SAN FRANCISCO, July 19 — On a screen somewhere in the city, a chatbot is telling a user, with perfect confidence, something that never happened. A citation link that leads nowhere. A historical date that is wrong.
A medical fact that is invented. This is the hallucination problem—an AI generating false or misleading information presented as fact—and it has followed the technology from obscure academic papers to dinner-table conversations since November 2022.
The term itself has a longer history than many realise. It was introduced in the PhD thesis of Eric Mjolsness back in 1986. In those early days, in computer vision, hallucination meant something different—adding detail to improve an image, such as generating a clear face from low-resolution inputs.
A notable work in this vein was the face hallucination algorithm published in 1999 by Simon Baker and Takeo Kanade. By the 2000s, the term had shifted.
In statistical machine translation, researchers began describing hallucinations as a failure mode. The 2010s brought the meaning closer to what users encounter today: the generation of factually incorrect or misleading outputs by AI systems.
The Point of No Return: 2015 and Beyond
The year 2015 marked two key moments. Researchers Saurabh Gupta and Jitendra Malik identified hallucinations in visual semantic role labeling tasks. Separately, computer scientist Andrej Karpathy used the word “hallucinated” in a blog post to describe his recurrent neural network language model generating an incorrect citation link.
Two years later, in 2017, Google researchers used the term to describe responses generated by neural machine translation models when they were unrelated to the source text. In 2018, it appeared in computer vision again, describing erroneous detection of non-existent objects through adversarial attacks.
Then came the chatbots. In July 2021, Meta warned during its release of BlenderBot 2 that the system is prone to hallucinations, defined as “confident statements that are not true.” The warning was a precursor to a much larger public reckoning. Following OpenAI’s release of ChatGPT in beta in November 2022, users complained that these chatbots embed plausible-sounding random falsehoods.
The complaints were loud enough that many news outlets, including The New York Times, started using “hallucinations” to describe these models’ frequently incorrect responses. By 2023, the Cambridge Dictionary had updated its definition of hallucination to include this AI-specific sense.
A Term Under Fire
Not everyone is comfortable with the metaphor. Some researchers have highlighted a lack of consistency in how the term is used, identifying alternatives such as confabulations, fabrications, and factual errors.
Definitions from OpenAI, published in May 2023, include “a tendency to invent facts in moments of uncertainty” and “a model’s logical mistakes.” That same month, CNBC described it as “fabricating information entirely, but behaving as if spouting facts.” The Verge, in February 2023, called it “making up information.” In scientific contexts, it is referred to as “probability distributions.” Journalist Benj Edwards, writing in Ars Technica, argues that “hallucination” is controversial but that some metaphor remains necessary.
He suggests “confabulation” as an analogy for creative gap-filling—a term that may better capture the process by which a large language model fills in missing information with invented data.
What Is at Stake
The distinction matters because the consequences are not theoretical. In July 2024, a White House report on fostering public trust in AI research mentioned hallucinations only in the context of reducing them. Not debating the term, but eliminating the phenomenon.
Symbolic AI models—rule-based systems that rely on explicit logic—generally do not produce hallucinations, unlike large language models, which generate text based on probability. That probability-based approach is what makes them powerful, and what makes them prone to fabricating information entirely while behaving as if spouting facts.
Detecting and mitigating errors and hallucinations pose significant challenges in high-stakes scenarios. Chip design. Supply chain logistics.
Medical diagnostics. In each of these fields, a confident falsehood from an AI system is not a minor inconvenience—it is a potential catastrophe.
The scene on the ground, for developers and researchers, remains one of urgency: build systems that can admit uncertainty, or risk the consequences of machines that do not know when they are lying.
