Hallucination

What it means

When people say an AI system "hallucinated," they mean it gave an answer that sounds right but isn't. IBM defines AI hallucinations as "instances where an AI system produces outputs that sound plausible but are factually wrong."¹

Here is an everyday example. You ask a chatbot for three books on supply-chain risk. It gives you a real-sounding author, a believable title, and a publisher, for a book that was never written. The made-up answer looks just like the real ones next to it. A confident tone tells you nothing about whether a claim is true.

Where the word comes from

The word first described a human experience. Cambridge Dictionary's older meaning is "to seem to see, hear, feel, or smell something that does not exist, usually because of a health condition..."²

In computing, the word started with images, not text. A survey published in ACM Computing Surveys says "The term “hallucination” first appeared in Computer Vision (CV)," and points to a 2000 paper by Simon Baker and Takeo Kanade. In that early use the word "carried more positive meanings, such as superresolution": the software filled in detail to sharpen a low-resolution image. The survey adds: "Such hallucination is something we take advantage of rather than avoid in CV."³

The meaning turned negative once the word moved to software that writes text. In 2020, Joshua Maynez and colleagues found that summarizing systems were "highly prone to hallucinate content that is unfaithful to the input document."⁴ After chatbots reached the public, Cambridge Dictionary named "hallucinate" its Word of the Year for 2023. It added a new meaning: when an AI "hallucinates, it produces false information."²

How others define it

• Researchers (survey of the field): "the generated content that is nonsensical or unfaithful to the provided source content," which the authors call "the most inclusive and standard" definition.³
• IBM: "AI hallucinations are instances where an AI system produces outputs that sound plausible but are factually wrong."¹
• Kalai and colleagues (2025): language models produce "plausible yet incorrect statements instead of admitting uncertainty."⁵
• Cambridge Dictionary: when an AI "hallucinates, it produces false information."²

Where they agree and differ

All four agree on the core idea: the output is false, and it looks believable.

They disagree about what the answer is checked against. The research definitions measure it against a source the system was given, so the question is whether the summary is faithful to the document.³⁴ The dictionary and IBM measure it against the world, so the question is whether the answer is true.¹² The difference matters at work. A system can repeat your document faithfully and still be wrong if the document is wrong.

They also disagree about the cause. Kalai and colleagues argue that hallucination is partly a learned habit: "Like students facing hard exam questions, large language models sometimes guess when uncertain," because the models "are optimized to be good test-takers, and guessing when uncertain improves test performance."⁵ If they are right, changing how systems are scored would reduce the problem. That is a hopeful finding, but it does not promise the problem will go away.

The strongest objection is to the word itself. Cambridge's announcement noted that the new meaning "illustrates a growing tendency to anthropomorphise AI technology,"² which means talking about a machine as if it had a human mind. A machine does not see or perceive anything. The objection is fair: the word can make a predictable design and scoring problem sound like an odd quirk of personality. We keep the word because it is the one readers meet in the news. Read it as "confident error."

In the news

In October 2025, the Associated Press reported that Deloitte Australia would "partially refund the 440,000 Australian dollars ($290,000) paid by the Australian government" for a report that contained apparent AI-generated errors. A University of Sydney researcher who spotted the problems told the AP: "I instantaneously knew it was either hallucinated by AI or the world's best kept secret." The revised report added "a disclosure that a generative AI language system, Azure OpenAI, was used in writing the report." Deloitte said the "matter has been resolved directly with the client."⁶ We are confident about these facts as the AP reported them.

The best-known earlier case came from a 2023 New York lawsuit against the airline Avianca. Bloomberg News reported that "The lawyers behind a court brief filled with phony case citations dreamed up by ChatGPT were fined $5,000."⁷

In both cases, professionals put AI-written work in front of a client or a court without checking it, and someone else found the errors.

What it means for your organization

The Knowledge Core argues that agents are individuals, imperfect like us.⁸ Hallucination is that imperfection at its most practical. It is also why the book's rule matters: [a person stays in the middle](/definitions/a-person-in-the-middle) of every chain of command.⁸ Checking is not a sign of distrust. It is a job, and someone has to own it. Before anything AI-assisted leaves your organization, a named person should check every source, number, and quote against the original. In both cases above, the people outside the firm found the errors first.

Ask your team: For each AI-assisted document we send out, who checks every source, and how would we know if they didn't?