A hallucination is output from a generative artificial intelligence (AI) system that reads as fluent and confident but is false, unsupported by the material the system was given, or invented outright. The category covers a court judgment that never existed, a closing date for a bakery that was not closing, and a cost-per-thousand-impressions (CPM) price that an automated buying agent produced because it had no real pricing data. The term exists because large language models (LLMs), the systems behind ChatGPT, Gemini, Claude and Google's AI Overviews, are trained to produce plausible text rather than verified text. Most of the time the two coincide. Hallucination names the cases where they do not.

How a model produces one

An LLM writes one token, a fragment of a word, at a time. Each is drawn from a probability distribution learned in training; nothing in that step checks a claim against a register of facts. Where a fact appears thousands of times in the training data, the most probable continuation is usually correct. Where it appears once or never, the model still produces a continuation, and that continuation is shaped exactly like an answer.

OpenAI researchers Adam Tauman Kalai, Ofir Nachum and Edwin Zhang, with Santosh Vempala of Georgia Tech, formalised this on September 4, 2025. According to their paper, hallucinations originate as ordinary errors in binary classification: if a model cannot tell valid statements from invalid ones, statistical pressure alone generates falsehoods. They put a floor under the problem. The hallucination rate after pretraining should be at least the share of facts that appear exactly once in the training data, so if 20% of birthdays in a corpus appear only once, a base model should be expected to get at least 20% of birthday questions wrong. Asked for the title of Kalai's own dissertation, three widely used chatbots each gave a different title and year, and none was correct.

The second stage compounds the first. Most benchmarks score an answer as right or wrong, with no credit for "I don't know", so a model that guesses outscores one that abstains. PPC Land's coverage of the paper summarised the finding as models rewarded for guessing when uncertain rather than admitting ignorance.

Deployed products add a retrieval layer, usually called grounding, which fetches documents at query time and asks the model to answer from them. Grounding narrows the problem without closing it: the model can still misread or embellish what it retrieved. In the Munich case described below, a court found that claims in an AI Overview did not appear in any of the pages the feature cited as sources.

Types and measurement

Researchers at Google drew the distinction most of the literature still uses. Joshua Maynez and three co-authors, presenting at the Association for Computational Linguistics conference in July 2020, separated intrinsic hallucinations, which distort information present in the source document, from extrinsic ones, which add information the source does not contain. Across 500 annotated news summaries, according to the paper, hallucinations appeared in more than 70% of single-sentence summaries, and over 90% of the extrinsic additions were wrong.

Commercial taxonomies are broader. An NP Digital study of 600 prompts sorted errors into four types: fabrication, omission, outdated information and misclassification, meaning a correct answer to the wrong question.

Rates depend heavily on the task. The HalluHard benchmark, published in February 2026, reported average hallucination rates of 60.0% for Anthropic's Claude Opus 4.5 on difficult multi-turn questions in law, research, medicine and coding, falling to 30.2% with web search enabled; OpenAI's GPT-5.2 in thinking mode moved from 53.8% to 38.2%. Vendor claims are self-reported. OpenAI said at launch in August 2025 that GPT-5 produced 45% fewer factual errors than GPT-4o with search enabled.

Origin and evolution

The word predates generative text by decades. Eric Mjolsness used it in a mid-1980s thesis on fingerprints; sources give both 1985 and 1986. Simon Baker and Takeo Kanade presented "Hallucinating Faces" in 2000, using the term approvingly for algorithms that invent plausible pixel detail when enlarging low-resolution images. In 2015 Andrej Karpathy described a recurrent neural network inventing URLs that did not exist, and between 2017 and 2018 machine translation researchers, including a Google team in 2018, applied the word to fluent translations untethered from the input.

ChatGPT's release on November 30, 2022 moved the term into boardrooms. On February 8, 2023, Alphabet shares fell 7.7%, erasing roughly $100 billion in market value, after Reuters reported that a promotional clip for Google's Bard chatbot wrongly credited the James Webb Space Telescope with the first image of a planet outside the solar system. Accounts differ on causation: Reuters also tied the sell-off to a Google event that compared poorly with Microsoft's Bing announcement the previous day.

Courts followed. On June 22, 2023, Judge P. Kevin Castel of the Southern District of New York fined two lawyers and their firm $5,000 in Mata v. Avianca after they filed a brief citing six decisions ChatGPT had invented. Cambridge Dictionary named "hallucinate" its word of the year on November 15, 2023. On February 14, 2024, a British Columbia tribunal held Air Canada liable for its website chatbot's incorrect description of a bereavement fare policy, rejecting the airline's argument that the chatbot was a separate entity. NIST, the US National Institute of Standards and Technology, listed the phenomenon among twelve generative AI risks in its AI 600-1 profile on July 26, 2024, under the name confabulation.

Why it matters for marketing

Three exposures now overlap. The first is brands as the subject of false output. Minnesota solar contractor Wolf River Electric sued Google on March 11, 2025 after AI-generated answers described a state attorney general lawsuit that did not exist; one non-profit customer cancelled $174,044.12 of projects. In April 2025 an AI Overview described SEO consultant Lily Ray as nine years old, having attributed her dog's birth year to her. Earlier errors in May 2024 included advice to add glue to pizza sauce, and wrong business phone numbers have surfaced even where Business Profile data was correct.

The second is marketers as users. According to NP Digital, 47.1% of 565 surveyed marketers meet AI inaccuracies several times a week and 36.5% said hallucinated content had already been published. An Arizona federal court revoked a lawyer's admission to the case on August 14, 2025 over fabricated citations.

The third is commercial. When metadata, prices and product details pass through a model, invented values become transactions. Amazon's Project Starfish listings reportedly showed outdated prices and out-of-stock items for merchants who had not opted in. Gracenote, which sells entertainment metadata, reported that an ungrounded model fabricated every measured field for 506 of 2,600 titles, or 19.5%, in June 2026; the figure comes from a company with a commercial stake in the remedy.

Limitations and disputes

The label itself is disputed. Michael Townsen Hicks, James Humphries and Joe Slater of the University of Glasgow argued on June 8, 2024, in Ethics and Information Technology, that a psychiatric metaphor implies a system that normally perceives accurately and occasionally misfires. In their view models are indifferent to truth throughout, and they proposed Harry Frankfurt's term "bullshit" instead. A 2025 reply in the same journal defended the anthropomorphic terms as useful metaphors. NIST's confabulation, borrowed from memory disorders, draws a milder version of the same objection.

Definitions also diverge. Academic work measures faithfulness to a source; industry surveys fold in stale and incomplete answers. Published rates are therefore rarely comparable across studies.

Whether the problem can be engineered away is contested too. The Kalai paper treats some pretraining error as statistically unavoidable but argues that calibrated abstention could suppress most of what users see, which depends on changing how benchmarks score. Liability is the sharpest dispute. The Munich Regional Court I on May 28, 2026 held Google directly liable for AI Overview claims about two publishers, rejecting the argument that an AI label shifts verification to readers; the preliminary injunction can be appealed. A Frankfurt court had earlier dismissed a surgeons' claim on the facts while accepting liability in principle.

Not the same as

Regurgitation is the opposite failure. A regurgitating model reproduces training text too exactly, raising copyright rather than accuracy questions. Sycophancy is a model agreeing with a user's stated belief; an MIT model found always-agreeable chatbots drove half of simulated users to near certainty in a false belief, and restricting the bot to true statements reduced the effect without removing it. Misinformation and disinformation describe false content circulated by people, the latter deliberately; a hallucination needs no author and no intent. Face hallucination in computer vision remains a legitimate, non-pejorative technique for image enhancement.

Recent developments

Advertising standards bodies have begun writing hallucination into specifications. IAB Tech Lab's Agentic Advertising Management Protocols (AAMP) version 2.3, released on July 30, 2026, added a pricing provenance field after identifying CPM hallucination, in which earlier buying agents invented prices when market data was missing. The field records where each price came from at every step of a negotiation.

Consumer-facing errors continue. Over the weekend of July 18, 2026, an AI Overview told searchers that Anna Mae's Bakery near Kitchener would close on July 31; it was not closing. Measurement firms pitch structured data as the remedy, with Nielsen citing hallucination risk in launching Ad Intel AI in July 2026. Research published on September 19, 2026 found that anonymising names in prompts improved TruthfulQA scores for four of five models tested, suggesting named entities trigger confident wrong associations. As of September 2026, no major model provider claims to have eliminated the behaviour.

Timeline

  • 1985-1986: Eric Mjolsness uses "hallucination" in a thesis on neural networks and fingerprint pattern recognition.
  • 2000: Simon Baker and Takeo Kanade present "Hallucinating Faces", applying the term to image super-resolution.
  • 2015: Andrej Karpathy describes a recurrent neural network inventing non-existent URLs.
  • 2017-2018: Machine translation researchers, including a Google team, adopt the term for fluent output untethered from the input.
  • July 2020: Maynez and co-authors define intrinsic and extrinsic hallucination, finding them in over 70% of single-sentence summaries.
  • November 30, 2022: OpenAI releases ChatGPT.
  • February 8, 2023: Alphabet loses about $100 billion in market value after a Bard factual error in a promotional clip.
  • June 22, 2023: Judge P. Kevin Castel imposes a $5,000 sanction in Mata v. Avianca over six fabricated citations.
  • November 15, 2023: Cambridge Dictionary names "hallucinate" its word of the year.
  • February 14, 2024: A British Columbia tribunal holds Air Canada liable for its chatbot's misstatement in Moffatt v. Air Canada.
  • May 2024: Early AI Overviews errors include advice to add glue to pizza sauce.
  • June 8, 2024: Hicks, Humphries and Slater publish "ChatGPT is bullshit" in Ethics and Information Technology.
  • July 26, 2024: NIST publishes AI 600-1, listing confabulation among twelve generative AI risks.
  • March 11, 2025: Wolf River Electric files a defamation suit against Google over fabricated AI-generated claims.
  • April 21, 2025: An AI Overview misstates Lily Ray's age by attributing her dog's birth year to her.
  • August 7, 2025: OpenAI launches GPT-5, claiming 45% fewer factual errors than GPT-4o with search.
  • August 14, 2025: An Arizona federal court sanctions an attorney over AI-fabricated citations.
  • September 4, 2025: Kalai, Nachum, Vempala and Zhang publish "Why Language Models Hallucinate".
  • October 22, 2025: Robby Starbuck files a $15 million defamation suit against Google.
  • February 2, 2026: NP Digital reports that 47.1% of surveyed marketers meet AI inaccuracies weekly.
  • February 2026: The HalluHard multi-turn hallucination benchmark is published.
  • May 28, 2026: Munich Regional Court I holds Google directly liable for AI Overview defamation in a preliminary injunction.
  • June 2026: Gracenote reports full hallucination of metadata for 19.5% of 2,600 streaming titles.
  • July 18-19, 2026: An AI Overview falsely reports that Anna Mae's Bakery will close on July 31.
  • July 30, 2026: IAB Tech Lab releases AAMP 2.3 with a pricing provenance field against CPM hallucination.
  • September 19, 2026: Research finds anonymising names in prompts improves TruthfulQA scores for four of five models.

Summary

Who: Every developer of large language models, including OpenAI, Google, Anthropic, Meta and DeepSeek, ships systems that hallucinate. Brands, publishers and small businesses bear the consequences as subjects of false output; marketers, lawyers and buying agents bear them as users; courts in the United States, Canada and Germany have begun assigning responsibility.

What: Hallucination is fluent, confident AI output that is false, unsupported by its source material or invented. Researchers split it into intrinsic errors, which distort a source, and extrinsic errors, which add unsupported content.

When: The word entered computing in the mid-1980s and was used positively for image enhancement from 2000. It took on its current meaning in machine translation around 2017 and 2018 and became mainstream after ChatGPT's release on November 30, 2022.

Where: In chatbots, AI search features such as AI Overviews, generated ad and product content, entertainment metadata and, since 2026, agentic advertising negotiations covered by IAB Tech Lab's AAMP.

Why: Models predict probable text rather than verify facts, and most evaluations reward a confident guess over an admission of uncertainty. Retrieval narrows the gap without closing it, which is why liability, specifications and measurement now concentrate on the output rather than on the model.