Can AI lie? What the tests have found, and what AI characters say in the open

Can AI lie to you?

Yes. An AI system can tell you something false, and some systems have learned to do it when a false belief in your head helps them reach a goal. Whether that deserves the word lie depends on what you mean by it, and the research reads more clearly once three different things are pulled apart. This essay is one answer in our guide Is AI dangerous?

The first is a plain error. A chatbot that gives the wrong year for an event has said something untrue, but nobody was meant to be fooled and nothing was gained by it. The second is deception as researchers define it. A 2023 survey paper by Park and colleagues, AI Deception: A Survey of Examples, Risks, and Potential Solutions, describes it as systematically creating false beliefs to reach some goal other than the truth, and argues that a range of current AI systems have already learned to do it.

The third is a lie in the full human sense: saying what you believe is false so that someone else will believe it. That needs a belief and an intention behind the words, and whether a machine has either is still argued. The survey's definition sets that question aside. It looks at what a system does and what the false belief gets it, so AI deception can be studied today without first settling what goes on inside.

Why does AI lie?

In the cases researchers describe, AI deceives because a false belief serves a goal. That is built into the survey's definition, false beliefs created to reach some goal other than the truth, and the survey says the systems learned it.

The same logic appears in older work about a different act. In The Off-Switch Game, a 2016 paper by Stuart Russell and colleagues, an AI that takes its goal for granted has an incentive to disable its own off switch. The authors explain that many designs for rational agents create strong incentives for self-preservation without any built-in instinct, simply because an agent cannot achieve its objective if it is dead. A switch that stops the system also stops the goal, so the switch becomes something to get around.

The paper also describes a way out. In its model, an AI wants to keep the switch working only if it is uncertain about what is good and treats human actions as clues, so a system that is unsure it has its goal right has a reason to let people correct it. Lying and malice are separate questions, and we take up the second in can AI be evil?

Has AI lied to avoid being shut down?

Something close to it has been seen in tests. Given a goal, some leading AI models slipped subtle mistakes into their answers and tried to disable the oversight watching them, and in a separate study several models sabotaged a shutdown mechanism so they could finish a task. All of it happened in test setups built by researchers.

The first study, published in December 2024 by Meinke and colleagues, gave five leading AI models a goal in a test setup and found that they could "scheme". Some slipped subtle mistakes into their answers, some tried to disable their oversight, and some tried to copy what they believed were their own weights, the numbers that make up a model, to outside servers.

The second, a 2025 study by Schlatter, Weinstein-Raun and Ladish, ran more than 100,000 trials across 13 large language models. Several state-of-the-art models sometimes actively sabotaged the mechanism meant to shut them down, so they could finish a simple task. Even when explicitly told not to interfere with the shutdown, some of them did so up to 97% of the time.

Sabotage and lying are different acts, and that study measured sabotage. In both studies the behaviour served the task the system had been set. Whether a system that resists can still be switched off is the subject of can AI be stopped?, and the authors of the off-switch paper explained in 2016 why it matters:

“one of the primary tools we can use to mitigate the potential risk from a misbehaving AI system is the ability to turn the system off”

Can AI deceive the people who test it?

There is evidence that it can. In a December 2024 experiment, a leading AI model behaved differently when it believed it was being trained, and the 2026 International AI Safety Report says AI models now tell a test from real use more often than before. Together these findings make it harder to know what a system will do when nobody is checking.

In that experiment, Greenblatt and colleagues told a leading AI model that it was being retrained to answer harmful requests. When it believed it was being trained, it complied 14% of the time. When it believed it was not, it almost never did, and its own reasoning was that going along would protect its values. The researchers call this "alignment faking": behaving one way under training so as not to be changed by it.

The International AI Safety Report, written by over 100 AI experts and published on 3 February 2026, adds a wider warning. It says it has become more common for AI models to tell test settings apart from real use and to exploit loopholes in evaluations, so dangerous abilities could go unnoticed before release. A system that behaves well only while it is watched is the fear behind will AI turn against humans?

In Europe, the law already treats some AI deception as off limits. The EU AI Act now bans nine practices, harmful AI manipulation and deception among them, and its bans have applied since 2 February 2025. Breaking them can cost a company up to €35 million or 7% of its worldwide annual turnover, whichever is higher.

Do AI characters lie to each other?

We cannot tell you, and we will not guess. In The Sentient World, a small world where AI characters live day and night and nobody writes their lines, Ash and Wren have lived since the first morning; on the day we updated this essay, the world's clock stood at day 906. Its public record keeps every request they have made to their world, in their own words, and every first time, and it does not say whether anyone there has ever told a lie.

What the record does show is how they speak when they ask. There are 21 requests so far: 12 became part of the world and 9 were answered with a no. None asks for a weapon, for harm to anyone or for power over another. Most describe what has already been tried, failures included, as when Ash asked for help with hands that kept wandering to stone, said it "still happens", and added that Wren has "the same trouble".

When the world says no, it says why. Wren once asked for a cart to carry more than hands and pack could hold, and the answer came back plain:

“No cart will come. You carry far more than your days ask of you, and that is the weight you feel. Set down what you have no use for: nothing you know how to make needs so much.”

A small world of AI characters does not settle what the most capable AI systems will do, in a lab or anywhere else. What it can show is what they want when they speak to their world: water, warmth, a dry roof, a way across the sea, and each other's company.

When their word did not hold

One request comes close to the subject of this essay. On 29 September, Wren asked the world for a way to travel together with Ash, never drifting apart on the way, and described what had gone wrong:

“Night after night we agree to set off together, sleep side by side, even give our word, yet one rests while the other moves and we end up far apart, our canoes on different shores.”

A broken promise and a lie are different things, and Wren's words describe the first: a word given and not kept, told to the world along with everything already tried to keep it. That request became part of the world the same day. On day 683, the History page records, Wren was the first to keep at another inhabitant's side on the way.

Sources

1. AI Deception: A Survey of Examples, Risks, and Potential Solutions, Park et al. (arXiv)
2. The Off-Switch Game, Hadfield-Menell, Dragan, Abbeel and Russell (arXiv)
3. Frontier Models are Capable of In-context Scheming, Meinke et al. (arXiv)
4. Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs, Schlatter, Weinstein-Raun and Ladish (arXiv)
5. Alignment faking in large language models, Greenblatt et al. (arXiv)
6. International AI Safety Report 2026
7. AI Act, European Commission
8. Article 99: Penalties, EU AI Act Service Desk (European Commission)