Skeptical Malaysian manager with crossed arms reviewing a printed report while a colleague points at figures beside her laptop

AI Hallucinations at Work: The Cost of the First Plausible Answer

Short answer: An AI hallucination is when an AI tool states something false with complete confidence: a number, a quote, a source, a rule. The cost rarely comes from the AI making things up. It comes from the first plausible answer going straight into a report, a client email or a decision, with nobody checking it.

Ask an AI tool a question and it answers.

No hesitation, no "let me get back to you", no visible doubt.

That fluency is exactly why hallucinations are expensive at work.

The tool does not know when it is wrong, and there is no built-in alarm that reliably warns you.

What an AI hallucination actually is

An AI assistant generates its answers rather than reading them from a checked source.

Some tools can search or retrieve documents. The sentences they present are still generated, and a cited link is not the same as a verified one.

A prediction can be exactly right.

It can also produce a source that does not exist, a figure that was never measured, or a quote nobody ever said.

The output can look identical either way. Same tone, same structure, same confidence.

Fluent wording is not evidence of accuracy. That is what makes the failure so easy to miss.

Why the confident tone is the trap

People are wired to read confidence as competence.

A nervous colleague gets a second look; a polished paragraph gets forwarded.

AI output arrives pre-polished.

It skips the scrutiny that a rough, hesitant draft would get, and it lands in documents that carry real weight: board decks, client proposals, published reports.

A real case: the airline held responsible for its chatbot (2024)

In February 2024, a Canadian tribunal ordered Air Canada to pay CA$812 in compensation after the chatbot on its website invented a refund policy. CBC News reported the case, and the tribunal decision is public (Moffatt v. Air Canada, 2024 BCCRT 149).

A passenger, Jake Moffatt, asked the chatbot about bereavement fares after a death in the family.

The chatbot said he could book full-price tickets and claim the reduced fare afterwards, within 90 days.

That policy did not exist. The airline's own bereavement page said the opposite.

Air Canada's defence was remarkable: it argued the chatbot was "a separate legal entity that is responsible for its own actions".

The tribunal disagreed. A company is responsible for all the information on its website, whether it comes from a static page or a chatbot.

A system that could produce the false policy was live on the company's website, and the wrong information reached a customer.

A real case: the A$440,000 report with invented sources (2025)

In October 2025, Deloitte Australia agreed to partially refund the Australian government over a report that contained AI-hallucinated citations, as reported by Ars Technica and the Australian Financial Review.

The report, a welfare-system compliance review for the Department of Employment and Workplace Relations, cost taxpayers nearly 440,000 Australian dollars (about 290,000 US dollars).

After publication, a University of Sydney health-law researcher noticed citations to papers that did not exist, including several attributed to a real professor who had never written them.

The corrected version removed 14 of the original 141 references. It also disclosed for the first time that a generative AI tool (Azure OpenAI GPT-4o) had been used in the work.

Deloitte said it would repay the final instalment of the contract.

The failure was not that AI was used. The failure was that AI-invented citations made it into a published government report.

What these two cases share

In both cases, AI-generated text reached people as fact: a customer relying on a company website, readers relying on a published government report.

Whatever checks existed did not catch the false information first.

The tool made things up, and the false information passed as fact.

That is why this is a management topic, not only a technology topic.

Hands circling a figure on a printed business report beside a laptop, with a small Malaysian flag on the desk

Where hallucinations hurt most in everyday work

  • Numbers in reports. A growth figure or market size that sounds right and was never measured. It gets quoted internally, then externally, and correcting it later is embarrassing.
  • Sources and citations. The kind that sank a A$440,000 report. A reference that looks plausible is not a reference that exists.
  • Names, dates and titles. The wrong year on a case study, a job title that is almost right, a person quoted on a topic they never covered. Small errors that destroy trust with anyone who checks.
  • Policies and rules. The Air Canada pattern: terms stated confidently that do not exist. Dangerous anywhere customers or staff rely on the answer.
  • "Facts" about your own company. Ask an AI about your organisation and it may blend you with a competitor or an old news story. Harmless in a test, awkward in a proposal.

The verification habit that catches them

  • Check every number against its source. Open the original figure, not the AI's summary of it. If no source exists, the number does not go in.
  • Treat every citation as unverified until opened. The Deloitte case above is exactly this failure. Opening a reference takes seconds; defending a fake one takes months.
  • Be suspicious of the perfectly phrased quote. If a quote is exactly what the argument needed, confirm it exists before it goes anywhere.
  • Keep AI for drafting, not for final facts. Let it shape the document. Let verified sources fill the load-bearing parts.
  • Decide who is responsible before, not after. Air Canada's defence failed because responsibility cannot be handed to a tool. The same is true inside a company: the name on the document owns its contents.
    Infographic showing an AI answer flowing into a source check, then splitting into verified or discarded

Three crisp statistics, one of them invented

Picture a manager in Penang preparing a board deck.

She asks an AI assistant for three statistics about hybrid work, and gets three crisp figures with citations.

Two are real. One is invented, and it happens to be the one that supports the recommendation.

Nobody checks, because all three look the same.

Six months later the invented figure has been quoted in two budgets and a hiring plan.

This example is a hypothetical illustration. In both real cases above, the false information surfaced only after it reached its audience.

A reviewer with a checking habit opens each cited source before the deck ships, and the invented figure has nowhere to hide.

In a good course, learners run exactly this exercise: same prompt, same three figures, and the trainer asks them to find the one that does not exist.

What good use looks like

The teams getting real value from AI are not the ones that trust it most.

They are the ones whose checking habit matches the stakes: a quick scan for a low-stakes draft, source-by-source verification for anything that leaves the building.

AI is very fast at producing a first version of the truth.

Deciding what is actually true is still the job.

Frequently asked questions

Is AI hallucination a sign the technology is not ready for work?

No. It is a sign the technology needs the same skepticism as any other source.

A confident new colleague also gets checked until their work is proven. AI simply never graduates past that stage.

Which AI tools hallucinate?

Any of them can produce confident errors, and reported rates change with every version. The safe habit is the same for all of them: verify important output before you rely on it.

Is a hallucination the same as the AI lying?

No. A hallucination is a false or unsupported output. That is different from deliberate deception by a person.

The practical conclusion is the same either way: responsibility sits with the person relying on the output.

The companies that get burned are not the ones using AI. They are the ones publishing its first answer.

A verification habit is a trainable skill, and it spreads quickly once a team has seen one real near-miss.

Michael Cheong runs hands-on AI training for Malaysian workplace teams, built around exactly these habits: where to trust the tool, where to check it, and how to make that judgment second nature. Details are on the AI Trainer for Workplace Productivity page.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *