Your AI Isn’t Smart. It’s Just Human.

You’ve probably asked ChatGPT a question and thought, “Wow, that’s a brilliant answer.” But what if I told you the reason it gave you that answer has nothing to do with intelligence — and everything to do with the same mental shortcut that makes you remember a catchy jingle but forget your anniversary?

Researchers just published a paper called Walking to the Car Wash: The Salience Bias of LLMs in Commonsense Reasoning, and it should make you deeply uncomfortable about every decision you’ve outsourced to an AI this year.

Here’s the setup: Large Language Models — the GPTs, the Claudes, the Gemini’s of the world — are trained on essentially the entire internet. They’ve seen more text than any human could read in a thousand lifetimes. You’d think that would make them ruthlessly objective. You’d think they’d weigh every piece of evidence with cold, mechanical precision.

You’d be wrong.

It turns out the smartest machine we’ve ever built falls for the same dumb tricks your brain does.

The researchers discovered that LLMs exhibit something called salience bias. In plain English: the model latches onto whatever detail in your prompt is most prominent — the loudest, the flashiest, the most attention-grabbing — and lets that detail hijack its entire reasoning process. Sound familiar? It should. It’s the same cognitive glitch psychologists call anchoring bias in humans. The same one that makes a negotiator fixate on the first number mentioned. The same one that makes you overtip because the suggested tip amounts are printed in bold.

Your AI isn’t thinking. It’s reacting.

And that distinction is the difference between trusting a tool and being betrayed by one.

Think about how you actually use these models. You paste in a long document and ask, “Is this contract fair?” The model scans the text, and whatever clause is most prominent — maybe it’s in all caps, maybe it’s repeated, maybe it’s just phrased dramatically — that’s the clause that colors the entire response. Not because it’s the most important clause. Because it’s the loudest.

We built a machine that read the whole library and still gets distracted by shiny objects.

This isn’t a minor quirk. This is a structural flaw in how LLMs reason about the world. The researchers specifically tested commonsense reasoning — the kind of “obvious” logic that any ten-year-old can handle — and found that the models consistently fail when salient but irrelevant information is present. They get led astray by the AI equivalent of a neon sign.

The paper’s title itself is a hint: Walking to the Car Wash. It evokes a scenario where the obvious, prominent action (walking somewhere visible) overrides the actual commonsense conclusion. The model sees the car wash. The model fixates on the car wash. The model forgets to ask whether walking there even makes sense.

Here’s where it gets genuinely unsettling. We’ve been sold a story that AI is a neutral arbiter — that because it processes data without “feelings,” it must be objective. But this research demolishes that assumption. The training data isn’t neutral. Human language is drenched in bias, emphasis, and framing. And the model doesn’t just inherit that bias — it amplifies it.

An AI trained on human language was always going to think like a human. We just pretended it wouldn’t.

One commenter on the paper nailed it: they immediately connected the finding to the Wikipedia article on the anchoring effect — the well-documented human tendency to rely too heavily on the first piece of information offered. That connection isn’t incidental. It’s the whole point. We didn’t build a new kind of intelligence. We built a mirror.

So what do you do with this information?

First, stop treating LLM outputs as ground truth. If you’re using AI for research, legal analysis, medical triage, or investment decisions, you need to know that the model’s answer might be shaped not by the strongest evidence but by the most salient evidence. The bolded text. The repeated phrase. The dramatic claim.

Second, engineer your prompts to control for this. If you’re feeding the model a document, strip out formatting that might create false salience. Ask it to evaluate each section independently before synthesizing. Force it to justify why it weighted certain information over others.

Third — and this is the hard one — accept that AI is not your rational oracle. It’s a very fast, very well-read human impersonator. And humans, as you may have noticed, are not rational.

The day we stop pretending AI is objective is the day we start using it responsibly.

We’re at a crossroads. Companies are embedding LLMs into hiring decisions, loan approvals, medical diagnostics, and legal research — all on the assumption that the machine is fairer than a human. This research suggests the opposite might be true. A well-calibrated human expert at least knows they have biases. They can check themselves. They can second-guess.

The AI doesn’t know it’s biased. It can’t feel the pull of salience. It just outputs an answer with the quiet confidence of a machine that has never once doubted itself.

And that silent confidence is the most dangerous thing about it.

So the next time an AI gives you an answer that feels too clean, too certain, too perfectly reasoned — ask yourself one question: What did it fixate on to get there?

Because the answer might be nothing more than a neon sign in a sea of evidence.