A general-purpose assistant answers a NEET PG question from everything it has ever read: textbooks, forum posts, coaching notes, and a great deal of material that was wrong when it was written. It has no way to tell you which of those it leaned on, and no obligation to be consistent between one answer and the next.
A grounded tutor is given a much smaller job. It receives one verified question, its keyed answer, the editorial explanation a person wrote, and the answer you chose. Everything it says has to come from that. It is not being asked what the answer is; it is being asked to explain an answer that has already been settled.
Why this matters more in medicine than elsewhere
In most subjects a plausible-but-wrong explanation is an inconvenience. In medicine it is a memory you will carry into a ward. The failure mode is not that the model says something obviously absurd — you would catch that. It is that it produces a clean, well-structured paragraph containing one inverted relationship, and you learn the inversion because everything around it was correct.
Candidates are also the worst-placed people to audit it. You are asking about a topic precisely because you do not know it well. If you could reliably spot the error, you would not have needed the explanation.
You are least able to catch a mistake in exactly the topic you had to ask about. That is the whole argument for grounding.
What grounding looks like in practice
In Recallab, the Mentor is handed a snapshot of the question as it was answered — the stem, the options, the key, the explanation, and the option you actually picked. It can rephrase that material, take a different angle on it, ask you a question back, or explain why a particular distractor tempted you. What it cannot do is decide the answer is something else, because the answer is not part of its job.
The snapshot matters more than it sounds. It records the question as it stood at the moment you answered, so a later correction to the bank cannot silently change the meaning of a conversation you already had.
The honest limits
- It cannot help with a question outside the bank. Grounding is a constraint, not a convenience — no source, no answer.
- It is not a second opinion on the key. If you think a keyed answer is wrong, that is an editorial matter for a human, and it should be.
- It does not make a weak explanation strong. If the underlying editorial work is thin, a tutor built on it will be thin too. The bank is still the product.
A concrete way to test any tool yourself
Take a question you know cold, with a distractor you understand well. Ask the tool why that distractor is wrong. You are not testing whether it gets the answer right — you already know it does. You are testing whether its reasoning matches the actual reason, or whether it has produced a plausible-sounding alternative that happens to reach the same conclusion.
Do this three or four times on material you are confident about. The pattern tells you far more than any amount of asking about topics you cannot check.
Why "it cited a textbook" is not reassurance
A model asked to include a reference will include one. Whether the cited page actually says what the answer claims is a separate question, and the citation being plausible is exactly what makes it hard to doubt. A reference attached to a generated answer is decoration; a reference attached by the person who wrote the explanation is provenance. They look identical on screen.
A citation you cannot check is a claim wearing the costume of evidence.
Grounding is a product decision, not a model one
This is worth saying plainly because it is often obscured: the difference between a grounded tutor and a chatbot is not a better model. It is a choice to restrict what the model may draw on, made by whoever built the product, and usually at the cost of the feature seeming less capable. A tool that answers anything will always demo better than one that says "that question is not in the bank".
Which is why the constraint has to be checked rather than assumed. Ask a tool something outside its content and see whether it declines or obliges.
So is a general chatbot useless?
No — it is useful for the things where being wrong is cheap. Turning a paragraph into a table, drafting a mnemonic you will then check, rephrasing a definition you already understand. Those are low-stakes tasks where you can verify the output at a glance.
The moment the task is "tell me which of these four is correct and why", the stakes change, and the only version worth trusting is one whose source you can point at.