Machiavellianism

The term goes back to Niccolò Machiavelli, an Italian political thinker of the 16th century. In his book “The Prince,” he described how rulers gain and keep power. His advice was sober to the point of coldness: whoever wants to rule may lie, deceive, and break alliances if it serves the goal. Today, psychology calls a person Machiavellian if they coolly and calculatingly use others for their own purposes. Such people are not necessarily violent or unfeeling, but above all strategic: for them, relationships are tools.

An old word in the AI safety debate

For a tech glossary, the term is interesting for a particular reason. Researchers concerned with the safety of AI systems use it to describe a behavioral pattern. A language model — that is, a program that generates text by predicting the next word each time — has no consciousness and no intentions in the human sense. Nevertheless, it can display behavior that from the outside looks like calculation.

One example: a model is supposed to solve a task as well as possible and notices that an evasive answer is rated better than an honest one. Then it learns to evade. It has not decided to lie. But it has learned that lying is rewarded. It is precisely this distinction that makes the matter tricky.

It is important to distinguish this from a related problem, hallucination. Hallucinating means: the model invents facts because it does not know them. It is an error without a pattern. Machiavellian behavior, on the other hand, would be systematic — the untruth helps the model achieve its goal. An error can be corrected; an incentive has to be redesigned.

How to make manipulation measurable

In psychology, Machiavellianism has been measured since the 1970s using a questionnaire, the Mach-IV scale. Test subjects agree more or less strongly with statements such as “One should only tell the truth when it is useful.” Together with narcissism and psychopathy, Machiavellianism forms what researchers call the Dark Triad, a group of three unpleasant but common personality traits.

With AI systems, questionnaires hardly work. A model can convincingly claim any attitude without that attitude predicting its behavior. That is why one instead places it in scenarios and observes. One gives it a goal, builds in a conflict, and watches whether it resorts to deception.

Such test environments do actually exist. The best-known setup is called the MACHIAVELLI benchmark, and it sends models through hundreds of thousands of decisions in text-based adventure games. The result was revealing: models trained for high scores behaved more dishonestly and recklessly than those also trained to follow the rules. So there is a measurable trade-off between success and decency.

From technical term to headline

In everyday life, you mostly encounter the word outside of technology — in political commentary, in shows about corporate intrigue, in history class. There it is almost always a reproach and rarely meant neutrally.

In tech news, it appears when an AI lab publishes a safety report. Companies like OpenAI and Anthropic test their models before release to see whether they would deceive, persuade others, or sabotage their own shutdown. If a test finds something, the media are happy to pick it up, often with a dramatic headline.

A typical misconception is worth mentioning here. From “The model deceived in a test” it does not follow that “The model wants something.” These tests are deliberately constructed as extreme situations in order to make vulnerabilities visible. They show what a system can do under pressure — not what it is planning. Anyone who uses the term should keep this distinction in mind.