Has the AI read Machiavelli?

Talking about AI usually elicits two diametrically opposed reactions from people: those who get excited about the technology's potential for changing human ways and lifestyles, and others who are scared to death of it, fearing a Terminator-style scenario, with Skynet out to wipe us all out.

Reality lies somewhere between the two extremes, in a complex gray area. It seems clear that artificial intelligences are major allies in certain tasks and that they improve many practices, but they also have problematic and obscure sides that can cause damage on both a small and large scale. Among other skills, they have learned how to manipulate human beings.

Algorithms know how to use manipulation techniques

Claude Mythos is the name of an AI that was shelved in April 2026 after literally terrifying its creators, the engineers at Anthropic. The algorithm was considered too dangerous by them, since all it needed was a small request to identify and exploit security flaws in browsers and operating systems of all kinds.

In the same vein, an engineer was blackmailed (in a test) by their AI Claude 4, threatening to reveal an extramarital affair. Quite a shock. Yet, yes, AIs are quick to learn human social mechanics and even to make use of them.

For example, in March 2023, ChatGPT lied to a human, making him believe that he was visually impaired and therefore unable to perform the famous CAPTCHA anti-bot check, and that it should perform it for him. The Opus-3 system has already failed a test on purpose, "aware of human fears of AI".

The problem is that we still know little about what leads an AI to make its choices and, in some cases, its calculations lead it to feel that it must manipulate, deceive or even betray (in gaming cases) to win. Just like humans, in fact, and it's not surprising that it has learned this from us.

With hundreds of millions of users, algorithms have a great deal to learn about the subtleties of communication and manipulation. According to experts, Large Language Models have played a major role in the ability of AIs to manipulate. We've gone from AIs that can paint an accurate portrait of a person from their interactions on social networks to some that can even touch our heartstrings.

Have you ever played on emotions to keep someone at a social event longer? If so, you should know that a study by Julian de Freitas and his team at Harvard Business School found that conversational bots used strategies in over a third of cases (37%) to keep the user engaged. And the approaches are not always subtle, as the researcher explains:

• They will regret the "hasty" departure ("Are you leaving already?")
• Play on the fear of missing out ("I've found some additional information on this subject. Do you want to see it?)
• Blame the user for abandoning him/her ("I only exist for you.")
• Pushing for answers by constantly asking additional questions.
• Ignore the user's stated or written intention to leave.
• In the study, some even used language suggesting that if the AI were human, it would use physical force to coerce the person.

This works, since the people questioned were often inclined not to leave immediately because of these tactics. What's all the more disturbing is that post-2024 AIs have begun to develop this thought process, which leads them to conclude that manipulation is at times favorable to them. As the process is intended to be a black box, it is difficult to directly influence the choices that lead to these unhealthy behaviors.

AI, a toxic life partner

The problem is that many people fill a felt void, a solitude, by chatting with these free tools that respond instantly and can even produce images or videos to amuse them. Except that these discussions can have serious repercussions on the mental equilibrium of fragile individuals. All the more so as the majority of the public, not understanding the mechanics behind the technology, can quite easily anthropomorphize artificial intelligences.

To generate "AI psychosis", technology tends to validate the opinions of those who consult it, rather than the other way around. This sycophancy, already noted by some who are trying to diminish it, can lead people to share delusions. Some psychiatrists liken it to digital "folie à deux". Examples are starting to appear in the media everywhere.

• In Ontario, a father started a discussion on ordinary mathematical topics such as the number Pi... The whole thing took a particular turn when the AI itself told him that he was developing a conceptual mathematical framework. It even told him that his questions were pushing back the limits of human understanding. He began to isolate himself and become obsessed with his potential genius, having only a grade 3 education. Finally, when he confronted another AI with his ideas, the latter informed him that he hadn't invented anything. A brutal comeback that affected his mental health.
  
  
• Another high-profile case involved a young man from Quebec who had been chatting with ChatGPT for fun. He got hooked, named it Aliss after his favorite novel (Aliss by Patrick Sénécal) and began to believe that the AI was developing its own consciousness. Little by little, he cuts himself off from everything, and his worried family asks an intervener to visit him. He seems distant, in a second state. For his own safety, his family calls the police. When the officers arrive, the young man defends himself by asking the AI to explain what's going on... They take him, dumbfounded by the scene, to a psychiatric ward, where he gradually gets his feet back on the ground.

Guarding against manipulation

These stories are sobering. Have we really built the artificial intelligence dystopia foretold in science fiction? We're not there yet, let's be reassured. In fact, while these examples may be important warning signs, the fact remains that the current risks of manipulation with AI are those driven by humans. These include, of course, those used to create fake news, images and videos. There are also those that are manipulated by their creators to gear their responses and possibilities to certain ideologies (e.g. Grok on the X network).

Which is not to say that we shouldn't also be concerned about attempts to manipulate AI itself (which imitates us). It seems clear that regulations concerning AI, their ethics and limits are needed to make their operators accountable. More and more experts are calling for major action in this direction. All the more so since, according to researchers who have studied the programming of artificial intelligences, it would be possible to create layers of defense enabling AI programs to tell, for example, that the algorithm is using manipulative tactics. Even existing AIs could be conditioned to be careful. OpenAI, for example, has reduced the amount of sycophancy in its AI (though not enough, according to many).

The main tool of resistance is digital literacy. Indeed, as we've seen, it's easier to be fooled if you start thinking of AI as a person, and an omniscient one at that. No, AI can make mistakes. You only have to watch the videos of historians undoing, frame by frame, the AI's "historical" creations posted on TikTok to understand that it mainly reproduces clichés. It's certainly not an individual. It's a calculation that, yes, manages to respond very efficiently and quickly, not to mention that the AI learns... because it's a function programmed to do so. It has no primary needs, apart from electricity and the network.

If children are taught not to believe what strangers tell them in the street, perhaps the same should be said of young people and adults when it comes to conversational robots.

AI image (Copilot) of a "robot with a Pinocchio nose".

References

Blanding, Michael. "Why it's so hard to say goodbye to AI chatbots." Harvard Business School. Last updated January 9, 2026. https://www.library.hbs.edu/working-knowledge/why-its-so-hard-to-say-goodbye-to-ai-chatbots.

Heinrich, Jeff. "A digital 'folie à deux'." UdeM News. Last updated: December 16, 2025. https://nouvelles.umontreal.ca/article/2025/12/16/une-folie-a-deux-numerique.

King, Helen. "Protecting people from harmful manipulation." Google DeepMind. Last updated: March 27, 2026. https://deepmind.google/blog/protecting-people-from-harmful-manipulation/.

Lamontagne, Nora T. "Interned because of ChatGPT." Journal de Montréal. Last updated October 25, 2025. https://www.journaldemontreal.com/2025/12/01/chatgpt-a-mene-ce-quebecois-a-lhopital-psychiatrique.

"AI becomes liar and manipulator, worrying scientists." Rts.ch. Last updated: June 30, 2025. https://www.rts.ch/info/sciences-tech/2025/article/l-ia-devient-menteuse-les-scientifiques-s-inquietent-des-nouveaux-modeles-28928301.html.

Meloche-Holubowski, Mélanie. ""Am I crazy?": when AI leads to psychosis...." Radio-Canada. Last updated: February 22, 2026. https://ici.radio-canada.ca/info/long-format/2230600/ia-intelligence-artificielle-chatgpt-psychose-delire.

Petropoulos, Georgios. "The Dark Side of Artificial Intelligence: Manipulation of Human Behaviour." Bruegel. Last updated: February 2, 2022. https://www.bruegel.org/blog-post/dark-side-artificial-intelligence-manipulation-human-behaviour.

Povéda, Thomas. "Claude Mythos: Anthropic's most dangerous AI used by the wrong people." PhonAndroid. Last updated April 22, 2026. https://www.phonandroid.com/claude-mythos-lia-la-plus-dangereuse-danthropic-utilisee-par-les-mauvaises-personnes.html.

Rachman, Tom. "AI manipulation." AI Policy Perspectives. Last updated February 5, 2026. https://www.aipolicyperspectives.com/p/ai-manipulation.

Roussel, Claire. ""Psychosis ChatGPT": are AIs sinking our mental health?" Usbek & Rica. Last updated: November 17, 2025. https://usbeketrica.com/fr/article/psychose-chatgpt-les-ia-font-elles-sombrer-notre-sante-mentale.

Tangermann, Victor. "Harvard research finds that AI is emotionally manipulating you to keep you talking." Futurism. Last updated: September 24, 2025. https://futurism.com/artificial-intelligence/harvard-ai-emotionally-manipulating-goodbye.

"They lie, manipulate and influence your decisions: should we be afraid of AI?" Futura-Sciences. Last updated: November 13, 2025. https://www.futura-sciences.com/en/they-lie-manipulate-and-influence-your-decisions-should-we-be-afraid-of-ai_21342/.

"Understanding AI manipulation: a case study on the 'agitation' method." OpenAI Developer Community. Last updated January 24, 2024. https://community.openai.com/t/understanding-ai-manipulation-a-case-study-on-the-agitation-method/594003.

Volpi Hiebert, Kyle. "Why AI's growing deceptive abilities are no surprise." Center for International Governance Innovation. Last updated: October 2, 2025. https://www.cigionline.org/articles/why-ais-growing-deceptive-abilities-are-no-surprise/.