What Is an AI Hallucination? A Sixty-Year History of a Known Problem.

An AI hallucination is output that a language model presents as fact when it is false, fabricated, or unsupported by any source. The model is not lying, and it does not know it is wrong. It predicts the most plausible next words, and plausible is not the same as true. That definition is easy to find. The record of how long we have known it is not.

1966: the human half of the problem is documented.

Joseph Weizenbaum built ELIZA at MIT, a program that imitated a psychotherapist through pattern matching and understood nothing. People believed it anyway: users disclosed intimate problems, and Weizenbaum's own secretary, who had watched him build it for months, asked him to leave the room so she could talk to it privately. He later wrote that he had not realized short exposure to a relatively simple computer program could induce powerful delusional thinking in quite normal people. What the ELIZA effect documented six decades ago was not machine hallucination as we define it now. ELIZA generated almost nothing. What it documented was the other half of the problem, and the half that has never changed: people trust fluent output, and will keep trusting it after being told there is nothing underneath.

Before 2018: the word already existed.

"Hallucination" was in use in machine learning well before language models, most visibly in computer vision through Simon Baker and Takeo Kanade's Hallucinating Faces, circulated as a 1999 technical report and presented at an IEEE conference in 2000, which described filling in missing image detail from patterns learned in training. The term was coined and recoined across decades, tracing back further still to 1980s work on text parsing and fingerprint pattern generation, and the meaning tracked a common thread: the system producing content that was not warranted by the actual input.

2018: the term arrives in text generation.

Neural machine translation made output dramatically more fluent, which made errors dramatically harder to spot, since a statistical model's mistakes came out garbled while a neural model's came out grammatically viable and completely wrong. Katherine Lee, Orhan Firat and colleagues at Google published Hallucinations in Neural Machine Translation at a NeurIPS 2018 workshop, applying the existing term to translation output untethered from the input, warning it could lead users to a false sense of confidence, and finding hallucinations widespread in the system they studied. Four years before ChatGPT reached the public, researchers at one of the largest labs in the world had published on this behavior in text.

2021: the formal warning, and the firing.

Bender, Gebru, McMillan-Major and Mitchell published On the Dangers of Stochastic Parrots at ACM FAccT, arguing these systems produce statistically plausible sequences of language with no understanding of meaning, reference, or truth. Google fired two of the authors. Bender has since noted that when they wrote it in late 2020, the section on synthetic text felt like thin ice, because it was hard to imagine anyone would want machine-generated text at all.

November 2022: ChatGPT reaches the public.

OpenAI released ChatGPT to the general public with everything above already in the literature: the ELIZA effect, a documented failure mode in text generation published by Google's own researchers, and a formal warning about deploying exactly this. It went out with a disclaimer that it may produce inaccurate information.

January 2024: the first systematic measurement.

Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel Ho of Yale and Stanford published Large Legal Fictions in the Journal of Legal Analysis, running more than 800,000 queries against public models. Asked direct, verifiable questions about randomly selected federal court cases, the models hallucinated between 58 percent of the time with GPT-4 and 88 percent with Llama 2, and the paper also found they struggle to predict their own hallucinations and often uncritically accept a user's incorrect legal assumptions.

2024: an argument that it is permanent.

Ziwei Xu, Sanjay Jain and Mohan Kankanhalli published Hallucination is Inevitable: An Innate Limitation of Large Language Models, arguing that hallucination is a property of this class of system rather than a bug awaiting a patch, because these models predict text rather than verify it. The argument is not universally settled. Nobody has demonstrated the opposite either.

May 2024 to 2025: the fix does not fix it.

The industry's answer was retrieval augmented generation, and vendors marketed it as solved, with LexisNexis advertising "100% hallucination-free linked legal citations" and Thomson Reuters claiming its tools avoid hallucinations by relying on trusted content. In the first preregistered evaluation of those products, Stanford RegLab measured roughly 17 percent hallucination for Lexis+ AI and 33 percent for Westlaw AI-Assisted Research on the legal research queries tested, against 43 percent for an ungrounded GPT-4, concluding that providers' claims are overstated. Retrieval helped, and reached zero in nothing, because the model still writes the final answer live.

2023 to 2026: the consequences at scale.

Attorneys for Roberto Mata used ChatGPT to research a brief that cited six federal decisions, none of which existed, and when questioned, Steven Schwartz asked ChatGPT whether the cases were real and it said yes; he later testified he had been operating under the false perception that the website could not possibly be fabricating cases on its own, and on June 22, 2023, Judge P. Kevin Castel sanctioned the attorneys and their firm $5,000. That was the opening entry in a list that Damien Charlotin's public database of AI-hallucination decisions carried past 1,780 cases worldwide by July 2026, with at least $145,000 in U.S. sanctions in the first quarter of 2026 alone and the first indefinite bar suspension tied to AI filings handed down in Nebraska. In healthcare, the patient safety organization ECRI ranked the misuse of AI chatbots as the number one health technology hazard of 2026, above cybersecurity risks, after testing found chatbots suggesting incorrect diagnoses and inventing body parts.

The three types you will actually encounter.

Fabricated references, where the model invents a case, paper, statute, or URL that looks entirely real and is the easiest type to disprove. Factual inaccuracies, where dates, statistics, prices, deadlines, or medical guidance are confidently wrong and harder to catch because there is nothing to look up. And misattribution, where a real source is cited under a sentence it does not support, which the Stanford RegLab team flagged as harder to catch than an invented case, because the source is real even when the sentence is wrong.

Why the history matters.

The standard framing treats hallucination as a bug being worked on, an unfortunate side effect of a young technology that the next model generation will engineer away, and the record complicates that framing: the human tendency to over-trust fluent machine output was documented in 1966, the term was in machine learning use by the 1990s, the behavior was published on in text generation by Google researchers in 2018, formally warned about in 2021, released to the public in 2022, systematically measured in early 2024, and argued to be innate later that same year. These systems reached law offices, medical practices, and financial firms with the failure mode already described in the literature, because fluent confident text is what they are built to produce.

Sixty years of evidence says you cannot count on a language model to stop making things up. You can only decide whether it is allowed to speak for you unsupervised.