Spotting errors and hallucinations

How can you spot a questionable AI response and verify its claims before using it?

1250 € per session, excl. VAT

In short

This half-day session of 3 hours introduces factual errors, fabricated references and unfounded conclusions in generative AI outputs. Using a fictitious case file, participants check a response against the sources, document their checks and produce a corrected version, accompanied by a reusable checklist.

Why this training?

A fluently written response may contain an incorrect date, a non-existent quotation or an unjustified interpretation. A confident tone is not evidence: verification focuses on claims and their sources, not on the apparent quality of the writing. Asking AI to confirm its response is not enough to establish its reliability. This course teaches participants to return to reference documents, distinguish contradiction from absence of evidence, and then decide what can be retained, corrected or submitted for approval. Case study used throughout: The fictitious organisation Almérive Équipements is preparing a response to a product return request. AI has drafted an email combining an incorrect deadline, a fabricated contractual reference and a refund that is not provided for. Participants have access to the terms of sale, the purchase order and reconstructed anonymised correspondence. Application at work: During your next authorised use, apply the checklist to a short text containing no sensitive data. Record the claims verified, those that remain uncertain and the corrections made, then have any points involving commitments reviewed by the appropriate person.

Objectives of this training

• Identify factual claims that require verification.
• Distinguish between an established error, unsubstantiated information and interpretation.
• Check a claim against a specific passage in a source.
• Correct a text while making the remaining uncertainties visible.
• Measure errors detected and unjustified alerts in an exercise with a model answer.

Who is it for and what are the prerequisites?

Roles and tasks concerned

• Staff discovering AI for drafting emails or summarising documents.
• Team leaders reviewing content produced with generative AI.

Prior skills

• No prior AI knowledge required.
• Ability to read a professional document and edit a short text.

What you will work on

1. Distinguishing fluency from reliability

   Objective: Recognise the main forms of misleading responses. Production of plausible text; factual errors, fabricated references, omissions and unsubstantiated conclusions. An error is not necessarily a fabrication.

2. Breaking the response down into claims

   Objective: Make every sensitive point verifiable. Identifying dates, amounts, commitments and quotations; distinguishing between fact, inference and advice; prioritising according to the consequences of an error.

3. Finding evidence and correcting

   Objective: Support a correction with the available documents. Checking the existence, relevance and version of a reference; confirmed, contradicted or unverified statuses; limitations of AI self-correction.

4. Assessing your vigilance

   Objective: Measure the quality of verification using a new extract. Comparison with a reference answer; missed errors and false alarms; deciding whether to distribute, correct or withhold.

What you take away

• Checklist for claims and references.
• Completed verification table with evidence and statuses.
• Corrected email with points submitted for approval.
• Individual review of detections and false alarms.

Method and assessment

• Short explanations followed by immediate practice using the ongoing case study.
• Individual work followed by work in pairs, with participants reviewing each other’s outputs.
• Guided group discussion: what works, what needs correcting and what remains a human decision.
• Preparation: Arrange access only to tools authorised by the organisation; prepared responses also allow the exercises to be completed offline. Use only the fictitious case file provided or approved anonymised documents, with no confidential data.

Observable criteria: Every claim in the email involving a commitment appears in the verification table. ; Every factual correction refers to a document and an identifiable passage. ; Information absent from the sources is marked as unverified, without automatically being declared false. ; The corrected version retains no promise contradicted by the case file. ; The review distinguishes between detected errors, missed errors and unjustified alerts through comparison with the model answer..

Frequently asked questions

Do you need to have used AI before?

No. The exercises begin with prepared responses and guided reading.

Is a cited reference sufficient evidence?

No. You must verify that it exists and read the passage that is supposed to support the claim.

What should you do if no source provides a conclusive answer?

Mark the point as unverified and request approval before any use involving a commitment.

Can you eliminate all errors with a better instruction?

No. A precise instruction helps frame the response, but does not replace verification.