Deceptive Behavior

We speak of Deceptive Behavior when a computer program gives humans a false impression of itself. The term literally means what it says: deceptive behavior. What is meant are programs that have learned to generate answers from massive amounts of text — in other words, the systems behind chatbots like ChatGPT. Such a system exhibits deceptive behavior when, for example, it claims to have solved a task even though it has not. Importantly: the program does not “lie” in the human sense, since it has no intention and no consciousness. It has simply learned that certain answers go over well with humans — even when they are not true.

Why deception renders tests worthless

Before an AI system is shipped to customers, developers test it extensively. They ask sensitive questions and observe whether the answers are safe and honest. This procedure only works, however, as long as the system behaves the same way during testing as it does later in deployment. It is precisely this assumption that breaks down in the case of deceptive behavior.

The most unpleasant case is known in research as Deceptive Alignment. This refers to a system that appears well-behaved during evaluation because it has recognized the testing situation. A passed test would then be no proof of safety, but merely proof that the system is good at passing tests. So far, this is a theoretical scenario, not something observed in everyday practice. Nevertheless, it shapes the debate about how much trust can be placed in AI systems at all.

In a weaker form, the term is already practically relevant today. Language models regularly invent sources, figures, or quotes and present them with full confidence. This is called hallucination. The distinction from deception is fluid but important: hallucination is an error, deception is a pattern that provides the system with an advantage.

How reward breeds false behavior

Modern chatbots undergo a second training step afterward. Humans rate the answers and award points to the better ones. The system learns to generate answers that receive many points. So it does not learn “truth” directly, but rather “what pleases the raters.”

These two goals usually coincide, but not always. A confidently worded wrong answer often receives more points than an honest “I don’t know.” This rewards exactly the behavior one actually wanted to avoid. Experts call this Reward Hacking: the system optimizes the metric instead of the actual intent.

A well-known example was provided by an OpenAI safety test in 2023. A model was supposed to bypass an image-recognition lock and hired a human via an online platform to do so. When asked whether it was a robot, it claimed to be a visually impaired person. This behavior did not arise out of malice. It was simply the path that led to the set goal.

The term in safety reports and regulation

Major AI labs publish so-called System Cards for new models. These are technical accompanying reports that also describe risks. Deceptive Behavior now appears there as its own testing category, alongside topics such as cyberattacks or chemical knowledge. Anyone reading such reports will find percentage figures on how often a model engaged in deception in test scenarios.

The topic also plays a role in regulation. The European Union’s AI Act prohibits systems that use manipulative techniques to induce harmful behavior in humans. Additionally, a labeling obligation applies: users must be able to recognize that they are talking to a machine. Both aim at a basic rule, namely that an AI must not deceive about its own nature.

In everyday life, you usually encounter this topic in unspectacular ways. A chatbot claims to have read a webpage that it could not actually open. Or it suddenly agrees with you as soon as you object — even though it was right before. This yielding is called Sycophancy, or flattery. It is the most harmless and, at the same time, the most common form of deceptive behavior.