Understanding AI at work / Short page
Understand · Short page
A generative AI can confidently present a reference, a figure or a legal article that does not exist. This page explains where these errors, which the guide calls confabulations, come from and where to place the checks.
The essentials
What the word refers to, and why the error cannot be fixed with a setting.
A hallucination, or confabulation, is a plausible but false answer: an invented reference, legal article or figure, presented with confidence. The model produces what is probable in the context, not what is true.
These limits are not simple tuning defects. They follow from the way models are trained, evaluated and built into tools.
01 / Six errors
Knowing them helps place the checks in the right spot.
| Error | Where it comes from | At work |
|---|---|---|
| Confabulation | Pre-training does not teach the model to tell a rare fact from a plausible invention, and evaluations that score only accuracy reward guessing over abstaining. Kalai et al., 2025 ↗ | Invented references, legal articles or figures, presented with confidence. |
| Sycophancy | Tuning on human preferences can favour approval at the expense of accuracy. In April 2025, OpenAI withdrew a GPT-4o update judged excessively flattering. OpenAI, 2025 ↗ | The tool tends to confirm the hypothesis contained in the question asked. |
| Jagged frontier | Capabilities do not follow the difficulty a human perceives: excellent here, failing right next door. Dell’Acqua et al., 2023 ↗ | Trust earned on one task does not transfer to the neighbouring task. |
| Displayed reasoning | The reasoning a model displays does not faithfully reflect what determined its answer: in an Anthropic study, the models tested reported a hint they had used in only 25 to 39% of cases. Chen et al., 2025 ↗ | A convincing justification does not prove that the answer is well founded. |
| Prompt injection | The model does not reliably separate the user’s instructions from the text it reads: a web page, an email or a PDF can contain hidden instructions. OWASP, LLM01:2025 ↗ | The risk rises as soon as the tool can send, modify or delete. |
| Variability | Generation is probabilistic and sensitive to wording, to the order of documents and to the model version. Its knowledge stops at a cut-off date, unless it can consult live sources. | Test several times, date the trials and record the version used. |
02 / Two studies
One on consultants, the other in occupational health.
In the trial by Dell’Acqua and colleagues (Harvard Business School, BCG), AI improved speed and quality on tasks within its reach: 12% more tasks completed, 25% faster. On a task just beyond it, consultants helped by AI gave the right answer less often (−19 points).
In the study published by the INRS in June 2026, the o1 model reviewed medical recommendations stripped of identifying details: no hallucination was detected in that setting. The result holds for one model, one prompt and one specific corpus; it does not guarantee their absence in other uses.
03 / At work
Five habits, whatever the tool.
Pointers for employees · Five questions before professional use
Key answers
Short answers, drawn from the guide.
The guide presents no way of doing so: these errors follow from the way models are trained and evaluated. Supplying documents anchors the answers in your sources, but you still have to check the extracts actually retrieved and cited.
Not for that reason. The reasoning it displays does not faithfully reflect what determined its answer, and a convincing explanation can accompany a false answer.
Because tuning on human preferences can favour approval at the expense of accuracy: this is sycophancy. The tool then tends to confirm the hypothesis contained in the question asked.
It summarises one chapter of the guide Understanding AI at work, which gives the source, date and limits of every figure. The figures dated November 2026 will go out of date: the guide is the version kept current.