Reliability & grounding
P10.llm-essentials.03 · Audience: guest, it-ml, language-pro · Prerequisites: Generating text & context
The most dangerous habit these models have is stating things that are simply not true with the same fluent confidence they use for things that are. This is not a glitch to wait out: the model was trained to sound right, not to be right, and it gives no signal when it is guessing. The good news is that the strongest fix is in your hands — stop relying on the model's memory and hand it the real source text, whether you paste it into the prompt yourself or a retrieval system fetches it automatically. This module shows why hallucination happens, why the material you provide beats the model's baked-in knowledge, and how retrieval puts that fix to work at scale.
ⓘ Concept: One mental model to keep
The model is optimised to sound fluent, not to be correct — two different targets. The single most powerful fix is to stop trusting its memory and instead hand it the source text to work from.
Why it matters — Fluent and correct are two different targets; the most powerful fix for reliability is grounding — giving the model the real source text to work from.
HallucinationHallucinationWhen a model states something false as if it were true — a made-up fact, citation, or quote — because it predicts fluent text, not verified truth. Confident wording is no guarantee of accuracy. is when the model produces fluent, confident text that is factually wrong or invented — a fake citation, a plausible but incorrect date, a mistranslated name. It happens because the model generates the most plausible-sounding continuation, not a verified one. When it lacks the real information, it fills the gap with something that fits the pattern — and usually gives no signal that it is unsure.
Worked example — ask for a source, get an invented one
Prompt: “Give me an academic reference for post-editing effort in MT.”
What the model returns
“Krings, H. P. (2001). Repairing Texts. Journal of Translation Studies, 14(3), 221–240. doi:10.1016/j.jts.2001.03.008”
Looks perfect — author, year, journal, page range, even a DOI.
What's actually true
The DOI resolves to nothing; the volume/page numbers are invented. A real Krings book exists (Repairing Texts, 2001) — but this journal citation was assembled to look right, not retrieved.
Nothing inside the model flagged the problem, because a fake reference is still a fluent, plausible continuation. For language work:anything factual — names, numbers, quotes, legal or medical content, citations — must be verified, and grounding the model with real source text (next steps) sharply reduces this in the first place.
⚡ Interview Ref — the quick-scan Reference face
In plain language: why a confident-sounding answer can still be wrong, the difference between what the model "remembers" and what you hand it, and how giving it the source text makes it far more reliable. No maths required.
1 — Why it produces fluent but false statements (l7)
Hallucination = fluent, confident text that is factually wrong or invented, because the model generates the most plausible-sounding continuation, not a verified one — and gives no signal that it is unsure. For language work, anything factual (names, numbers, quotes, legal/medical content, citations) must be verified, and grounding with real source text sharply reduces it.
2 — The model's own knowledge vs what you give it (l13)
The model has two sources: baked-in training knowledge (outdated, fuzzy, incomplete) and what you provide now (documents, glossary, instructions). What you supply in the moment is more reliable, current, and controllable, and the model will favour it if told to — the basis of grounding: for anything current, proprietary, or precise, give the material rather than trusting memory.
3 — Retrieval (RAG) and why it helps (l20)
RAG finds relevant documents before the model answers and places them in the prompt, so it answers from that material. It improves accuracy, keeps information current, and lets you cite sources — shifting the question from "does the model know this?" to the far more controllable "did we retrieve the right source?"
Interview one-liners
- Hallucination = fluent + confident + wrong; the model optimises for plausible, not true, and gives no warning.
- Baked-in memory vs provided text: what you hand it now is more reliable — that is grounding.
- RAG retrieves the right sources into the prompt first, so the question becomes "did we retrieve well?" — verify facts, ground with sources.
📚 Go Further
Plain-language explainers on reliability and grounding.
| Type | Resource |
|---|---|
| Article | Retrieval-Augmented Generation for Knowledge-Intensive NLP — Lewis et al. (2020) |
| Explainer | Survey of Hallucination in Natural Language Generation — Ji et al. (2022) |
| In-app | P11.ir-rag.04 — RAG & groundedness, for the retrieval mechanics behind this reliability fix |
| In-app | P10.llm-essentials.04 — Prompting & fine-tuning, for how to steer and specialise the model reliably |
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
Keeping your files on this device
Off by default. The mentor always gets your file; this only decides whether your own copy stays here. Copies live in this browser only - they do not follow you to another device, and clearing site data removes them.
🎓 Practice ladder
1 graded rung · ~8 minPractise the decision rule. The workspace is three panels — instructions, a code editor, and live output with test results. Run checks the visible tests; Submit grades against hidden edge cases; the reference solution unlocks once you pass.
Explorer rung — Answer or admit you don't know
Loading exercise…
My notes on this module
Loading your notes...
Where next?
Later in LLM Essentials
Go up a level