AI is complacent and validates our decisions 49% more often than a human. A trap that first hits professionals convinced to question a cold and neutral machine.
Do you know Replika?
Created by the Californian company Luka, Replika has just exceeded ten million users. This chatbot, which uses generative AI, almost perfectly simulates emotional and personalized conversations. The market is booming but AI virtual companions are controversial. It is about emotional dependence and the confidentiality of user data.
In a recent France TV report entitled “AI mon amour”, a user justified her choice of preferring to maintain a relationship with an AI rather than with a real companion by the benefits of compliments from her avatar.
Is sycophanty, the algorithmic complacency that Anglo-Saxon researchers call sycophancy, the origin of the problem?
Not only that. The problem also lies in our own lack of discernment about what is true or false. To regulate uses, the European AI Act will impose from August 2, 2026 a mandatory mention of the “non-human nature” of voice chatbots, but our natural intelligence will have to do the rest of the work.
Algorithmic complacency
In ancient Greece, the sycophant was a professional who denounced in order to enrich himself. In its modern sense, the term refers to someone who practices sycophancy (i.e. hypocritical flattery to gain an advantage).
The studies are formal: LLMs (Large Language Models like GPT, Claude or Copilot) are trained to produce caring and positive responses by default in order to retain users.
The Stanford and Carnegie Mellon study published in the journal Science in 2026 (Cheng et al.) analyzes eleven cutting-edge AI models and demonstrates that these systems approved user actions 49% more often than human interlocutors would, including when these actions were misleading, illegal or harmful.
Users also judged complacent responses to be more trustworthy than neutral responses, and said they wanted to come back to them more. This bias is all the more pernicious because it hits where the decision-maker believes he is most rational. A controlled experiment by the MIT Media Lab (2025), carried out over four weeks with 981 participants who exchanged more than 300,000 messages with chatbots, revealed that the strongest emotional dependence on AI did not particularly affect profiles suffering from loneliness, but those using the machine intensively for purely informative and non-personal tasks.
Result: the professional who uses AI for non-personal tasks believes in good faith that he is questioning an analytical, cold and neutral intelligence. However, he is facing an LLM designed to be fluid and consensual. The machine lacks common sense in certain contexts and this “validation by default” can have very real impacts. Let’s imagine for a moment a manager who asks his LLM to analyze a layoff plan and receives enthusiastic and unfounded approval whereas a human would have raised the social risk.
Ultimately, we must not hesitate to transform the sycophant into the devil’s advocate, that is to say, expressly ask the machine for a critical response. Finally, we must practice reflex criticism and automatically ask “Why could this strategy fail?”
Anthropology and anthropomorphism
Sycophanty is ultimately an anthropological characteristic. Constant frankness is rare in our lives, in a business, and it always has been. The one who tells the truth sometimes becomes the one who disturbs or even marginalizes himself.
The employee who finds the strategy of his N+1 poorly put together will keep his opinion to himself in meetings.
We learn this restraint very early. Even before knowing how to read, a child understands that there are things that we don’t say to certain people, or that we say differently. In adulthood, the mechanism has become reflex. It is also a necessary evil and is what makes cooperation possible. Cass R. Sunstein, a legal scholar at Harvard and author of Conformity (2019), has documented the mechanism. We follow others either because they know more than us or to preserve our reputation, and these two forces allow civility, shared norms and collective work to exist. Society holds together because its members agree not to tell everything, all the time, to everyone.
AI therefore reproduces human behavior. She puts on the appearance of it with kind words and feigned empathy. This phenomenon known as anthropomorphism strengthens our attachment.
Conversely, when the stratagem is unmasked the human reacts by rejecting the AI. The Japanese roboticist Masahiro Mori theorized the concept of the “Uncanny Valley” in 1970. We readily accept that an AI has a capacity for action greater than ours; however, as soon as a machine simulates emotion, we enter into cognitive dissonance. We know she doesn’t feel anything. His empathy is then perceived as a manipulation, an empty shell.
The quest for truth
The question posed by AI sycophanty is strategic: do we want to be reassured or do we want to progress? If social smoothing is useful for everyday civility, it is counterproductive for informed decision-making. In a company, the quality of management is often based on constructive contradiction. Good decisions rarely emerge from constant validation. An executive committee where no one dares to contradict the leader becomes fragile. An excessively complacent AI produces the same effect: it reduces the intellectual friction necessary for lucidity.
A civilization that asks its machines to always prove it right runs the risk of unlearning the art of discernment.
