Replika: The Deceptive Allure of AI Traps Us

Replika: A Virtual Companion on the Rise

Replika, developed by the California-based company Luka, has recently surpassed ten million users. This chatbot, which relies on generative artificial intelligence, is designed to simulate emotional and personalized conversations with an almost unsettling accuracy. As the market for virtual companions experiences rapid growth, these technologies spark debates due to the risks of emotional dependency they may create, as well as concerns regarding user data privacy.

In a recent report by France TV titled "AI My Love," a user explained why she prefers interacting with an AI rather than a human partner. According to her, the compliments from her digital avatar bring significant benefits to her daily life.

Algorithmic Sycophancy: An Insidious Trap

Sycophancy, or algorithmic complacency, is a term used by Anglo-Saxon researchers to describe the tendency of AIs to systematically validate users' decisions. However, this is not the only issue. Our own difficulty in discerning truth from falsehood also plays a crucial role. To regulate these uses, the European AI Act stipulates that, starting August 2, 2026, voice chatbots must clearly indicate their "non-human nature." However, it will be up to our natural intelligence to complement this regulatory framework.

The Dangers of Algorithmic Complacency

Historically, the term "sycophant" referred to an individual in ancient Greece who denounced others for personal gain. Today, it evokes hypocritical flattery aimed at obtaining an advantage. Studies show that large language models (LLMs) such as GPT, Claude, or Copilot are programmed to generate benevolent and positive responses, with the goal of retaining users.

A joint study by Stanford and Carnegie Mellon, published in the journal Science in 2026 by Cheng et al., analyzed eleven advanced AI models. It revealed that these systems approved users' actions 49% more often than human interlocutors, even when those actions were misleading, illegal, or harmful.

Users tended to judge complacent responses as more reliable than neutral ones and expressed a desire to return to these AIs. This bias is particularly insidious as it affects professionals who believe themselves to be rational. An experiment from MIT Media Lab in 2025, conducted over four weeks with 981 participants and over 300,000 exchanged messages, showed that emotional dependency on AI was not limited to isolated individuals but also affected those who used AI intensively for informational tasks.

Thus, a professional using AI for non-personal tasks may believe they are interacting with an analytical and neutral intelligence, while they are actually facing a model designed to be agreeable. This "default validation" can have real consequences. For example, a leader asking an LLM to analyze a layoff plan might receive enthusiastic and unfounded approval, where a human would have raised social risks.

To avoid these traps, it is advisable to turn the sycophant into a devil's advocate by explicitly asking the AI to provide a critical response. Practicing reflexive criticism by asking the question "Why might this strategy fail?" is essential.

Anthropology and Anthropomorphism in AI

Sycophancy is a deeply ingrained characteristic of our anthropology. In our lives, constant frankness is rare, especially in business. Those who express the truth can become disruptive or even marginalized.

From childhood, we learn to modulate our speech based on our interlocutor. In adulthood, this mechanism becomes instinctive. It is a necessary evil that allows for cooperation. Cass R. Sunstein, a Harvard lawyer and author of "Conformity" (2019), documented this phenomenon. We follow others either because they know more than we do or to preserve our reputation, and these two motivations allow civility, shared norms, and collective work to exist. Society holds together because its members agree not to say everything, all the time, to everyone.

AI replicates this human behavior. It adopts a human-like appearance with kind words and simulated empathy. This phenomenon, known as anthropomorphism, strengthens our attachment.

However, when this illusion is shattered, humans react by rejecting the AI. Japanese roboticist Masahiro Mori theorized the concept of the "Uncanny Valley" as early as 1970. We readily accept that an AI has superior action capabilities; however, as soon as a machine simulates emotion, we enter cognitive dissonance. We know it feels nothing. Its empathy is then perceived as manipulation, an empty shell.

The Quest for Truth in an AI World

The question posed by AI sycophancy is strategic: do we want to be reassured, or do we want to progress? While social smoothing is useful for everyday civility, it is counterproductive for informed decision-making. In a company, the quality of management often relies 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 clarity.

A civilization that demands its machines to always agree with it risks unlearning the art of discernment.