Humans and AI Alike Fill Gaps with “Plausibility” Under Constraints

Why inference systems — human and artificial — trade accuracy for coherence under pressure

This article was originally written in my native language, Japanese, and has been translated into English by AI. You can read the original Japanese version here.

Imagine you’re suddenly asked for your opinion in a meeting. Or the clock is ticking on an exam. Or your boss or client demands an explanation on the spot. In moments like these, we stitch together fragments of memory, experience, expectations of the other person, and the mood of the room to construct an answer that at least holds together.

It isn’t necessarily a lie. In fact, the person answering is often genuinely trying their best. “It was probably like this,” “Given the flow of the conversation, this seems the natural conclusion,” “I think I heard something similar before.” These vague materials are instantly assembled into a single explanation.

And sometimes, it’s wrong.

What Is AI Hallucination, Really?

When we hear the term “AI hallucination,” we tend to think of it as a flaw unique to machines. An AI cites papers or legal cases that don’t exist, even though it has no actual knowledge of them. It offers explanations with little basis in a calm, composed tone. The answer is wrong, but the writing is curiously polished.

That is certainly dangerous. And yet, there’s something about it that resembles humans.

From a cognitive science perspective, AI hallucination is said to be closely related to a human psychological phenomenon called confabulation. According to a research article in The Conversation, both the human brain and AI generate “plausible guesses” when filling knowledge gaps — making similar kinds of mistakes. However, the underlying mechanisms are fundamentally different.

Large language models (LLMs) are not systems that retrieve truth; they are systems that statistically predict “what word comes next.” As MIT Sloan’s analysis points out, LLMs are optimized for “plausibility” rather than accuracy. What answer naturally follows a question? What phrasing feels convincing to a human? What structure seems intellectually credible? As a result, AI errors are often beautiful — grammatically correct, clearly organized, and calmly delivered. The problem is that fluency and correctness are two different things.

Humans Fill Gaps Too

Humans, too, struggle to tolerate emptiness. Not knowing. Not remembering. Not having enough to judge by. Remaining silent in that state is harder than it sounds. Silence might make you look incompetent. It might make the other person uneasy. It might stall the conversation.

So we fill the gaps. We patch holes in memory with guesses. We fill the lack of evidence with rules of thumb. We fill uncertain judgments with plausible narratives.

Words like “probably,” “generally speaking,” “if you think about it normally,” and “I think” are, in principle, markers of uncertainty. Yet as we speak, those guesses gradually begin to take on the face of conviction. When an explanation fits together cleanly, it feels correct. The human brain is not only a device for storing facts — it is also a device for constructing stories.

Try this: have someone say “silk” ten times in a row, then ask, “What do cows drink?” Most people instinctively answer “milk” — even though they know perfectly well the correct answer is water. The word “silk” has primed their mind, and “milk” flows out before the actual answer can catch up.

(A similar trick works in Japanese, where “pizza” (piza) and “knee” (hiza) sound remarkably alike. Say “pizza” ten times, point to your elbow, and ask “What’s this?” — many people say “knee” instead of “elbow” (hiji), pulled there by the phonetic echo still ringing in their head.)

This isn’t a lack of knowledge. It’s the brain’s own mechanism by which prior context pulls a “plausible answer” to the surface. Constraints don’t only arise from lack of time or information. They’re also created by cognitive priming. And in those moments, a “contextually fitting answer” comes out of our mouths before an accurate one does.

A paper published in PLOS Digital Health proposes reconceptualizing AI hallucination as “confabulation.” Human confabulation occurs when we fill gaps in memory — but AI also generates statistically “plausible” answers to fill in the blanks when it encounters patterns absent from its training data. The underlying mechanisms differ, but the act of “filling gaps with narrative” is common to both.

Fluency and Correctness Are Different Things

A confident person isn’t necessarily right. A quick respondent isn’t necessarily a deep thinker. A person who explains things clearly doesn’t necessarily truly understand. We routinely treat the smoothness of speech as evidence of intelligence. And AI reproduces that smoothness at an extremely high level.

According to MIT Sloan’s analysis, users tend to build trust in AI based on fluency, tone, and authoritative delivery, and are prone to overlooking inaccuracies when not corrected. In other words, fluency and confidence don’t just fail to serve as evidence of correctness — they can actively make errors harder to detect.

So viewing AI hallucination merely as “evidence of lacking intelligence” is a bit too simple. Rather, it may be a failure common to any system that must make inferences under constraints. Not enough time. Not enough information. Silence is not permitted. An answer must be given. In those moments, both humans and AI fill the missing parts with “plausibility.”

Where the Difference Lies

Of course, there are significant differences between humans and AI. Humans have bodies, experience, and accountability. If we’re wrong, we feel embarrassed. A mistake at work costs us credibility. Hurting someone leaves us with regret. Contact with reality gradually corrects our thinking.

AI carries little of that weight. When it gives a wrong answer, the AI itself feels no shame. It bears no responsibility. It feels no pain. As the PLOS Digital Health paper notes, human confabulation has an element of “emotional defense,” which connects to self-awareness and motivation to correct. AI’s “confabulation,” however, has no such emotional dimension. While it makes statistically similar errors, it is fundamentally different in that it has no intrinsic motivation to self-correct.

Society Already Has Mechanisms for Verification

Yet the same is true, in essence, for human answers as well. Expert opinions undergo scrutiny. Research papers go through peer review. Newspaper articles are copyedited. Medicine has second opinions. Society has built systems of verification and correction on the premise that humans make mistakes.

What the AI era requires is not to single out AI for special suspicion. It is to start from the premise that both humans and AI make mistakes.

“Plausible answers” are useful. They serve as starting points for thinking. They move discussions forward. They function as provisional maps for venturing into the unknown. But the moment we mistake them for truth itself, usefulness turns into danger.

To Avoid Being Intoxicated by Plausibility

The way we look at AI is, at the same time, the way we look at humans. And probably, the way we look at ourselves.

How much evidence lies behind a fluent answer? How much verification underlies a confident tone? How much reality remains outside a tidy narrative?

What we need is not to deny plausibility. It is to avoid being intoxicated by it.

Both humans and AI fill gaps with “plausibility” under constraints. That is not a defect of intelligence — it is a common property of any inference system operating under constraints. Understanding that property, and cultivating the habit of verification and correction: that, perhaps, is the condition for honest thinking — for humans, and for the humans who use AI.