Responsible Use of Generative AI at Work · Unit 3 of 5 · about 10 minutes
Unit 3: Accuracy, hallucination and the duty to verify
When AI output leaves your hands, the errors in it become your errors. The tool cannot be accountable; you are. This unit gives you a routine that makes verification fast enough that you will actually do it.
What hallucination looks like in practice
"Hallucination" is the industry term for confident fabrication. It does not look like nonsense. It looks like: a statistic with a plausible number and no real source; a citation to a report, article or court case that does not exist; a real person attached to a job title they never held; a legal requirement stated one clause wrong; a product feature your own company does not offer.
Fabrications are most likely exactly where they are most damaging: specific facts, numbers, names, dates, quotations and references. Generic advice is usually safe; checkable specifics are where the risk concentrates.
The check-before-use routine
Before any AI-assisted output leaves your hands, run four questions. With practice this takes under a minute.
- Stakes: who sees this, and what happens if it is wrong? A brainstorm for your own use needs no checking. A client deliverable, a regulatory submission, a public post, or anything with your organisation's name on it gets the full routine.
- Specifics: highlight every checkable claim: numbers, names, dates, quotes, references. Verify each against a source the AI did not write.
- Source: did the facts come from material I supplied, or from the model's memory? Supplied facts need checking against the original once. Memory facts need independent confirmation every time.
- Sign-off: would I stake my name on this if I had written every word myself? Because once it is sent, that is precisely the position.
The asymmetry that catches professionals
Verifying takes minutes. Recovering from a published fabrication takes far longer: corrections, apologies, and a durable dent in how much your work is trusted afterwards. Professionals in several countries have been sanctioned for filing AI-invented citations in court, and companies have had to correct public claims nobody checked. In every case the check would have taken minutes.
Scenario: the conference slide
A marketing lead asks an AI tool for "three statistics on SME AI adoption in Southeast Asia" for a conference deck. The tool supplies three, each with a named research firm. Two are real but outdated. One study does not exist. A journalist in the audience searches for it during the talk and posts that the company is citing invented research. The correction travels further than the talk did.
The fix costs ninety seconds: search each statistic, keep the two that check out with their real dates, drop the third.
Key takeaways
- Fabrication concentrates in specifics: numbers, names, citations, quotes.
- The routine is stakes, specifics, source, sign-off.
- Facts from the model's memory need independent verification every time.
- You are accountable for AI output the moment you send it onward.
Knowledge check
Q1. An AI tool gives you a paragraph for a client report including the line "According to a 2025 Deloitte study, 61% of regional SMEs now use AI weekly". What must happen before this reaches the client?
A named source in AI output is a claim to verify, not a verification. If the study cannot be found, the claim cannot be used.
Q2. Which output needs the LEAST verification effort?
The stakes question comes first. Private working notes carry minimal risk; the other three are outbound, specific and attributed, which is exactly where fabrication does damage.