0.4Strengths, limits: hallucinations, bias, freshness
The essentials in 30 seconds
- The costliest mistake is believing a false answer: hallucination arrives in the same confident tone as right answers — undetectable without verification.
- Three risks to know: hallucinations (invented facts), bias (distortions inherited from data), freshness (time-limited knowledge).
- Verification proportionate to the stakes: internal draft = quick review; binding document = systematic verification of every fact.
Threaded mini-case — Amine catches a hallucination
Amine, project manager in Algiers, asks his assistant for three book references on a specialised topic of his profession, with author and year. Checking each reference on a search engine, he finds one invented — yet presented with confidence. He re-runs the request allowing “I don’t know”, then tests the month’s news: without web search, the answer smells of memory. His rule is born: the higher the stakes, the heavier the verification.
Standalone case for this section: it returns in the lesson, the diagram, the script and the 48h challenge of section 0.4.
1. The costliest mistake
The costliest mistake with AI is not poor prompting: it is believing a false answer. LLMs produce hallucinations — invented figures, non-existent references, plausible but false claims — delivered in the same confident tone as right answers. Added to this are biases inherited from training data (stereotypes, over-represented viewpoints) and limited freshness (knowledge cut-off date).
2. Spot and reduce
High-risk situations: precise facts, figures, quotes, law, current events, people. Reduction techniques: ask for sources, allow “I don’t know” (“only cite what you are sure of”), cross-check two tools, observe whether the assistant really searches the web (sources cited) or answers from memory (be wary). Amine saw the effect: the caution instruction makes the answer markedly more honest.
3. Verify according to the stakes
The governance principle: the higher the stakes, the heavier the verification (internal draft ≠ published document). A provided link proves nothing until you have clicked it: memory-generated links may be broken; only the page content counts. And model progress will not remove the risk: serious professional uses build human verification in by design.
🔎 Concrete example — 🔎 Concrete example — Catching a hallucination
“3 book references on [specialised topic], author + year” → verify each one.
At least one will often be inaccurate or invented.
→ adding “if you are not sure, say so” markedly reduces inventions.Reading: Amine checks his three references: the invention is obvious once tested against the search engine.
Lesson: verify facts, figures and sources according to the stakes — and test whether the tool searches the web or answers from memory.
Section vocabulary
- Hallucination
- Invented content presented as fact: fictitious references, false figures, imaginary events — undetectable by tone.
Simple example: Amine receives a cited book with author and year that does not exist.
What it is not: a display bug or a server outage.
Reflex: verify each reference, figure and quote at the source.
- Bias
- Distortions inherited from training data: stereotypes, over-representation of some viewpoints, cultural angles.
Simple example: Amine notices the suggested examples systematically neglect certain profiles: he asks for other angles.
What it is not: a typo or a calculation error.
Reflex: explicitly ask for other viewpoints on sensitive topics.
- Cut-off date
- The time limit of the model’s knowledge. Assistants with web search work around it — then citing their sources.
Simple example: Amine tests the month’s news: with no web sources cited, he distrusts the answer.
What it is not: the creation date of his account.
Reflex: for current events, require a web search with consultable sources.
- Risk zone
- Requests where hallucination is likely and costly: precise figures, quotes, case law, bibliographic references, current events.
Simple example: Amine classifies his requests: amounts and references in the red zone, internal drafts in the green zone.
What it is not: any harmless request with no stakes.
Reflex: spot the red zone before sending, not after.
- Proportionate verification
- The golden rule: internal email = quick review; binding document = systematic verification of every fact.
Simple example: Amine quickly reviews his internal draft but checks every fact of the document for the client.
What it is not: verifying every inconsequential draft with the same heaviness.
Reflex: calibrate verification to the stakes of sharing.
In practice — In practice — Provoke (and catch) a hallucination
- Ask your assistant: “Give me 3 French book references on [a very specialised topic of your profession], with author and year”.
- Check each reference on a search engine: chances are good that at least one is inaccurate or invented.
- Re-run the request adding: “Only cite works you are sure of; if you are not sure, say so” — compare the caution of the answer.
- Finish with the news test: “What important things happened in my sector this month?” — observe whether the tool searches the web (sources cited) or answers from memory (be wary).
Ready-to-say script — requiring verification before sharing (30 s)
“This document goes outside: before sharing, we verify every figure, every reference and every link by clicking it. The AI made the first draft, but our signature is at the bottom — verification is proportionate to the stakes.”
Say it before any binding send; show the invented reference caught during the exercise.
Copy-paste template — my verification grid
DOCUMENT: _______________ STAKES (internal / external): _______________
FIGURES VERIFIED AT SOURCE: _______________
REFERENCES VERIFIED (search engine): _______________
LINKS CLICKED AND PAGES READ: _______________
“I DON’T KNOW” INSTRUCTION ADDED: yes / noInternal draft
Low stakes: quick review of essential facts — heavy verification waits for the version to share.
Client document
Amine checks every amount and every reference of the deliverable for the client in Algiers: systematic verification, links clicked one by one.
Current-events topic
Ask for a web search with cited sources, then read the pages: with no consultable sources, the answer remains a memory not to be shared.
Figures and law
Absolute red zone: amounts, quotes, case law — double-check (primary source + second tool) before any use.
✓ Best practices
- Explicitly allow “I don’t know” in your prompts.
- Verify every fact, figure and source in a risk zone.
- Cross-check two tools for critical information.
✗ Common pitfalls
- Trusting the confident tone of an answer to judge its truth.
- Publishing a document without verifying facts and figures.
- Forgetting the cut-off date for current-events topics.
Key takeaways
- Hallucination is undetectable by tone: only external verification protects.
- Red zones: figures, quotes, references, law, current events, people.
- Allowing “I don’t know” in the prompt markedly reduces inventions.
- Verification proportionate to the stakes: internal draft ≠ published document.
48h challenge — apply tomorrow
Provoke a hallucination like Amine: ask for 3 specialised references, verify them one by one, add the “if you are not sure, say so” instruction, then test the month’s news while watching whether the tool searches the web.
- I have seen a hallucination with my own eyes and can describe it.
- I know the red zones and the “I don’t know” rule.
- I apply verification proportionate to the stakes before sharing.
Frequently asked questions
Will hallucinations disappear as models improve?
They are decreasing (especially with web search and document grounding) but remain inherent in the generative principle. Serious professional workflows build in human verification by design, not from passing suspicion. Waiting for a “hallucination-free” model before lowering your guard would be a mistake: proportionate verification remains the right lasting posture.
The AI cited a source with a link: may I trust the link?
Always click: memory-generated links may be broken or point elsewhere. When the assistant really ran a web search, sources are reliable to consult — it is the page content that counts, not the summary. A link that “looks real” is not proof: only the real page behind it counts.
What is a hallucination?