Five reasons the answers differ
| Cause | What it looks like | Who to follow |
|---|---|---|
| Convention | Different sign convention, notation, or naming for the same thing | Your lecturer, always — conventions are local |
| Scope | The model gives the general case; your course covers a restricted one | Your lecturer, but understand the general case too |
| Level | A graduate-level treatment of an undergraduate topic | Your lecturer for the exam; the model for insight |
| Regional or disciplinary variation | Different jurisdiction, different clinical guideline, different school of thought | Your lecturer — the exam is set in one context |
| The model is simply wrong | A confident, specific claim that isn't in any source | Your lecturer, obviously |
Only the last row is a hallucination. The other four are real differences between a general-purpose answer and the answer your course is teaching, and treating all disagreements as model errors means missing what they're telling you.
The examinable answer is your lecturer's
This is the rule that resolves nearly every case, and it's not about who's cleverer. Your lecturer wrote the exam, chose the syllabus and — often — wrote the mark scheme. An answer that's correct in the field's general terms but uses different notation, a different classification or a different framework than the one taught will lose marks even when nothing in it is false.
So: learn your course's version for the exam, and use the model's version to understand why your course made the choice it did. The second part is genuinely valuable — knowing that your lecturer's framework is one of three tells you where its edges are.
How to tell a convention difference from an error
- 1
Ask whether both can be true at once
If the two answers are the same content with different labels, signs or ordering, it's convention. If one says a mechanism happens and the other says it doesn't, it's a factual conflict.
- 2
Ask the model directly about the discrepancy
"My course defines this differently — is there more than one convention here?" This resolves a large share of cases in one message, and the answer usually names both conventions.
- 3
Check the textbook your course actually set
The set text almost always sides with your lecturer, because that's usually why it was set. Two sources agreeing settles it.
- 4
Look for the scope condition
Many apparent conflicts dissolve into "true in general, but not under the assumption your course makes". Finding the missing assumption is often the most useful thing in the exchange.
- 5
If it's still unresolved, ask in office hours
"I've seen this presented two ways — which does this course use?" is a good question and staff answer it happily. See office hours.
Where models are most and least reliable
- Most reliable: stable, widely-documented material — standard derivations, well-established mechanisms, textbook definitions, code and maths that can be checked.
- Less reliable: anything recent, contested, or specific to a jurisdiction, institution or guideline version. Clinical guidelines and legal rules change and models are trained on a mixture of versions.
- Least reliable: specific numbers, citations, dates and attributions — the classic fabricated citation failure, and the one that costs marks most directly.
- Unknowable to it: your lecturer's emphasis, your department's convention, and what's actually on your paper. No model has that, and it's often what the difference is about.
Grounding removes most of it
A model answering from general training will give the field's general answer. A model answering from your uploaded lecture notes will give your course's answer, because that's the material in front of it — which is the whole argument for grounded AI in a study context.
This changes the failure mode usefully: instead of "is this true in general?", the question becomes "does the source say this?", which you can check in seconds against a cited page. It doesn't eliminate error, but it makes error findable.
The disagreement is worth something
Two credible sources disagreeing is a good prompt. Working out why they differ forces you to identify the assumption, scope condition or convention that separates them — and that's precisely the kind of understanding that distinguishes a strong exam answer from a memorised one.
In several subjects, writing "under convention X this is positive; the course uses Y, hence the sign here" is a mark-scoring sentence. Noticing the conflict is what gets you there, so don't just resolve it and move on — elaborative interrogation is exactly this move made deliberate.
When your lecturer is actually wrong
It happens — a slide with a typo, an out-of-date guideline, a genuine slip. If you have a well-sourced reason to think so, the move is to ask rather than to assume, and to ask about the specific claim rather than to open with the correction.
"I found this stated differently in the current guideline — am I misreading the slide?" gets a real answer. And for the exam, write what the course taught unless you've had it confirmed, because the mark scheme was written from the same slide.