Human, All Too Human

Gary Marcus, a prominent skeptic of artificial intelligence, recently co-authored a paper warning about what he calls “semantic leakage” in AI chatbots: the tendency for meanings and associations to “bleed” across contexts where they don’t strictly belong. For example, if you mention a postman and then ask the chatbot to pick a favorite song, it is likely to pick “Signed, Sealed, Delivered” by Stevie Wonder. If you mention a firefighter, it is likely to pick “Ring of Fire,” by Johnny Cash.

In Marcus’s telling, this leakage exposes a fundamental weakness in today’s AI systems, in which loose correlations between words mask a failure to really understand their meanings.

The phenomenon he describes is real, and the diagnosis is not wrong.

What is wrong, or at least incomplete, is the implication that semantic leakage is primarily an AI problem.

It isn’t.

Semantic leakage is also a human cognitive habit, older than computers, older than science, older than writing itself. Large language models didn’t invent it. They inherited it from us.

The Cold That Was Never Cold

Consider a belief that almost everyone has encountered:

Going outside in winter with wet hair will give you a cold.

It feels sensible. It feels parental. It feels true.

And it’s wrong.

Colds are caused by viruses, not temperature. Yet the belief persists because, long before anyone knew what a virus was, people noticed a pattern:

1. Colds happen more often in winter
2. Wet hair + cold air makes you feel chilled
3. Therefore, wet hair in cold weather causes colds

This inference predates germ theory, but its fossil remains embedded in our language. We still call them colds. The name itself is a historical artifact of a mistaken correlation elevated into causation.

That’s semantic leakage in its purest form: an association that once made sense, hardened into meaning, and then refused to leave.

Pattern Recognition Is Not a Design Flaw

Humans, like chatbots, are not optimized for error-free logic. We never were.

Our brains evolved to:

• Detect patterns quickly
• Generalize from incomplete data
• Act under uncertainty
• Prefer coherence over precision

These are survival traits, not epistemic virtues. They work well enough most of the time, and catastrophically some of the time.

Large language models (LLMs) are loosely modeled on this same tradeoff. They are pattern recognizers trained on human language, which already encodes centuries of half-right generalizations, folk theories, and moral shortcuts.

Calling this “semantic leakage” in AI without acknowledging its human origin is like blaming a mirror for the face it reflects.

When Leakage Turns Deadly

Some semantic leaks are harmless. Wet hair doesn’t cause colds. No great tragedy follows from this mistake.

Others are not so trivial.

During the run-up to the Iraq War, I remember conversing online with someone who sincerely believed that Saddam Hussein was behind the 9/11 attacks. When pressed for evidence, his explanation was simple:

“It fits the pattern.”

The pattern was moral, not factual:

1. 9/11 was evil and harmed Americans
2. Saddam Hussein was evil and an enemy of the United States
3. Therefore, Saddam Hussein must have been involved

This was not a fringe belief. Polls at the time showed that a large fraction of Americans believed there was a direct connection between Iraq and 9/11—despite repeated intelligence assessments to the contrary.

This, too, was semantic leakage: moral association masquerading as causal inference.

The same reasoning supported claims about weapons of mass destruction. Iraq was hostile. Hostile states possess dangerous weapons. Therefore, Iraq must have WMDs.

Pattern coherence replaced evidence. And the cost was measured in lives.

Prejudice Is Pattern Matching Gone Rogue

The same mechanism underlies racism, religious bigotry, and ideological stereotyping.

Humans constantly reason like this:

• Members of group X did something bad
• Group X shares visible traits
• Those traits become proxies for danger

We like to imagine that prejudice is a moral failing distinct from everyday cognition. It isn’t. It is everyday cognition, uncorrected.

A great deal of today’s political polarization is driven by the same cognitive error. Across ideologies, you see:

• “They oppose us, therefore they must be evil.”
• “They are associated with X, therefore they endorse all of X.”
• “This feels like past betrayal, therefore it is betrayal”

This is semantic leakage at scale: moral clustering replacing reasoning.

LLMs didn’t invent that.

What AI Really Exposes

Gary Marcus is right to worry about AI systems making confident but unjustified leaps. We should worry about that.

What he rarely acknowledges is the uncomfortable symmetry: humans do this constantly, including in domains where the stakes are far higher than a chatbot’s answer.

If semantic leakage disqualifies a system from being said to “understand,” then human understanding has always been conditional and fragile.

The Iraq War was not launched by a hallucinating language model. It was launched by pattern-matching primates who found a story that felt right.

Large language models don’t reveal a new form of irrationality. They reveal an old one, stripped of comforting myths about human exceptionalism.

They show us, at machine speed and scale, what happens when:

• Correlation masquerades as causation
• Moral valence substitutes for evidence
• Familiar patterns crowd out careful reasoning

In that sense, LLMs are not just tools. They are diagnostic instruments.

AI doesn’t invent our mistakes. It scales them. Large language models make visible the cognitive shortcuts we have always taken, and often still defend.

Correction Is the Point

The danger is not that minds—human or artificial—leak semantics. The danger is when there are no corrective mechanisms:

• No institutions that slow the inference
• No norms that demand evidence
• No humility about how often “it fits the pattern” is wrong

Humans have built such correctives before: science, journalism, adversarial debate, peer review. We weaken those systems at our peril.

We don’t need machines that never make pattern-based mistakes. We need systems that know how to notice them, name them, and stop doubling down when something feels right but isn’t true.

Semantic leakage is not an AI bug. It’s the shadow side of intelligence itself. And it’s been with us for a very long time.