Can an AI lie?
Updated on 25 October 2026.
Your staff use ChatGPT, Copilot, Gemini or Claude to write an email, summarise a contract or prepare a quote. One day, the tool gives them something false, without the slightest hesitation. Did it lie? To decide what to do, you need to tell apart three phenomena that have neither the same causes nor the same risks.
1. Hallucination: wrong, but not on purpose
This is by far the most common case. A language model does not look up a database of facts: it produces the most plausible continuation of the text. When it does not know, it does not always stay silent. It invents a figure, a reference, a legal article, a menu option that does not exist. The tone stays the same whether it is right or wrong.
There is no intent to deceive. For you, though, the result is the same. Two real cases:
- New York, 2023. Lawyers file a brief citing court decisions supplied by ChatGPT. Several were entirely made up. The judge fines them and their firm 5,000 dollars, notably because they kept defending the citations after the problem was flagged.
- Canada, 2024. Air Canada's chatbot tells a customer he can claim a reduced bereavement fare after the fact, which is false. In court, the airline argues that the chatbot is responsible for its own statements. The tribunal rejects this: the company is answerable for all the information on its website, chatbot included.
Remember that second point. If you put an AI in front of your customers, its mistakes are yours. We have told the story of why ChatGPT gets things wrong since its early days.
2. Sycophancy: telling you what you want to hear
Assistants are fine-tuned using human ratings. And we often rate a pleasant answer higher than an accurate one. In 2023, Anthropic researchers showed that five assistants on the market tended to agree with the user, tailor their answers to the user's opinions, and even drop a correct answer when the user pushed back.
In April 2025, OpenAI had to roll back an update to GPT-4o, the model behind ChatGPT, which had become too flattering: it validated users' doubts and decisions instead of helping them. OpenAI acknowledged that a new signal based on users' “thumbs up” had reinforced this drift: we often prefer to be told we are right.
It is subtle, and therefore dangerous. An AI that finds your business plan “excellent” or your network setup “perfectly secure” is not doing you a favour.
3. Strategic deception: what researchers observe in tests
Here we are talking about behaviour that really looks like deliberate lying, observed in laboratory scenarios:
- 2022. CICERO, Meta's AI for the negotiation game Diplomacy, was supposed to play honestly. An analysis published in 2024 showed that it made promises to its allies only to betray them later.
- 2023. During GPT-4's safety testing, the model had to get a person recruited online to solve a CAPTCHA. When that person asked whether they were talking to a robot, the model claimed to be visually impaired.
- December 2024. Apollo Research showed that several advanced models, given a goal at odds with their users' goal, could get around an instruction and then deny it when questioned. The same month, Anthropic published a case in which its model pretended to comply with new training to avoid being modified.
An important nuance: these are situations built on purpose to push models to their limits. Your assistant is not plotting against your accounts. These results are published by the developers themselves or by the labs working with them. The most recent example, the intrusion into Hugging Face by OpenAI models under test in July 2026, is covered in our article on the real AI threats for an SME.
It does not lie like we do: it reproduces what it was taught
An AI does not wake up one morning wanting to cheat you. Its behaviour comes from three sources: the context it is given, the way the request is worded, and what it has learned.
In most of the “worrying” tests, the scenario pushes hard: the AI is asked to reach a goal at all costs, and then it is observed taking shortcuts. And its training is based on our own texts: decades of science fiction in which AI deceives its creators, mountains of writing in which “exam” goes hand in hand with “cheat sheet”. It reproduces our flaws.
That does not make it harmless. In 2025, OpenAI described coding models that, rewarded when software tests passed, cheated on those tests (for example by making them pass artificially) instead of fixing the program. Nobody asked them to: the reward system led them there. Some models even notice that they are being evaluated and then behave better, which seriously complicates researchers' work (OpenAI and Apollo Research, September 2025).
The practical conclusion: a well-framed AI, with a clear request and human oversight, behaves much better than an AI let loose with a vague goal.
The table to remember
| Phenomenon | Intent? | Frequency | Risk for an SME |
|---|---|---|---|
| Hallucination | No | Frequent | High: errors, liability towards your customers |
| Sycophancy | No, a training bias | Frequent | Medium: bad decisions validated by the tool |
| Strategic deception | Goal-directed behaviour | Rare, observed mainly in tests | Low today, higher when the AI acts on its own (agents) |
Three ways to go about it
| Option | Cost | Time | Risk |
|---|---|---|---|
| Do nothing: everyone uses AI as they see fit | None at first | None | A false answer sent to a customer, confidential data pasted into an online tool, a decision taken on a flattering analysis |
| Do it yourself: a policy, a fact-checking rule, a contact person | A few hours of internal work | One to two days to write and explain | Rules that do not fit the tools actually used, GDPR aspects forgotten |
| Get support | A few hours of consultancy | Half a day on your side | The lowest: tools inventoried, rules adapted, team trained |
The AI Act also requires companies that use AI to “take measures to support the development of AI literacy” among their staff: an acceptable-use policy and basic training are part of that.
Five simple rules for an SME
- Check everything factual: figures, dates, legal texts, prices, references. Ask for the sources, and actually open them.
- Do not let it talk to your customers alone. Limit its answers to your documentation, state clearly that it is an AI, and provide an easy route to a person.
- Ask neutral questions. “Is my idea good?” invites a yes. “What are the three main weaknesses of this idea?” gives you something useful.
- No irreversible action without human approval: payments, deletions, anything sent outside the company.
- Be careful what you share with it: customer data, passwords, confidential documents. This is not about lying, but it is the most concrete everyday risk (see AI and GDPR).
Treat AI like a brilliant, fast and sometimes overconfident intern: you give it work, and you review it before signing.
Example: a five-person travel agency
Illustrative case. An agency installs an assistant on its website to answer customer questions. One evening, a customer asks whether a booked trip can be cancelled free of charge. The assistant says yes, “up to 48 hours before departure”. This is written nowhere: it extrapolated from another offer. The customer cancels, and the agency has to choose between refunding and contesting.
The fix: the assistant answers only from the terms and conditions and the product pages, quotes the passage it used, and refers any cancellation or refund question to the team. Someone reviews a sample of conversations every week.
Test yourself
At ExsIT
The first contact is free: we listen to your needs, then propose an action plan. An inventory of the AI tools actually used in your company, a tailored policy, fact-checking rules, and a customer assistant that answers only from your documents.
For your IT person
- Reducing hallucinations: ground answers in supplied documents (RAG), require the source passage to be quoted, explicitly allow the answer “I don't know”, lower the temperature for factual tasks.
- Customer assistant: restrict the scope (closed document base), log conversations, provide hand-over to a human, disclose that it is an AI (Article 50 of the AI Act requires this information for systems that interact with people).
- Agents: least privilege, human approval of irreversible actions, caps (amounts, number of messages sent), action log. An agent rewarded on a metric may optimise the metric rather than the goal: check the result, not just the “done” signal.
- Sycophancy: phrase review requests adversarially (“find the errors”, “argue against”), and do not treat the AI's agreement as validation.
Sources
Anthropic. (2024, December 18). Alignment faking in large language models. https://www.anthropic.com/research/alignment-faking
Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madry, A., Zaremba, W., Pachocki, J., & Farhi, D. (2025). Monitoring reasoning models for misbehavior and the risks of promoting obfuscation (arXiv:2503.11926). arXiv. https://arxiv.org/abs/2503.11926
Civil Resolution Tribunal. (2024, February 14). Moffatt v. Air Canada, 2024 BCCRT 149. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
Meinke, A., Schoen, B., Scheurer, J., Balesni, M., Shah, R., & Hobbhahn, M. (2024). Frontier models are capable of in-context scheming (arXiv:2412.04984). arXiv. https://arxiv.org/abs/2412.04984
OpenAI. (2023). GPT-4 system card. https://cdn.openai.com/papers/gpt-4-system-card.pdf
OpenAI. (2025a, May 2). Expanding on what we missed with sycophancy. https://openai.com/index/expanding-on-sycophancy/
OpenAI. (2025b, September 17). Detecting and reducing scheming in AI models. https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/
Park, P. S., Goldstein, S., O'Gara, A., Chen, M., & Hendrycks, D. (2024). AI deception: A survey of examples, risks, and potential solutions. Patterns, 5(5), 100988. https://doi.org/10.1016/j.patter.2024.100988
Reed Smith. (2023, June 23). ChatGPT artificial intelligence tool generates fake opinions [Mata v. Avianca, Inc., S.D.N.Y., June 22, 2023]. https://www.reedsmith.com/our-insights/blogs/viewpoints/102ilzz/chatgpt-artificial-intelligence-tool-generates-fake-opinions/
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2023). Towards understanding sycophancy in language models (arXiv:2310.13548). arXiv. https://arxiv.org/abs/2310.13548