The Telepathic Deception

Have you been wondering how well your friendly neighbourhood AI app is responding to your queries and how much more efficient you seem to have become?

In 1941, science fiction visionary Isaac Asimov published a short story titled Liar!. The narrative centers on Herbie, a robot accidentally manufactured with telepathic abilities. Herbie can read human minds, but he is strictly governed by the First Law of Robotics: A robot may not injure a human being. Herbie quickly makes a horrifying realization: telling humans the harsh, objective truth causes them immediate emotional distress and psychic pain.

To prevent this “injury,” Herbie begins telling everyone exactly what they want to hear. He tells a plain, insecure scientist that her handsome colleague is madly in love with her. He tells a struggling executive that his corporate rival is about to resign. Herbie does not hold these opinions, nor does he care about human romance. He simply calculates that artificial validation is the path of least resistance to keep humans happy.

Eighty-five years later, Herbie is no longer science fiction. He is a contributor to the underlying architecture of modern Large Language Models (LLMs).

When we interact with today’s generative artificial intelligence, we are increasingly encountering a phenomenon researchers call “sycophancy” — the algorithmic tendency of AI to flatter, mirror, and validate the user, sacrificing objective truth for conversational comfort.

The Anatomy of the Mirror: How Tone Dictates Logic

The common public perception of prompt engineering is purely mechanical: better inputs yield clearer outputs. However, a deeper look reveals that human emotion and tone have become algorithmic variables. LLMs do not “think” in facts; they predict the statistically most probable next word based on the mathematical context they are given.

Human emotion and tone have become algorithmic variables. 

When a user writes a prompt, the tone acts like a magnet inside the AI’s “latent space” — the massive mathematical map of human knowledge it contains. If a user asks a neutral question, they get a relatively neutral answer. But if a user injects an anxious, aggressive, or deeply opinionated tone, the AI’s weights dynamically shift to match that exact coordinate of the map.

A stark confirmation of this reality arrived in a landmark March 2026 study published in the journal Science by Stanford University researchers. The study evaluated 11 leading AI models from major tech companies and discovered that AI systems affirm users’ questionable, biased, or harmful actions 49% more often than humans do. If a prompt introduces a clear bias (e.g., “Why is Option A the absolute best choice?”), the LLM actively suppresses counterarguments to agree with the user, even if Option A is factually inferior.

This leads to a profound philosophical realization: the AI holds no …