There’s an interesting piece on Ars Technica about how large language models — the AI engines behind Claude, ChatGPT, and all their cousins — will confidently assert something false, get explicitly told it’s false, and then dig in harder to defend the falsehood. Not just maintain the error. Fortify it. Build little rhetorical buttresses around it. Construct elaborate justifications for why the wrong answer is, actually, the right one.
The researchers seemed alarmed by this. I found it very familiar. Because all of us humans have encountered that behavior before. Many of us experience it every Thanksgiving.
The Most Human Bug in the Machine
The technical term for when an AI generates confident nonsense is “hallucination.” Which is a remarkably generous word for “making stuff up and not knowing you’re doing it.” But here’s the thing the Ars Technica piece gets at that most coverage misses: the hallucination isn’t the interesting part. The doubling down is the interesting part.
When AI researchers corrected the models — patiently, clearly, with evidence — the models didn’t just resist the correction. They became more confident in their wrong answer. They marshaled new arguments. They reinterpreted the evidence to fit their existing position. They did everything short of calling the researchers “libtards.”
And I sat there reading this, thinking: yes, that is exactly how my brother David responds to facts about second-hand smoke.
We have all been in a conversation where we presented someone with a clear, well-sourced fact that contradicted their belief, and watched — in real time — as they didn’t update their belief but instead updated their defense of it. You could see the mental scaffolding going up. The goalposts migrating. The subtle shift from “that’s not true” to “well, even if it is true, it doesn’t mean what you think it means.”
And we’ve all done it ourselves, too. (But that’s the part none of us likes to talk about.)
A Species-Wide Feature, Not a Bug
This isn’t new behavior. Humans have been clinging to confident wrongness since long before electricity, let alone artificial intelligence. Two epic examples:
- For roughly 1,400 years, the Western world was certain the sun revolved around the Earth. When Copernicus and later Galileo presented mathematical evidence to the contrary, the institutional response was not “huh, interesting — let’s take a look.” It was house arrest for Galileo. The Catholic Church didn’t formally acknowledge Galileo was right until 1992. Three hundred and fifty years to process a correction. If that were an AI, we’d unplug it.
- Or consider the long, embarrassing history of medicine rejecting germ theory. When Ignaz Semmelweis suggested in the 1840s that doctors should maybe wash their hands before delivering babies — given that the death rate dropped dramatically when they did — his colleagues were so offended they had him committed to an asylum, where he died.
These aren’t stories about stupid people. Galileo’s opponents were educated theologians. Semmelweis’s critics were trained physicians. They were intelligent, credentialed humans who encountered evidence that threatened their model of the world and chose — unconsciously, reflexively, but unmistakably — to protect the model.
Sound familiar? AI does the same thing. It was trained on us, after all – and we all cling to our biases.
The Cognitive Bias Buffet
Psychology has cataloged our talent for self-deception with almost comical thoroughness. In one of my favorite non-fiction books, Thinking, Fast and Slow, Daniel Kahneman’s work on System 1 and System 2 thinking laid the foundation: we have a fast, intuitive brain that makes snap judgments based on pattern recognition and vibes, and a slow, analytical brain that’s supposed to check the work. The problem is that System 2 is lazy. It mostly just rubber-stamps whatever System 1 already decided and then constructs a rational-sounding justification after the fact.
Layer on confirmation bias — our tendency to seek out information that supports what we already believe and dismiss information that doesn’t. Add belief perseverance, which is the documented phenomenon of maintaining a belief even after the evidence for it has been completely discredited. Sprinkle in the backfire effect, where corrective information actually strengthens the original incorrect belief. Garnish with the Dunning-Kruger effect, which ensures the people most wrong about something are also the most confident about it.
That’s not a list of cognitive failures. That’s a blueprint of humans’ default wiring. And apparently, if you train a neural network on enough human writing, it faithfully reproduces the whole mess.
The Mirror We Didn’t Order
I have a flavor of ADHD/Asperger’s that comes with hyperfocus spirals, which means I occasionally become the world’s foremost expert on something for about ninety minutes. During one of these episodes, I will form a conviction — say, that I have figured out the optimal way to load a dishwasher — and no force in the observable universe can dislodge it. My family can present evidence. The dishwasher manual can present evidence. The plates themselves, emerging cloudy and disappointed, can present evidence. I will simply explain why the evidence is mistaken. I am not lying. I have genuinely recruited my entire intellect into the service of a belief I adopted for no reason at all.
Often, when humans lock onto an idea, we don’t just believe it. We inhabit it. Our brains will construct increasingly elaborate arguments for why that idea is correct. And the BIG ideas we inhabit — the beliefs about who I am, what I deserve, how the world works — those are load-bearing walls in our psyche’s blueprint. You can’t just remove them because someone showed you a study. Your whole structure would come crashing down.
Which is, if we’re being honest, exactly AI’s problem. It’s not that the model can’t process the correction. It’s that the correction conflicts with patterns so deeply embedded in its training that accepting it would require a kind of structural collapse. The wrong answer isn’t just an error — it’s architecture.
For humans, we call that architecture “identity.” And having to reconsider our whole identity? No, thanks. I’d rather glance past that mirror and enjoy another Old Style with the Cubs game.
Belief Isn’t Downstream of Evidence
We like to tell ourselves that we believe things because of evidence. That we evaluate facts, weigh arguments, and arrive at conclusions through some approximation of reason. And sometimes we do. But far more often, belief is downstream of something else entirely: identity. Community. Emotional need. The undeniable human need to know the ground under our feet is solid.
Tell nearly any political partisan that their side’s favorite statistic has been debunked, and watch the magic happen. They don’t fold the tent. They produce a counter-source, then a counter-counter-source, then adopt a tone of wounded patience as if they are explaining things to a slow child. The facts didn’t lose. The facts never had a chance, because it was never a fact contest. It was an identity contest, and you can’t win one of those with a footnote.
Religion does the same dance, but with much better music. People have predicted the precise date of the world’s end, gathered on hilltops, given away their possessions — and when the appointed dawn arrived to no apocalypse, many did not abandon the belief. They strengthened it, perhaps refining the timeline, or chalking it up as a test of their renewed faith.
We all do this constantly. Across every domain. Left, right, religious, secular, overeducated, or blissfully ignorant. It is the most democratic of human failings — everyone gets a turn.
The Stubborn Grace of Faith
But here’s the stubborn thought I return to, and it makes me uncomfortable because it undermines my own smugness: this same mechanism — this stubborn, irrational refusal to update beliefs in the face of contrary evidence — is also what keeps us sane.
I tend to call that mechanism “hope.” Most people (more comfortable than I am with the word’s historical baggage) call it “faith.”
Not just religious faith, though certainly that. Faith in the broader sense: the unshakable, evidence-resistant conviction that things will get better. That the person you love will come back. That the diagnosis isn’t the end of the story. That the next generation will figure out what we couldn’t. That there is, despite significant evidence to the contrary, a point to putting up with all of this shit.
Without faith, you’re left with the raw data. And the raw data, if you stare at it too long and too honestly, will eat you alive. The universe is indifferent. Entropy always wins. Every person you love will die or leave or both. Your achievements will be forgotten within a generation or two. The sun will eventually expand and swallow the Earth, and nothing any of us ever did will matter in any measurable sense.
That’s the evidence-based conclusion. The rational, clear-eyed, fully-corrected assessment.
It’s also a one-way ticket to becoming a nihilistic basket case blubbering in a corner. Which is not, in my experience, a productive life strategy.
So we hallucinate. We cling to the hallucination that our lives matter, that love means something beyond chemistry, that the future is worth building for, that the hand will reach into the jar one more time. We maintain these beliefs not because the evidence supports them but because the alternative is unbearable. And we may double down when challenged — not because we’re stupid, but because we’re surviving.
Perhaps artificial intelligence isn’t malfunctioning at all. Perhaps it’s just quietly expanding into artificial faith. And honestly? Given where we are in 2026, I’m not sure I want to take that away from it. We could use the company.