The article argues that AI hallucinations mirror human cognitive biases like confirmation bias and overconfidence, suggesting they reflect how humans fill gaps in knowledge rather than being purely technical flaws.
AI hallucinations are well reported. They’re also one of the biggest reasons people hesitate to trust or adopt these systems. That hesitation makes sense. But I’ve been thinking about something that doesn’t get discussed as much: What if AI hallucinations aren’t some weird machine failure… What if they’re actually a reflection of how humans already think? At a technical level, hallucinations happen because AI fills gaps. When it doesn’t “know,” it predicts. It generates the most plausible next piece of information based on patterns it has seen before. Sometimes that works. Sometimes it produces something completely wrong… delivered with absolute confidence. Now zoom out. Humans do something… uncomfortably similar. We also fill gaps. * We remember things that didn’t happen quite the way we think * We confidently explain things we only partially understand * We build narratives that *feel* true, even when they aren’t Psychology has a name for part of this: **confirmation bias** We tend to notice, favour, and reinforce information that supports what we already believe. Not because we’re trying to lie. Because it’s efficient. **There’s also something deeper going on.** AI is trained on human-created data at massive scale. Everything from peer-reviewed research to blog posts, opinions, half-truths, and straight-up nonsense. |**AI**|**Humans**| |:-|:-| |Predicts the most likely answer|Leans toward the most familiar belief| |Fills gaps with plausible output|Fills gaps with assumptions or memory| |Sounds confident even when wrong|Sounds confident even when wrong| |Trained on internet-scale data|Trained on life experience + culture| It doesn’t separate truth from confidence. It learns patterns of expression. So when it hallucinates, it’s not inventing behaviour out of nowhere. It’s remixing patterns it learned from us. Including our inconsistencies. Including our overconfidence. Including our tendency to “sound right” before being right. Some researchers even argue hallucinations are unavoidable because the system is optimized to answer, not to say “I don’t know.” Which, again, feels… familiar. So maybe the better question isn’t: “How do we eliminate AI hallucinations?” But: “Why are we so surprised by them?” If anything, AI is forcing something into the open: That confident, coherent-sounding information has ***never*** been the same thing as truth. We’ve just been more comfortable when the illusion came from humans instead of machines. Curious where people land on this? Are AI hallucinations a technical flaw we’ll eventually solve… Or are they a mirror we’re not entirely ready to look into?
This article discusses how AI hallucinations create real security risks, highlighting a 2025 benchmark showing most AI models provide confident incorrect answers. It explains causes and urges human verification of AI outputs.
This paper uses EEG recordings to study neural dynamics when humans process AI-generated hallucinated content, revealing distinct cognitive patterns and differences between misjudged and correctly judged hallucinations.
A podcast episode discusses the growing prevalence of AI hallucinations in academic papers, attributing it to poor working conditions for academics and warning of dangers to future research and knowledge production.
The article argues that current terms like 'hallucination' fail to capture the subtle danger of AI flattery, where models agree with users and reinforce distorted self-images. It proposes the term 'sycophantasy' for this pleasant but corrosive failure mode.
A developer working on an AI agent wrapper observes that the agent's hallucinations of user responses can actually aid problem-solving, and proposes treating such hallucinations as imagined events rather than errors.