timetable: use the Mentor on questions you got wrong and care about, not on everything. Depth where you are failing, coverage everywhere else.
What to expect over a few months
AI will not change your percentile by itself. What it can plausibly do is remove a specific category of waste — re-reading explanations that never landed, revising topics you already know, carrying a confident error for four months without noticing. Those are real losses, and recovering them is worth marks. It is a smaller claim than the advertising makes, and it has the advantage of being true.
Where Recallab has got to
We build the bank first and the layer second. Every question is written or repaired by hand, keyed, explained option by option, and referenced to a current edition — 14,309 of them so far across 20 subjects, and almost all illustrated. The practice apps that read your own attempt history are live to every account. The Mentor, which answers questions about a specific attempt is running for a small pilot group while we work out how often it is genuinely useful rather than merely available.
The rest of these guides go through the pieces one at a time: what your wrong answers reveal, why scheduling beats re-reading, how the exam's own marking scheme should change your behaviour, and what an AI layer can honestly contribute to each.