The specific failure citations prevent
A general chatbot answering a question about your module is answering from its training data, which contains the internet's aggregate version of the topic. That is often right. Where it's wrong is subtle and expensive: your lecturer's preferred definition, the notation your department uses, the three of five causes your syllabus lists, the simplifying assumption your course makes and the wider literature doesn't.
None of those errors look like errors. They look like fluent, confident, correct-sounding paragraphs, and you will write them in an exam and lose marks with no idea why. This is different from the crude hallucination problem of invented facts — it's the quieter version, and citations to your own material are the only thing that catches it.
What counts as a real citation
| Behaviour | Real grounding? | Why |
|---|---|---|
| Links each claim to a page in a document you uploaded | Yes | You can verify in seconds. This is the standard. |
| Names the document but not the location | Partly | Better than nothing; verification takes minutes, so nobody does it. |
| Cites web pages it searched | No, for coursework | Grounded in the internet, not in your syllabus. Useful for other things. |
| Produces a bibliography of plausible-looking references | No — actively dangerous | The classic fabricated citation failure. |
| Says "your material doesn't cover this" | Yes — the strongest signal there is | A tool that can refuse has an actual boundary. One that always answers doesn't. |
That last row is the one to test first. A grounded system knows the difference between what your documents contain and what it happens to know, and will tell you when you've asked about something outside them. A system that cheerfully answers anything is not consulting your material at all — it's decorating a general answer with the vocabulary of your upload.
How to test a tool in five minutes
- 1
Upload one document and ask something it definitely doesn't cover
A related topic from a different week. A grounded tool says the material doesn't cover it. An ungrounded one writes you a confident essay. This single test eliminates most of the market.
- 2
Ask a question where your course differs from the standard treatment
Every course has one — a non-standard notation, a definition your department argues about, a model your lecturer rejects. Check which version comes back.
- 3
Follow one citation and check the page really says it
Not that the page is about the topic — that the specific claim is on it. Citation theatre, where the link is real and the claim isn't on that page, is common enough to be worth catching.
- 4
Ask the same question twice in different words
Grounded answers converge because they come from the same passage. Ungenerated ones drift, because they're being composed fresh each time.
- 5
Ask it to quote the exact sentence it used
A retrieval-based system can. A system reconstructing from memory produces a paraphrase and calls it a quote — which is the same failure mode as a fabricated reference, just smaller.
How grounding actually works
The mechanism is retrieval-augmented generation: your documents are split into passages and indexed; your question is used to find the most relevant passages; only those passages plus your question are sent to the model, with an instruction to answer from them and say so if they're insufficient. The model isn't recalling your syllabus — it's reading it, each time, in front of you.
That's why the citation is available at all: the system knows which passages it sent. If a tool can't tell you where an answer came from, it's usually because there was no retrieval step and nothing to point at. RAG explained for students covers this in more depth, including where it still goes wrong.
Where grounding still fails
Honesty matters more than marketing here, so: retrieval is not a guarantee of correctness. The failure modes that remain are real.
- Retrieval misses. If the relevant passage isn't found, the answer is built from the wrong pages and cited confidently to them.
- Your source is wrong. A grounded tool faithfully reproduces the error in your lecture notes. Grounding guarantees provenance, not truth.
- Synthesis across passages can introduce a connection neither passage made. Citations show you both sources and not the leap between them.
- Scanned documents that OCR badly produce garbled retrieval, and the citation will point at a page whose extracted text was nonsense.
What grounding buys is not infallibility. It's that every one of those failures is *detectable in four seconds*, because there's a page number to look at. An ungrounded tool's failures are undetectable by construction.
What to use ungrounded AI for
General models are excellent at things where your syllabus isn't the authority: rephrasing an explanation you didn't follow, generating analogies, drafting practice questions in a style you specify, or checking whether your written argument is coherent. The rule is simple — use ungrounded AI for form, grounded AI for facts. Thaxas vs ChatGPT sets out where each is genuinely better.