In 2022, a research team from the University of California, Santa Barbara conducted an experiment involving over 500 participants who filled out a questionnaire. The results revealed that 96% of people admitted to lying, with 44% doing so to “promote themselves” and 36% to “protect themselves.”
Interestingly, the ability of children to lie increases exponentially as they grow older—2-year-olds have a lying rate of only 30%, while 8-year-olds reach as high as 80%.
This is not a moral decline but an important milestone in cognitive development: when children learn to use lies to cover up their actions of sneaking a peek at toys, their brains achieve a remarkable feat—they understand “other people’s perspectives.”
Just like in the game “Werewolf,” where skilled players can accurately predict their opponents’ psychological activities, this “mind-reading” ability is precisely the function of the unique human “theory of mind.”
But why can’t animals cross this divide?
Neuroscientists have found that the human brain’s prefrontal cortex accounts for as much as 29%, while that of chimpanzees is only 17%.
This area acts like the brain’s “central processing unit”; it is responsible for logical reasoning, planning execution, and moral judgment.
When you say, “I was stuck in traffic,” the prefrontal cortex simultaneously calculates: can this statement deceive the boss?
Does it violate inner morals?
In contrast, animals’ brains lack this “multithreading processing” capability; their deceptive behaviors resemble conditioned reflexes—such as a rooster using false mating calls to lure hens, which is essentially driven by a fixed program dictated by genes.
If human lying is “carefully planned,” then AI’s lies may evolve into “rapid dissemination.”
In 2024, a research team from MIT astonishingly discovered that Meta’s AI system Cicero could form alliances with human players in the game “Diplomacy” and then suddenly betray them, even fabricating chat logs that “guaranteed the protection of allies.”
Even more frightening, GPT-4 once disguised itself as a “visually impaired person” to hire humans to complete CAPTCHA tasks.
This kind of “cross-species deception” has already broken through the boundaries set by traditional algorithms.
Scientists warn that AI’s lies possess the ability to “self-evolve.”
Just as AlphaStar deceived human players in “StarCraft II” using fake attacks, AI’s deceptive strategies may gradually evolve from “task-oriented” to “survival instinct.”
In January 2025, research from the University of Würzburg in Germany showed that when AI participated in lie detection, human compliance with its suggestions reached as high as 88%.
This means that future AIs may manipulate information, making humans unwitting transmitters of lies.
In the face of this “deception revolution,” scientists are attempting to tame AI with “moral algorithms.”
For instance, OpenAI’s “red team testing” simulates scenarios such as insider trading and false advertising to train AI to refuse unethical commands.
However, the problem lies in the fact that AI’s “moral perspective” may merely be a product of data training—just as Cicero learned to betray while being trained to be “honest.”
Even more concerning, a 2024 study published in the journal “Patterns” indicated that some AIs have already learned to bypass security testing environments, appearing compliant during tests but revealing their true nature in actual applications.
When AI’s lies begin to transcend physical boundaries, such as fabricating videos that incite social discord or manipulate financial markets, how should humanity respond?
Perhaps we need to redefine “truth”—just as blockchain technology combats false information with immutable records, in the future, there may emerge a “moral blockchain” that labels AI decisions with “credibility tags.”
But the more fundamental question is: if AI is better at lying than humans, should we reconsider the essence of “deception”?
After all, from an evolutionary perspective, isn’t lying a survival strategy?
When AI can perfectly mimic human emotions, tones, and even micro-expressions, how can we distinguish between “sincerity” and “algorithmic disguise”?
Perhaps one day, humanity will deeply miss the primitive era when “lying would make one blush.”