We’ve Been Hallucinating for a While
But hallucination is neither bug nor feature
When I was in high school, a teacher asked how many states are in the United States. I said fifty. He said I was wrong. The answer was fifty-two. You must include Alaska and Hawaii. I had included them. I told him so. He pulled rank. He was the teacher. It was settled for him. He attempted to humiliate me. He only humiliated himself.
The room mostly did not agree with him. I know, I asked around afterward. But you could not tell in the class. Nobody else stepped to the plate to correct him after I did.
Forty years later I called my high school buddy to make sure I had not made up a false memory. He remembered it too. We forgot the class. We remembered the teacher’s name.
We did not hallucinate.
The word hallucination was coined by Thomas Browne in 1646. An English physician. He needed a word for a vision that was off. Not a lie. Not a dream. Something the mind produces while convinced it is perceiving the world. From the Latin alucinari, to wander in the mind.
Browne watched humans hallucinate in the 17th century. Three centuries later, John Tait watched a machine do it.
Tait described a program called FRUMP, short for Fast Reading Understanding and Memory Program. The name already promised what the program could not do. FRUMP read news wires and summarized them. One day, it processed a sentence about San Francisco being shaken by the death of Mayor Moscone. FRUMP filed the story under earthquake.
Tait called it “the hallucination of matches.” The match was stronger than the meaning.
That was 1982. Forty years before ChatGPT made the word a public concern. ChatGPT was released in November 2022. Within months, courts were sanctioning lawyers, professors were failing students, and editors were retracting articles. The same hallucinations entered medical records. ChatGPT had invented cases, quotes, and sources. Some humans had passed them on as facts. AI was giving us something we had not expected: smarter hallucinations. In 2023, Cambridge Dictionary named the verb, hallucinate, its word of the year. The new definition included AI hallucinations.
It sounds like a glitch. A misfire. Something the machine did while briefly disconnected from reality. The framing protects the system. The system is fine. ChatGPT just hallucinated. Sometimes humans confabulate too. The answer comes first. Then we make up the explanation.
What about the teacher in 1985? By Browne’s definition, he did not hallucinate. He believed it. He was not alone. The “fact” of fifty-two states was in circulation in the 1980s. Before the web, you went to the library for fact-checking or asked an authority figure, if you had access to one. The teacher was often the only authority in the room. The error survived. I had experienced a phenomenon that would not have a name for another twenty years.
Can you hallucinate if others experience the same thing?
The biological hallucination definition no longer holds. Hallucination now describes a system that is wrong, not corrected, and passed as knowledge. The burden of seeing clearly falls on the human, who read in good faith and assumed what they read was factual.
In 2009, a writer named Fiona Broome noticed that she had a clear memory of Nelson Mandela dying in prison in the 1980s. Other people online agreed. They remembered the funeral. They remembered the news coverage. They were certain.
Mandela was released from prison in 1990. He became president. He died in 2013, at 95, in his home.
Broome called it the Mandela Effect. The pattern kept showing up. Mr. Monopoly has never worn a monocle. Darth Vader never said “Luke, I am your father.” The line is “No, I am your father.” Some still think the United States has fifty-two states. People remember otherwise. Many of them with absolute confidence.
No machine was involved. No teacher pulled rank. Just a mistake that traveled. One person remembered wrong. Another agreed because they half-remembered the same thing. A third repeated it because the other two sounded sure. Somewhere along the way, the wrong version stopped being a mistake and started being memory.
The classroom has thirty kids. The social cost of disagreeing with the teacher is immediate. Most stay quiet. Some students could have walked out believing it.
The Mandela Effect operates on millions. The social cost is harder to see. You are not correcting a specific person. You are correcting a shared memory that has already hardened into fact. By the time the error is old enough to be named, too many people believe it for correction to travel cleanly.
We have more fact-checking tools now. Will it be easier?
We test Artificial General Intelligence (AGI) the way my teacher tested me: not to see if it can think, but to see if it provides the expected answer. Developers use benchmarks with technical names like MMLU, ARC-AGI, or HLE. These acronyms act as a shield. They discourage the question. The benchmarks do not track new ideas or complex problems. They check obedience.
Yann LeCun, Meta’s former AI chief, has admitted that his team used different models for different benchmarks. “Results were fudged a little bit,” he told the Financial Times. We are not testing cognition. We are checking compliance.
You rate an AGI based on the answer you expect to hear. Is that how you grade an essayist writing on AGI? A diplomat in a tense situation? A researcher with unexpected data?
AI nepotism: models feeding off models. When a hallucination enters the data, the next model learns it as fact. Error becomes norm. Some call it Habsburg AI, after the dynasty that went extinct through inbreeding. The model reproduces itself with itself until no outside signal can get in. Even before social media, rumours found a way to propagate. Urban legends too. Now, humans needn’t take part.
If ChatGPT hallucinates and nobody reads it, is it a hallucination? No. The hallucination needs a reader. And each ChatGPT session is its own environment. The same prompt can produce different answers between sessions, unlike traditional software.
Bugs used to have addresses. You found the cause. You wrote the fix. Large language models do not fail by bug. The error is not located. There is nothing to patch.
The industry has a response. Responsible AI. Human in the loop. Guardrails. Reinforcement learning from human feedback. A solution devised by the marketing department. The work looks like correction. It is closer to polish. A glass eye fills the socket. It does not see. You can keep it shining. It is still blind.
Meanwhile, the teacher still makes the questions and decides the answers. If he is wrong again, it is only a hallucination if he is caught... otherwise it becomes knowledge.