Ungrounded versus grounded, concretely
| Ungrounded | Grounded | |
|---|---|---|
| Source of the answer | Patterns learned during training | Passages retrieved from your documents |
| Can cite a location | No — there's nothing to point at | Yes, the page it retrieved |
| Behaviour outside its knowledge | Produces something plausible | Says the material doesn't cover it |
| Whose version of the topic | The internet's aggregate | Your lecturer's |
| How you verify | Re-derive it yourself | Click the citation |
| Consistency across rephrasings | Drifts — recomposed each time | Converges — same passages retrieved |
How it works, briefly
Your documents are split into passages and indexed by meaning rather than by keyword. When you ask something, the system searches that index, pulls the handful of most relevant passages, and sends them to the model along with your question and an instruction to answer from those passages and to say so if they're insufficient.
The model isn't recalling your syllabus — it's reading it, freshly, each time. That's why a citation is available: the system knows exactly which passages it supplied. A tool that can't tell you where an answer came from usually has no retrieval step at all. RAG explained for students covers the mechanics in more depth.
What grounding fixes
- Course-specific divergence. Your lecturer's definition, your department's notation, the three of five causes your syllabus lists. This is the most valuable and least discussed benefit.
- Fabrication. Not eliminated, but massively reduced — the model has the text in front of it rather than reconstructing from memory.
- Undetectable errors. A cited error can be caught in four seconds. An uncited one is invisible by construction.
- The confident gap-fill. A grounded system can decline, which is the single most useful behaviour a study tool has.
- Staleness. Your material is your material, regardless of when the model was trained.
What it doesn't fix
Worth being precise, because grounding is often marketed as if it made a tool infallible.
- Retrieval misses. If the search doesn't surface the right passage, the answer is built from the wrong pages and cited confidently to them. This is the commonest real failure.
- Wrong sources. A grounded system faithfully reproduces the error in your notes. It guarantees provenance, not truth.
- Synthesis leaps. Combining two passages can produce a connection neither made. The citations show both sources and not the inference between them.
- Bad extraction. A scanned PDF that OCRs poorly gets indexed as garbage, and the citation points at a page whose extracted text was nonsense.
- Citation theatre. Some systems cite a page that's merely on-topic rather than the page containing the claim. Spot-check this.
Testing whether a tool is genuinely grounded
- 1
Ask about something your upload definitely doesn't cover
A related topic from another week. A grounded system says so. An ungrounded one writes a confident answer. This single test separates most of the market.
- 2
Ask where your course diverges from the standard treatment
Every course has something — a notation, a contested definition, a model your lecturer rejects. Check which version comes back.
- 3
Follow a citation and verify the specific claim
Not that the page is about the topic — that the sentence supporting the claim is on it.
- 4
Ask the same thing twice, worded differently
Grounded answers converge because they retrieve the same passages. Ungrounded ones drift.
- 5
Ask for the exact quote it used
A retrieval system can produce it. A reconstructing system paraphrases and calls it a quote.
Grounded isn't always what you want
There are jobs where restricting the model to your documents is a limitation rather than a feature, and it's worth knowing which is which.
| Task | Use |
|---|---|
| "Explain this the way you'd explain it to a child" | Ungrounded — you want a different framing, not your notes back |
| "What did my lecturer say the three conditions were?" | Grounded — this is exactly the case |
| "Give me an analogy for this concept" | Ungrounded — analogies aren't in your syllabus |
| "Generate practice questions on this module" | Grounded — otherwise you practise someone else's syllabus |
| "Is my argument coherent?" | Ungrounded is fine — this is about form |
| "Summarise this chapter" | Grounded — and check the citations |
The rule that holds: ungrounded for form and framing, grounded for facts. Thaxas vs ChatGPT sets out where each is genuinely better rather than pretending one dominates.
Why it matters more for some subjects
Grounding matters most where the cost of a confident error is high and where courses diverge from the general treatment. Clinical subjects, law, and anything with jurisdiction- or department-specific conventions sit at that end — a fluent, plausible, wrong answer about a drug dose or a legal test is genuinely dangerous, not merely inconvenient.
It matters least for well-standardised introductory material where the internet's version and your lecturer's are the same. Even there, though, the ability to check is worth having — see AI study tools that cite sources for the practical argument, and AI and medicine's hallucination risk for the high-stakes case.