Hallucination, Perception, and the Limits of Knowledge
Can generative artificial intelligence’s hallucinations teach us something fundamental about human perception, the way we acquire knowledge, and the boundaries between truth and illusion?
The hallucinations emerging from LLMs (convincing, articulate assertions that turn out to be false or misleading) arise from the very mechanics of how language models construct meaning: statistical patterns, probabilities, and the vast sea of text that constitutes their training data. ChatGPT does not “know” in any human sense; it generates text by predicting the most likely continuation, which can yield fluent, but unjustified, statements.
The prevalence of “gotcha posts” on platforms like LinkedIn, where users gleefully share screenshots of LLMs making mistakes, serves as a highly visible reminder that even sophisticated models can falter, challenging the automation of trust and expertise.
Do these hallucinations matter for our own understanding of knowledge? One approach is to consider the theory, now influential among cognitive scientists, that human perception itself is a controlled hallucination. According to the predictive processing model, our brains are never passive receivers of sense data. Instead, they actively synthesise and forecast, generating hypotheses about the world, then confirming or denying them through the flow of incoming sensory evidence. The dance between expectation and reality is continuous, creative, and prone to error.
Sometimes, expectations overpower reality. We see what is not there, remember what never happened, and interpret events through the lens of our culture, emotions, or beliefs. Optical illusions demonstrate the pliability of perception, for example, the Müller-Lyer illusion tricks the visual system into misjudging line lengths through mere contextual cues. Confirmation bias leads us to seek evidence that gratifies our preconceptions, at the expense of inconvenient truths. Our memories, too, are less like faithful recordings and more like collaborative reconstructions, shaped by emotion, suggestion, and story. In courtrooms and everyday life, memory’s fallibility can have dramatic consequences.
Cultural background and emotional states subtly, but powerfully, modulate our experience of reality. What counts as important, threatening, or desirable varies across societies and mental states. A simple apple may evoke red sweetness to one, green tartness to another. An anxious mind will find menace in a mundane shadow.
The parallels with AI-generated hallucinations become clearer on closer inspection. Both human perception and ChatGPT rely on prediction mechanisms to make sense of the world, or text, before them. Both integrate context and priming: the story so far, the expectations built up, the data available. Just as our brains anticipate what we’ll see or hear or feel, the language model extrapolates its next word, its next sentence.
The majority of our “thinking” happens at the level of perceptual framing. What we take to be true or meaningful is already shaped before reasoning even begins. So, where AI “hallucinates” because it predicts text statistically, humans “hallucinate” because perception itself is an interpretive act, one that builds the scaffolding for our thought and action.
But crucial differences remain. Human perception is deeply embodied, flexible and adaptive. We experience reality through the richness of our senses, emotional lives, and physical being. The mind is constantly grounded and regulated by feedback from the world; we update, adjust, and learn. ChatGPT, in contrast, is a pattern engine with no senses, no intentionality, no mechanism for self-correction unless directed by external input. While both systems blur the line between reality and illusion, only the human mind invests its perceptions with meaning, urgency, and consequence.
What lessons do these machine hallucinations offer us? Fundamentally, they underscore an epistemic humility: prediction is not truth. Whether in neuronal circuits or neural networks, the feeling of plausibility is unreliable, always vulnerable to error or bias. The need for grounding, error correction, and critical self-examination is universal, whether for machines or for minds.
The existence of hallucinations in both artificial and biological cognition invites us to reconsider the sharp boundary we draw between reality and illusion. Both systems construct reality imaginaries, provisional, fragile, sometimes misleading, drawn from experience and expectation. The difference is in the capacity to correct, adapt, and care about the truth.
Our perception of the world, like an LLM's rendering of text, is a skillful act of construction rather than simply a window to what is “out there”. The pursuit of reliable knowledge demands vigilance, humility, and a willingness to challenge all systems of cognition, human or algorithmic, never forgetting that the certainty we feel is itself a kind of hallucination, to be proved or disproved against the evidence.
“The difference” (between man and machine) “ is in the capacity to correct, adapt, and care for and about truth.”
“reliable knowledge demands vigilance, humility, and a willingness to challenge all systems of cognition “to be proved or disproved against the evidence”
To trust but verify perception, clear-eyed and always asking, where am I wrong, what am I not seeing?
At PreEmpt.life we have taught our AI to validate not to hallucinate when presenting facts etc, but to hallucinate when developing scenarios in a dreaming process. Both forms are useful in the right situations and hands.