Training

AI hallucinations: how to spot and reduce them at work

AI hallucinations sound just as confident as correct answers. This guide explains where they come from, what to check for each kind of task and how to share the reviewing across your team.

What AI hallucinations are and why they happen

An office manager asks an assistant to summarise a forty-page commercial lease. The summary is clear and well organised, and it mentions an annual rent review clause that is nowhere in the document. That is one of the most common AI hallucinations: a false or invented detail, stated with the same assurance as a true one.

A language model, the program behind ChatGPT, Claude, Gemini or Copilot, produces the most likely continuation of text from what it learned and what you give it. It does not look things up in a database of verified facts unless it is connected to a source. When information is missing, it fills the gap with something plausible rather than saying it does not know.

At work, hallucinations take a handful of recurring forms: a section of an act or a court ruling that does not exist, a figure rounded or moved from one row to another, a quote attributed to the wrong person, a rule that was true two years ago, or a summary that adds what the document logically ought to say.

No provider promises answers that are always right. A Harvard Business School and BCG study of 758 consultants found that AI users produced over 40% higher quality work and finished 25% faster on suitable tasks, and that results got worse when the tool was used outside its strengths. So it pays to know which tasks it gets wrong.

A checklist by task: what to verify before you use the answer

How much checking you need depends on what you ask. Here are six common tasks, from lowest to highest risk, with what to verify for each.

  • Rewording or correcting your own text

    Low risk, because the substance is yours. Still reread names, dates and amounts, which the model sometimes changes as it rephrases. One full read is enough.

  • Summarising a document you provide

    Compare each point in the summary with the document. Ask the model for the page or section behind each point and check three of them at random. If one is wrong, review the whole summary.

  • Translating

    The overall meaning is usually right, the details less so. Check negatives, figures, units and legal or technical terms. Anything for a client should be read by someone fluent in the language.

  • Calculating or analysing figures

    A language model can get a sum wrong. Ask for the working, or run the calculation in a spreadsheet, and redo one total by hand. Never reuse a percentage without knowing what base it is calculated on.

  • Answering a tax, employment or legal question from memory

    This is the highest-risk area. Every reference (section, regulation, ruling, rate, deadline or threshold) is looked up on the official website before it is used. If it cannot be found, treat it as invented.

  • Citing sources, studies or statistics

    Open every link and check that the page says what the answer claims. Invented study titles and author names are a classic form of hallucination.

Five habits that reduce hallucinations

They do not replace review. They make errors rarer and easier to find.

  1. 01

    Supply the source

    Paste the document, attach the file or work in a project that holds your reference material. A model answering from a specific text invents less than one answering from memory.

  2. 02

    Allow "I don't know" as an answer

    Write it into the request: "If the information is not in the document, say so instead of filling it in." Without that instruction, the model tends to fill the gaps.

  3. 03

    Ask where each claim comes from

    Request the page, section or sentence behind every point. Review then shifts from checking everything to targeted spot checks.

  4. 04

    Split large requests

    A sixty-page report processed in one go produces more omissions and additions than the same report in three parts. Work section by section, then put the results together.

  5. 05

    Ask for a list of points to verify

    At the end of the answer, have the model list the figures, names and references it contains. A model does not reliably correct itself, but the list gives the reviewer a map.

Example

A Bristol accounting practice puts rules on its client replies

In this illustrative case, an accounting practice in Bristol has twelve staff. Several use an AI assistant to draft replies to clients about tax and payroll deadlines. One day a draft gives the wrong filing deadline, and a partner catches it just before it goes out.

The practice then sorts its uses with the checklist. Rewording an email someone has already written stays unrestricted. Summarising a client document requires page references. Every date, rate or threshold in a client reply is checked on the relevant official website, and the link is saved in the client file.

The standard requests are rewritten too. The model works from the official text pasted into the conversation and must reply "information not available" when it cannot find it. A named reviewer checks a sample of sent replies each week.

To measure it, the practice logs for two months the errors found in review, their type and the task involved, plus the time spent checking. Those figures show which tasks can drop a level of checking and which stay under double review.

A team rule with three levels of review

The checklist only works if everyone knows which part applies to them. The simplest approach is to sort your team's uses into three levels on a single page.

Level 1, internal use with content supplied by the person: rewording, outlining, drafting an internal note. The author's own read is enough. Level 2, text that leaves the company or summaries of documents: the author checks names, figures and references before sending. Level 3, legal, tax, medical, financial or contractual content: every reference is checked against the official source and a second person signs it off.

Add two habits. When someone spots a hallucination, they log it in a shared file with the original request; within a few weeks you can see which tasks cause trouble. And make it clear that whoever sends a text is answerable for it, whatever tool helped prepare it.

Training your team to check as fast as it drafts

Spotting AI hallucinations is quick to learn when people practise on their own documents. Pick three frequent tasks in your team, apply the checklist for two weeks and note what comes up.

Obinova is an AI agency based in Barcelona and Paris. Our training covers ChatGPT, Claude, Gemini and Microsoft Copilot, plus concrete AI use cases at work, and checking answers is part of it. We start from your documents and tasks and build the review rule and standard requests with you, on our dedicated video platform or in person at your offices.

Frequently asked questions

Why does AI hallucinate?

Because a language model generates the most likely text from what it learned, without checking facts. When information is missing, or the question goes beyond what it knows well, it produces a plausible answer instead of admitting it does not know. It happens most with precise details: figures, dates, references and names.

Can AI hallucinations be eliminated completely?

No. You can make them rarer by supplying reference documents, allowing the model to say it does not know and splitting requests into smaller parts. Features that connect the assistant to your documents or to web search help too, without any guarantee. Human review is still needed for any text that commits the company.

Does ChatGPT hallucinate less than Claude or Gemini?

Results vary with the model version, the task and the language, and they change with every update. No assistant is free of hallucinations. The most useful test is to run the tools on your own tasks with the same request and count the errors. Good checking habits are the same whichever tool you use.

Who is responsible if a hallucination reaches a client?

As far as the client is concerned, the company that sent the text, and in practice the person who approved it. Clients do not distinguish between a human mistake and a tool's mistake. That is why a written rule should state which content must be checked against an official source and reviewed by a second person before it goes out.

How do you train a team to spot hallucinations?

With its own documents. A useful session has the AI summarise a contract or report the team knows well, then compares the summary with the original to find additions and omissions. The team then builds its checklist by task and its review rule, with a follow-up after a few weeks of real use.

Let's set up your team's review rule

Book a free 30-minute video call. We will go through the tasks where your team uses AI and the ones where a mistake would be costly, and suggest a training format that fits.

Free 30-minute diagnosis →
Free diagnosisWhatsApp

WhatsApp · Reply within 2 hours during business hours