LONDON, July 21 — The term “hallucination” has become a fixture in the lexicon of artificial intelligence, yet its migration from a niche computer-vision procedure to a shorthand for AI-generated falsehoods is a journey that reveals as much about the technology as it does about the people building it.
The technical reality is that the word first appeared in the field — specifically in computer vision — in the 1980s. It crops up in Eric Mjolsness’s 1986 PhD thesis, where “hallucination” meant adding detail to improve an image. The phrase was later given legs by the “face hallucination” algorithm from Simon Baker and Takeo Kanade, published in 1999. The detail that matters: back then, it was a constructive trick, not a flaw.
That changed in the 2000s, when researchers in statistical machine translation began using the term to describe a specific failure mode. By the 2010s, it had settled into its current meaning: factually incorrect outputs. In 2015, Saurabh Gupta and Jitendra Malik identified hallucinations in visual semantic role labelling. That same year, computer scientist Andrej Karpathy — writing a blog post about his recurrent neural network language model — used “hallucinated” to describe a generated citation link that was simply wrong.
The pattern hardened. In 2017, Google researchers applied the term to neural machine translation responses that bore no relation to the source text. A year later, it appeared in computer vision again, this time for erroneous detection of non-existent objects via adversarial attacks.
From technical quirk to public concern
The public face of hallucination arrived with conversational AI. In July 2021, Meta warned that its BlenderBot 2 system was prone to “confident statements that are not true.” OpenAI released ChatGPT in beta in November 2022, and users immediately noticed plausible-sounding falsehoods embedded in its content. The New York Times and other outlets adopted the term. By 2023, the Cambridge Dictionary had updated its definition of “hallucination” to include the AI-specific sense.
As ever, the detail that matters is context. Hallucinations pose significant challenges in high-stakes domains such as chip design, supply chain logistics, and medical diagnostics. Critics — including some software engineers and statisticians — argue the term unreasonably anthropomorphises computers. Under the hood, a language model doesn’t “see” things; it generates text based on statistical patterns. Calling it a hallucination, the critics say, muddies the technical reality.
A July 2024 White House report on fostering public trust in AI research mentioned hallucinations only in the context of reducing them. That narrow framing suggests policymakers are treating the problem as a bug to be squashed, not a feature of the architecture.
Symbolic AI models generally do not produce hallucinations; large language models do. The contrast is sharp. Researchers have noted inconsistency in usage, offering alternative terms such as confabulations, fabrications, and factual errors.
What to watch next: the debate over terminology is unlikely to settle soon, but the underlying challenge — how to trust a system that confidently produces falsehoods — will only grow as language models embed deeper into high-stakes work. The word may change; the problem will not.
Sources
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Hallucination in AI: From 1986 Thesis to ChatGPT’s Falsehoods”, “description”: “LONDON, July 21 — The term “hallucination” has become a fixture in the lexicon of artificial intelligence, yet its migration from a niche computer-vision procedure…”, “datePublished”: “2026-07-21T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Sebastian Smith”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “citation”: [{“@type”: “CreativeWork”, “url”: “https://en.wikipedia.org/wiki/Hallucination_%28artificial_intelligence%29”}]}
