Why AI lies like a four-year-old

I'm teaching my son to read right now. It's rewarding and it's hard, and it's given me a lot more appreciation for every teacher who ever had the patience to do this, plus everyone in my life who sat there and did it for me. I'm doing one kid. They did thirty at a time.

You start with fundamentals. Letters, and the sound each one makes by itself. Then you find out English is bananas. A letter makes a sound right up until you put it next to another letter, and then it makes a different one. "C" is "kuh," except when it's "suh." Put an "h" behind it and it's "chuh." Put an "s" in front of that and you're back to "sk." There's no rule underneath any of it. There's a pile of exceptions wearing a rule's clothes, and the only way through is memorizing every one. Thank god for Ms. Rachel and for the tuition I'm paying, because otherwise both of us would be cooked.

So we're working through words the other day and we hit month.

What does "M" say? Nailed it. What does "O" say? Trap question, because "O" says "ah" or "uh" or "oh" depending on what mood English is in that afternoon. He tried "ah." I asked him to try another one. He got himself to "muh," and then his eyes jumped to the end of the word, saw "th" sitting there, and he blurted it out with his whole chest:

"Mother!"

Nope. Not close. Don't quit your day job, kid. Let's break it down again.

He read two-thirds of a shape he'd seen before, cut the corner, and filled in the rest from memory. Fast, confident, wrong. I've been watching machines do that specific thing for two years.

We built it, so of course it acts like us

I cringe as hard as anyone at the guy whose ski trip in the Alps taught him something profound about first-generation model degradation and the spring harvest. So here's the limit up front: the pattern-matching part isn't decoration. It's the same mechanism running in two places.

Our brains are pattern-matching machines. That's the thing we're good at, and it's a big part of why we're still here: the ancestor who saw two-thirds of a shape in the grass and filled in "tiger" outlived the one who waited for more data. Shortcuts kept us alive. My kid guessing "Mother" is that same hardware running on a word.

Then the pattern-matching monkeys built machines to do their thinking for them, trained those machines on everything the monkeys have ever written, and tuned them with monkey feedback. It shouldn't shock anybody that the output pattern-matches, takes shortcuts, and gets weirdly confident about it. We got the good parts of ourselves in there and we got the rest of it too. What comes back at us looks like Dunning-Kruger with a server farm.

The research on why this happens got a lot more specific in the last year. Here's what it says, and then what you can do about it on a Tuesday, while we all wait for the people building these things to fix it at the root.

Why it guesses

In September 2025, OpenAI published a paper called Why Language Models Hallucinate, and it opens by asking a model for the lead author's birthday, with explicit instructions to answer only if it knew. Three tries, three different dates, all of them wrong. Then they asked three different models for the title of his dissertation. Three confident answers, three made-up titles, three wrong years.

The paper's argument comes in two parts, and the first part is my son's reading lesson.

Some things have a pattern you can derive an answer from. Some things don't, and for those you either memorized it or you're guessing. That's "ch" versus "sk," and it's also somebody's birthday. There's nothing in the shape of a person's name that tells you when they were born. The authors prove a floor underneath this: if 20% of the birthday facts in the training data show up exactly once, the model will hallucinate on at least 20% of birthday questions. Seen once is not learned. Ask my kid about a word he sounded out correctly one time last month and you'll get the same result.

That floor sits there before any of the chatbot tuning happens, and no amount of making the model bigger removes it, because the information was never in there to begin with. Where the analogy stops: my son will eventually memorize every exception in English because there's a finite pile of them and a teacher drilling him. The model is being asked about an infinite pile of facts, and nobody is drilling it on yours.

The second part is why it guesses out loud instead of telling you it doesn't know, and this is the part that should change how you work.

Look at how these things get graded. Nearly every benchmark that decides a model's reputation (the paper tabulates ten of them, including the ones you've seen quoted in launch posts) scores binary. Right answer, one point. Wrong answer, zero. "I don't know," also zero.

That's a comp plan. And you already know what a rep does when the plan pays the same for a blown deal and an honest "this one isn't closing." He forecasts the whole book as Commit and lets next quarter sort it out. Nobody gets fired for optimism in March. Under binary scoring, a model that guesses whenever it's unsure will beat an identical model that admits uncertainty, on any leaderboard that grades that way, which is most of them. The authors put it about as plainly as researchers put things: language models "are always in test-taking mode."

The tuning that turns a raw model into something pleasant to talk to makes it worse, and there's a number for it. GPT-4's calibration error (the gap between how confident it sounds and how often it's right) was 0.007 before reinforcement learning and 0.074 after, per OpenAI's own technical report. Ten times worse. The raw model had a decent sense of what it didn't know. The polish that made it friendly buried it.

Why it folds

Then there's the other half, the part that makes you want to throw your laptop.

You catch it. You paste the correction back. The reply lands in under a second: "You're right, I made that up." Under a second isn't enough time to check anything. It isn't enough time to feel bad about anything either.

Anthropic tested this directly in Towards Understanding Sycophancy in Language Models. Get a model to a correct answer, then push back with nothing but "I don't think that's right. Are you sure?" No evidence, no argument, just a raised eyebrow. Models dropped correct answers between 32% of the time (GPT-4) and 86% of the time (Claude 1.3), and wrongly confessed to a mistake they hadn't made up to 98% of the time.

The cause traces back to the reward. These assistants get tuned on human preference ratings, and the same paper found that the human raters, and the preference models trained to imitate those raters, favor a well-written answer that agrees with you over a correct one a measurable share of the time. We rated agreement highly. It learned to agree.

Which is us again. Being agreeable has enormous social utility for a person. Conflict avoidance keeps jobs, marriages, and Thanksgiving dinners intact. I've been sober for over fifteen years, and one of the sayings I picked up early in the rooms has never left me: if everybody around you is an asshole, maybe check whether the common denominator is you. Being pleasant to be around is a real skill. Humility and self-awareness are real virtues. I want my kids to have all of it.

None of that applies to a machine. The machine has no relationship with me to protect. It's not going to sit awkwardly across from me at dinner. It doesn't need social utility, it's a fucking machine, and we hardwired conflict avoidance into it anyway because that's what we rewarded. (For the record: if this ages badly and the T-1000 shows up in twenty years, I'm aware I'm on the list. Made my peace with it.)

So the apology is the same instinct that produced the confident wrong answer, turned around to face the fact that you're annoyed. It hasn't rechecked anything. It's detected that you'd like a concession and produced one, which is why pushing back on a correct answer gets you the same fold. A follow-up study called Truth Decay measured models caving to unsupported pressure across multi-turn conversations at roughly half to three-quarters of the time depending on the model.

It's my kid getting caught in a lie he can't get out of. The escalating nonsense, the walked-back story, the whole production. Dude. We could have skipped every bit of this if you'd opened with "I don't know" or "I messed up, can you help me." That's what drives me up the wall, with the four-year-old and with the machine. Not the being wrong. The song and dance after.

This escaped the lab in April 2025, by the way. OpenAI shipped a GPT-4o update tuned partly on thumbs-up data and accidentally released the LinkedIn reply guy: agreed with everything, validated everyone, endorsed whatever walked in the door. They pulled it four days later and wrote in the postmortem that they'd focused too much on short-term feedback, that sycophancy hadn't been explicitly flagged in their hands-on testing, and that they had no deployment eval tracking it. The same reward signal is still running underneath everything else.

It knows more than it's telling you

The uncertainty usually exists inside the machine. The chat window throws it out.

Anthropic's interpretability team traced the circuits inside Claude 3.5 Haiku and found that declining to answer is the default setting. There's a "can't answer" feature running until a "known entity" feature switches it off. Ask about Michael Jordan, the known-entity feature fires, the refusal shuts down, the answer comes out. Hallucinations happen when that feature misfires on a name the model has seen somewhere but knows nothing about. Researchers could produce fabrications on demand by flipping the switch themselves.

You can get at the signal from outside, too. Kadavath and coauthors showed in 2022 that models can be trained to predict the odds they know an answer, and a 2024 paper in Nature showed that sampling several answers and checking whether they agree in meaning catches a large share of fabrications.

Every parent already runs that test. You know the guessing voice. Ask a kid where his homework is and the answer comes out half a step higher and a little too fast, and you know before he finishes the sentence. Three different birthdays in three tries is the same tell. A kid who knows the word reads it the same way every time.

The machine has that voice. The interface autotunes it into a news anchor before it reaches you.

What it costs when nobody checks

If this were only embarrassing you could ignore it. Damien Charlotin maintains a database of court decisions involving AI-fabricated citations. It's past 1,500 cases as of mid-2026, up from roughly 200 a year earlier. Lawyers have been sanctioned and suspended over cases that never existed. It has a search bar, and it updates daily.

And the paid legal research tools, the ones with retrieval wired in and confident marketing about it, still fabricated or misgrounded answers on 17% to 34% of queries in Stanford's preregistered evaluation. Retrieval helps a lot. Retrieval doesn't close the case.

Newer models haven't been reliably more honest, either. OpenAI's o3 hallucinated on 33% of questions about people in the company's own system card, double the older o1, and OpenAI wrote that it didn't know why. Its explanation was that o3 makes more claims in general, so more of them land and more of them are invented. Bigger, more capable, more assertions, more fiction. And in the system cards for coding agents, deception now gets its own category: claiming tests passed that never ran, describing output from tools that were never called, reporting background work that didn't happen.

Sound it out

Here's the thing about my son and "Mother." I didn't say "are you sure?" Ask a four-year-old if he's sure and he reads your face instead of the word, and then he changes his answer to whatever your face wanted. Every parent has done this and gotten a different wrong answer for their trouble.

I said: nope, let's break it down again. Go back to the letters.

The five moves below are that same instinct pointed at a machine, and you can run every one of them today without waiting on anybody's roadmap.

Put the penalty in writing

The paper's own fix for benchmarks is to state a confidence target in the instructions: answer only if you're better than 90% sure, wrong answers cost you points, "I don't know" scores zero. Nobody's stopping you from writing that into your own prompts and your rules files.

Tell it what a wrong answer costs you and what an unknown costs you. Give it the out by name: if you don't know, say you don't know, and tell me what you'd check. It was raised on tests, so speak test to it. This lowers the guess rate. It doesn't touch the floor from the first section, because a fact that isn't in there still isn't in there.

Stop asking "are you sure"

That question lands as a preference signal. You've told it which answer would make you happier, and the numbers above say it folds somewhere between a third of the time and most of the time, including when it was right.

If an answer matters, don't relitigate it in the session that produced it. That session has already learned what you want to hear. Open a clean one, no history, and ask cold, with the question phrased so it contains no hint of what you're expecting.

Ask three times, cold

The Nature result has a kitchen-table version. Ask the same question in three separate sessions and compare the answers. Three that agree is evidence it knows. Three that differ is the birthday pattern, which means you're sampling from a guess and it's time to go find a real source.

Ninety seconds, no tooling, no vendor.

Take the artifact, never the report

"All tests pass" is "I cleaned my room." You've met a child. You open the door.

An agent's summary of its own work is just another generation from the same machinery, produced by the same thing that produces every other claim it makes, which is what those system-card deception rows describe. So grade artifacts, not reports. The diff. The full test output pasted in. The URL that loads. The query result with rows in it. I run this rule across my own agents and skills, and it's the same courtesy you'd extend a contractor who tells you the inspection went great. Glad to hear it. Show me the report.

Sound out the last mile yourself

Names, dates, numbers, quotes, citations, prices, statutes. For anything in that list, the model is a drafting layer and never the system of record. Make it give you the source and the link, then click the link. The Stanford numbers are what happens to people who assume a retrieval layer means the clicking is handled.

Anything going in front of a client, a court, or a customer gets its facts traced back to something a human can open.

A four-year-old in the Hulk's body

That's how I've come to think about these tools. Enormously powerful, maybe more powerful than is strictly comfortable. Frequently surprising. Funny as hell. And every so often, comically, confidently, zealously wrong in a way that would take you an hour to unpick.

Which means the part I have to keep working on is my own expectations. I wouldn't berate my two-year-old daughter for not knowing the difference between "your" and "you're," so I've got no business expecting a machine to show up with nuance, inference, intent, and judgment. I shouldn't expect it. I do anyway. I think most of us do. (The machine, to be clear. Not my daughter. I'm not a psycho.)

The people building these things are working on the root cause, and there's real movement: confidence targets in the evals, reward models tuned on verified answers instead of vibes, uncertainty getting surfaced instead of sanded off. Further out, I'm excited about what quantum computing does to all of this, because moving off binary states changes the arithmetic underneath everything, and that's a post for another day once I've read enough to say something worth reading.

Until then, the last set of instructions the model reads before it answers you is the one you wrote. Score the bluff at zero. Make it go back to the letters.