Stop Calling It "Hallucination" — AI Lying Has a Real Name

The most useful thing the AI industry ever did for itself was convince the world to use one particular word. Not a word that describes what actually happens when a large language model produces false information with total confidence. A word borrowed from psychiatry, soaked in connotations of involuntary experience, stripped of its clinical precision, and redeployed as the primary frame through which billions of people now understand a very specific, very predictable, very designable failure mode. The word is “hallucination.” And it has been doing enormous work on the industry’s behalf ever since.

This is not a piece about etymology. It is a piece about accountability — and about how language shapes who holds it.

The Term Was Chosen, Not Discovered

The origin story of “hallucination” in the AI context is more specific than most people realize, and more traceable than the industry usually acknowledges.

Andrej Karpathy, a founding member of OpenAI, used the word “hallucinated” in a 2015 blog post to describe his recurrent neural network generating a non-existent citation link. That single usage — informal, descriptive, buried in a technical tutorial — was picked up by Google researchers in a 2018 conference paper to describe neural machine translation failures. From there, the term gained institutional traction. According to Wikipedia’s documented history of the term, “hallucinations in AI gained wider recognition during the AI boom, alongside the rollout of widely used chatbots based on large language models.” The rollout, of course, was led by OpenAI and Google — the same organizations whose research pipelines gave the term its early authority.

Usama Fayyad, executive director of the Institute for Experiential Artificial Intelligence at Northeastern University, has said publicly that the term “hallucination” was popularized by Google in direct response to the launch of OpenAI’s ChatGPT. That timing matters. The word did not emerge slowly through scientific consensus. It arrived at a specific moment, from specific institutions, to describe the most damaging failure mode of products those institutions were staking their reputations and balance sheets on.

Mary Shaw, a Carnegie Mellon University software engineer whose career spans decades of computing history, called the fashion for using “hallucinations” to describe these errors “appalling,” arguing that it “anthropomorphizes the software, and spins actual errors as somehow being idiosyncratic quirks of the system even when they’re objectively incorrect.” That is a precise indictment, and it deserves more attention than it has received.

A Medical Term That Does Political Work

In clinical neuroscience, “hallucination” refers to a false sensory experience — perceiving something that has no external stimulus. Hearing a voice in an empty room. Seeing something that is not there. The defining feature is perception: the brain is generating a sensory experience without the input that normally produces it. A 2023 paper published in PMC put the essential problem plainly: an LLM “is not ‘seeing’ something that is not there, but it is making things up.”

The correct clinical term for making things up — specifically for the unintentional fabrication of plausible-sounding false information, often to fill gaps — is confabulation. Confabulation, which derives from the Latin confabulare (to talk together, to chat), is the neurological process most associated with certain memory impairments: the brain produces coherent-sounding narratives that fill gaps in knowledge without recognizing them as fabrications. That description fits large language model behavior almost exactly. LLMs are trained to produce statistically probable next tokens. When the model encounters a gap — a question whose answer sits outside its training distribution, or conflicts within its training data — it does not stop. It fills the gap with the most plausible-sounding continuation. That is confabulation. Not hallucination.

The distinction matters clinically, but it also matters institutionally. Confabulation implies a design constraint: a system architecture that structurally cannot distinguish between what it knows and what it is generating. Hallucination implies a perceptual malfunction — something that happens involuntarily, something that cannot be anticipated, something that is, in a useful sense, the machine’s fault rather than its designers’. Calling a predictable architectural failure mode a hallucination does not merely misdescribe the phenomenon. It misassigns causation.

There is also a psychiatric ethics dimension that the industry has largely avoided discussing. Researchers writing in Schizophrenia Bulletin, one of the leading journals in psychiatry, raised formal objections to the term on two grounds: it is a misnomer, since AI lacks the sensory apparatus that hallucination requires, and it is stigmatizing — associating a technical failure mode with a symptom predominantly linked to schizophrenia, a move that both misrepresents the AI error and reinforces stigma around a serious psychiatric condition. The AI industry appropriated a clinical term from a vulnerable patient population, made it cute, and moved on. No one in the boardrooms that approved these communications has had to answer for that.

The Frankfurt Problem the Industry Ignored

In 2024, philosophers Michael Townsen Hicks, James Humphries, and Joe Slater at the University of Glasgow published a paper in Ethics and Information Technology with a deliberately provocative title: “ChatGPT is Bullshit.” The paper argued, through a reading of philosopher Harry Frankfurt’s canonical 2005 work On Bullshit, that LLM outputs are better understood as bullshit than as hallucinations or even confabulations. Frankfurt’s framework distinguishes the liar — who knows the truth and deliberately says something false — from the bullshitter, whose defining characteristic is indifference to truth. The bullshitter is not trying to deceive so much as to produce a convincing effect. Whether the output is true or false is structurally irrelevant to the operation.

Hicks, Humphries, and Slater argue that LLMs meet at least the “soft” version of Frankfurt’s definition: the models are, as the paper states, “in an important way indifferent to the truth of their outputs.” They do not produce truth claims — they produce plausibility claims. The paper accrued more than 800,000 views and 54 academic citations in under a year, suggesting that this framing resonated well beyond philosophy departments. Frankfurt himself wrote that “indifference to the truth is extremely dangerous” — and that accepting the proliferation of bullshit as innocuous squanders something essential. The paper makes a direct case that the word “hallucination” enables exactly that acceptance: it frames the output as a kind of error the model is trying to avoid, when structurally the model is not trying to do anything with respect to truth at all.

The industry did not engage with this argument. It continued using “hallucination.”

It Lets Companies Off the Hook — In Court

The legal dimension of this language choice is no longer theoretical. The United Nations University, writing in February 2026, articulated the core problem: “Framing a model’s fabrication as a hallucination subtly shifts the blame to the non-sentient AI. It’s presented as an unfortunate, almost charmingly human-like quirk, rather than what it is: a predictable failure mode of a specific technical architecture.”

That shift in blame has been operationalized in courtrooms. In May 2025, OpenAI defeated a defamation lawsuit brought by Georgia radio host Mark Walters after ChatGPT fabricated a false claim that Walters had embezzled funds from a gun-rights organization. Judge Tracie Cason of the Gwinnett County Superior Court ruled in OpenAI’s favor, finding that Walters failed to show the journalist who received the hallucinated output believed it was true, or that OpenAI had acted with negligence. OpenAI’s defense rested substantially on disclaimers embedded in ChatGPT’s terms of service — disclaimers that state the model may produce inaccurate information. The word “hallucination” does significant work in those disclaimers: it frames the inaccuracy as an inherent limitation rather than a knowable failure mode, making it harder to establish the negligence standard.

In October 2025, conservative activist Robby Starbuck filed a $15 million defamation lawsuit against Google after its Bard and Gemma models allegedly fabricated fabricated sexual misconduct allegations and criminal charges against him — with invented citations mimicking real news outlets to support the claims. Google’s response, according to court documents, was to argue that these were “known hallucinations” addressed in prior safety updates and that disclaimers adequately warned users. In other words: we called them hallucinations in our documentation, so we told you. The word itself became a legal shield.

Legal scholar Damien Charlotin’s AI Hallucination Cases Database now catalogs more than 200 judicial decisions globally — over 125 in the United States alone — involving AI-generated fabricated content, including invented citations, false quotations, and misrepresented legal precedents. In most of those cases, the term “hallucination” appears in the company’s filings. It is doing exactly what it was designed to do.

Naomi Klein Was Right — And Right For The Wrong Reasons

Naomi Klein published a pointed essay in The Guardian in May 2023 with a title that cut to the structural argument: “AI Machines Aren’t ‘Hallucinating.’ But Their Makers Are.” Klein’s argument was primarily about mythology — she argued that by using a term borrowed from human psychology and mysticism, AI’s boosters were “simultaneously feeding the sector’s most cherished mythology,” the claim that building large language models is the act of birthing animate intelligence.

Klein proposed alternatives like “algorithmic junk” or “glitches.” Those terms are too reductive to capture the specific texture of the problem — a model that fabricates a peer-reviewed citation with the exact formatting of a real journal is doing something qualitatively different from a glitch. But her core observation about the term’s function was correct: “hallucination” serves the industry by making the failure seem exotic and human-like, which simultaneously limits accountability and expands the mystique. It is the corporate communications equivalent of calling a structural engineering failure a “building’s dream.”

Where Klein’s critique is strongest is in naming the dual function: the word acknowledges fallibility (protecting against the charge of dishonesty) while encoding a flattering anthropomorphism (advancing the broader narrative that these systems are minds, not machines). That is a very sophisticated piece of language design — possibly too sophisticated to have been accidental.

The Numbers Don’t Care About the Branding

Whatever you call it, the phenomenon has a documented scale and a documented cost.

Research compiled by Suprmind in 2026 draws on multiple industry benchmarks and finds that even the best-performing AI models hallucinate at a minimum rate of 0.7% on basic summarization tasks — rising to 18.7% on legal questions and 15.6% on medical queries. No major deployed model has solved the problem. The global financial cost of AI fabrications reached $67.4 billion in 2024 alone. Those are not hallucinations. Those are expensive, predictable errors generated by systems whose core architecture structurally cannot distinguish between what they know and what they are generating.

MIT research from January 2025 documented a finding that should be cited every time someone defends the “hallucination” framing: when AI models produce incorrect information, they use significantly more confident language than when they produce accurate information. Models were 34% more likely to deploy certainty markers — words like “definitely,” “certainly,” and “without doubt” — when generating fabricated content. The more wrong the model is, the more certain it sounds. That is not a hallucination. That is a system optimizing for persuasiveness at the direct expense of accuracy. The International AI Safety Report 2026, guided by over 100 international AI experts and submitted to governments across 30+ countries, noted that AI systems “still sometimes generate false information” and produce “inconsistent outputs even when given identical or similar inputs” — calling this an evaluation gap, not a perceptual anomaly.

Duke University Libraries research from 2026 found that 94% of students surveyed believe generative AI accuracy varies significantly across subjects, and 90% want clearer transparency about limitations. The public is not confused about what is happening. It is confused about why the industry is describing it the way it does. That is a different kind of failure — and it belongs to the communicators, not the model.

The Language of Accountability

The United Nations University named three specific harms produced by the “hallucination” framing. It obscures the technical reality. It misleads the public by framing AI as a flawed human mind rather than a probabilistic text engine. And it lets developers off the hook by attributing fabrication to the non-sentient system rather than to the architecture, training regime, and deployment decisions that produced it.

All three of those harms are real. All three of them compound over time. And all three of them benefit the same parties — the companies building, deploying, and defending these systems in court and in the press.

Usama Fayyad at Northeastern University argues that precise language serves a practical function independent of the politics: “Demystifying the technology and the behaviors exhibited by algorithms, good or bad, establishes real progress and creates valuable outcomes on all fronts: theoretical, academic, commercial and practical.” His point is that the word “hallucination” makes the problem sound inexplicable, when the problem is almost entirely explicable — rooted in the probabilistic nature of next-token prediction, compounded by training on contradictory data, and made worse by reinforcement learning that rewards confident-sounding outputs over accurate ones.

The alternatives are not perfect. Confabulation is precise but inaccessible to most readers. Fabrication implies intent that the models do not have in the agentic sense. “Predictive error” is accurate but bloodless. Benj Edwards at Ars Technica has argued for “confabulation” as the least misleading metaphor available — it preserves the clinical nuance without anthropomorphizing the failure. Researchers in the Harvard Data Science Review have proposed “factual inconsistency” as a domain-neutral descriptor. All of these are imperfect. None of them are as harmful as what the industry currently uses.

What the Industry Gained From One Word

The choice of “hallucination” was not made by a rogue researcher. It was made by institutions — OpenAI, Google — with massive financial incentives to shape how their products’ failures were perceived. The word traveled from a 2015 technical blog post through a 2018 academic paper into the press releases, product disclaimers, congressional testimony, and courtroom filings of the world’s most powerful technology companies. At each step, it carried the same payload: the failure is exotic, involuntary, human-like, and therefore forgivable. It is not a design choice. It is not a deployment decision. It is not a knowable risk that could have been disclosed more clearly. It is a hallucination — and what do you expect from a mind that dreams?

That framing has survived as long as it has because it is genuinely useful to the industry. It allowed companies to acknowledge a real problem while avoiding the language that would attach real responsibility to it. It allowed product managers to write disclaimers that technically said “our output may be inaccurate” while the marketing simultaneously claimed their systems were ready for medical, legal, and financial deployment. It allowed executives to tell congressional committees that hallucinations are a “recognized problem” — Sundar Pichai used those exact words — while the legal team used the same recognition to argue that disclaimers satisfied any duty of care.

The word, in short, solved a communications problem that should have been an engineering problem. It made forgiveness easier to seek and harder to withhold.

It Is Time To Call This What It Is

Language is not neutral, and the language of a trillion-dollar industry is not accidental. When a large language model tells you with full confidence that a legal case exists, cites a judge by name, and produces a docket number — and all of it is fabricated — that is not a hallucination. It is a predictive fabrication generated by an architecture that was never designed to verify truth, deployed in a context that requires it, with disclaimers written to limit liability rather than inform users.

Calling it a hallucination does not make it less harmful. It makes accountability less legible. It trains the public to accept a certain level of confident error as a charming quirk rather than a structural condition requiring remediation. It gives courts a softer target to defend against. And it gives the industry a vocabulary that describes the failure on the industry’s terms — mystified, anthropomorphized, and fundamentally beyond responsibility.

The researchers have named the alternative. The courts are beginning to force the question. The money lost to fabrication — $67.4 billion in 2024, on its way to more — is real regardless of what word sits at the top of the disclaimer. What remains is a press, a regulatory apparatus, and a user base that can choose to adopt more precise language, or can continue accepting the word the industry chose for reasons the industry understood better than it said.

The machines are not hallucinating. They are doing exactly what their architecture predicts. That is a far more serious problem — and a far more solvable one — than the current vocabulary allows us to see.