By a Lead Data Scientist
I was in a boardroom last week, listening to a VP of Operations explain why he wanted to fire their new automated customer support vendor.
“The AI lied to me,” he said, slamming his hand on the mahogany table. “I asked it if we had the Q3 inventory in stock, and it looked me in the eye—metaphorically—and told me we did. It lied. It’s dishonest.”
I sighed. I had heard this a thousand times.
“It didn’t lie to you, Dave,” I said. “It predicted a sequence of words that statistically correlated with your question based on its training data. It doesn’t know what ‘inventory’ is. It doesn’t know what ‘truth’ is. It’s a spicy autocomplete.”
Dave looked at me like I was the one hallucinating. “But it sounded so confident,” he whispered.
This interaction is the single biggest problem in Artificial Intelligence right now. Not compute power. Not energy consumption. Not copyright law.
The problem is that we treat math like it’s a person.
We use words like “hallucination,” “thinking,” “reasoning,” and “understanding.” We give these models names like Claude and Gemini.1 We say “please” and “thank you” in our prompts.
And because we project humanity onto these systems, we get burned. We trust them when we shouldn’t. We get angry at them when they fail. We attribute intent to what is fundamentally just a probability distribution.
It is time to stop anthropomorphizing the math. Calling a factual error a “hallucination” isn’t just a quirky metaphor—it is a dangerous misnomer that is costing businesses millions of dollars and confusing an entire generation of users.
Here is why we need to strip the humanity out of AI to actually use it well.
The Myth of the “Hallucination”
Let’s start with the word itself.
When a human hallucinates, it means their sensory processing has broken down. They are seeing something that isn’t there. But the implication is that normally, they see reality. A hallucination is a deviation from the truth.
When an LLM (Large Language Model) makes up a fact, we call it a “hallucination.”2
This frames the AI’s default state as “Truthful.” It implies that the AI knows the truth, but for some reason—maybe it’s tired, maybe it’s glitchy—it drifted away from it.
This is a fundamental misunderstanding of how LLMs work.
An LLM has no concept of “truth.” It has no concept of “reality.” It does not have a database of facts that it checks against.
It is a Next-Token Predictor.
Imagine a machine that has read every book in existence. If you type, ” The capital of France is…”, the machine calculates the probability of the next word.
- “Paris” (99.9% probability)
- “London” (0.001% probability)
- “Banana” (0.00001% probability)
It picks “Paris.” Not because it knows geography. Not because it looked at a map. But because in its training data, the word “Paris” follows “Capital of France” millions of times.
Now, imagine you ask it: “Who was the first CEO of Google’s Mars Division?”
There is no Google Mars Division. But the model doesn’t know that. It just sees the pattern: “Who was the [Title] of [Company]?”
It knows that usually, this pattern is followed by a name. So it calculates the most statistically probable name that fits the context of “Google” and “CEO.”
It might say: “Larry Page.”
It didn’t “hallucinate.” It didn’t “lie.” It successfully completed the pattern. It did exactly what it was built to do: generate a plausible-sounding sequence of words.
From the model’s perspective, “The capital of France is Paris” and “The CEO of Mars is Larry Page” are the same operation. They are both just high-probability token strings. One happens to align with physical reality; the other doesn’t. The model cannot tell the difference.
The Danger of the “Liar” Framing
When we call these errors “lies” or “hallucinations,” we attribute intent.
The lawyer who famously used ChatGPT to write a legal brief—and ended up submitting fake case law to a judge—fell into this trap. He asked ChatGPT to find cases. ChatGPT invented them.
The lawyer was shocked. “I asked it if the cases were real,” he told the judge, “and it said yes!”
He treated the AI like a junior associate. If a junior associate gives you a fake case, they are lazy or malicious. You can fire them. You can scold them.
But you can’t scold math.
Because the lawyer anthropomorphized the tool, he assumed a baseline of “truth-seeking.” He assumed that if the AI didn’t know, it would say, “I don’t know.”
But LLMs are not truth-seekers. They are plausibility-seekers. They are designed to sound human, not to be right.
In philosophy, Harry Frankfurt wrote a famous essay called “On Bullshit.” He argued that there is a difference between a Liar and a Bullshitter.3
- The Liar knows the truth and actively tries to hide it.
- The Bullshitter has no regard for the truth at all.4 They just want to sound convincing.
LLMs are the ultimate Bullshitters. They don’t care if the answer is right; they care if the answer looks right.
When we treat them as “Liars” (who occasionally “hallucinate”), we try to fix them with morality. We write prompts like: “You are a helpful and honest assistant. Do not lie.”
This is like yelling at a calculator to “be more careful.” It’s useless. The calculator doesn’t know what careful means. It just crunches numbers.
Re-Framing: From “Mind” to “Stochastic Engine”
So, if we stop thinking of them as messy humans, how should we think of them?
We need to treat them as Stochastic Engines. (Stochastic just means “randomly determined”).5
When I work with my engineering team, I ban the word “hallucination.” We use the term “Confabulation” or “Prediction Error.”
This shift in vocabulary changes how you build software.
1. You Stop Trusting the Output
If you think you are talking to a “Smart Assistant,” you trust it.
If you think you are using a “Probabilistic Text Generator,” you verify everything.
You build Grounding Systems. You don’t just ask the LLM for the answer. You ask the LLM to write a search query, run the query against a trusted database (RAG), and then ask the LLM to summarize only the data from that search.
2. You Stop “Reasoning” with It
I see so many prompt engineers trying to argue with the model. “No, you got that wrong, please try again and think harder.”
Sometimes this works (because it changes the context window), but often it just leads to the model “apologizing” and then making up a new lie to please you.
Instead, you fix the input. You realize the model isn’t “confused”; it just lacks the statistical signal to generate the right token. You feed it better examples. You adjust the temperature (randomness). You treat it like a machine that needs calibration, not a person who needs a pep talk.
3. You Embrace the Creativity
Once you accept that the model is just guessing the next word, you realize that “hallucination” is actually a feature, not a bug.
That same mechanism that makes it invent fake legal cases is the mechanism that allows it to write a poem about a cyberpunk toaster. It is connecting unrelated concepts based on probability.
If you want accuracy, you clamp down on the randomness. If you want ideas, you turn it up. But you don’t expect the “Idea Machine” to be the “Fact Machine.”
The “Vibe” Trap
The hardest part of killing the anthropomorphism is that the models are getting better at faking it.
Claude Opus 4.5 feels incredibly human. It uses nuance. It expresses hesitation. It says things like, “I feel like that approach might be risky.”
It does not feel. It does not have risk assessment. It has merely ingested millions of conversations where humans expressed hesitation in similar contexts, and it is mimicking that pattern.
This is the “Vibe Trap.” The better the UI/UX of the personality becomes, the harder it is to remember it’s just math.
We are entering an era of “Agentic AI,” where we give these models access to our emails, our calendars, and our bank accounts.6 We are saying, “Go act on my behalf.”
If we continue to view them as “Digital Employees,” we are going to have disasters. We will have agents that negotiate contracts they don’t understand, or delete databases because the probability distribution suggested that rm -rf was the next logical command.
Conclusion: Respect the Alien
I am not a luddite. I use LLMs every single hour of every single day. I love them.
But I love them the way I love my microwave.
My microwave is amazing. It excites water molecules to heat my burrito in 60 seconds. It is a miracle of engineering.
But I don’t ask my microwave for life advice. I don’t get mad at my microwave if it makes the burrito too hot. And I certainly don’t attribute “intent” to the magnetron.
We need to treat LLMs with the same cold, clinical respect.
They are Alien Intelligence. They process information in a way that is fundamentally different from biology. They are not messy, flawed humans who sometimes “hallucinate.” They are precise, mathematical engines that sometimes predict the wrong token.
The moment you stop trying to make friends with the chatbot is the moment you start actually getting value out of it.
Stop calling it a hallucination. Call it a wrong guess. And then check the math.