AI Hallucination

Phenomenon where an AI model—usually an AI chatbot powered by Large Language Models (LLMs) or computer vision programs—produce a confident but factually inaccurate or nonsensical responses or perceives objects and patterns that do not exist
Kayla Rose van Kooten 2026-10-16

Explication

AI “hallucinations” occur as a side-effect of how AI programs work, as they are not grounded in any experience of the world but rather built to statistically predict words or detect patterns. Within the realm of AI, the term “hallucination” has lacked consistency in its usage and definition, as well as a precise and universally accepted definition (Maleki et al.). Alternative terms for “hallucinations” have been proposed by researchers both within and outside of the field, such as confabulations, delusion, stochastic parroting, factual errors, and fabrication (Maleki et al.). 

The first use of the term “hallucination” has been traced back to a dissertation in 1986 by Eric Mjolsness entitled “Neural Networks, Pattern Recognition, and Fingerprint Hallucination” (Mjolsness). In 2000, it was used positively in a paper for the generation of high-resolution enhancement of low-resolution inputs in face recognition (Baker and Kanade). In these early usages, “hallucinations” were seen as something positive and valuable and were used to describe the ability of computer vision programs to add details and denoise images, image inpainting, as well as image synthesis (Xiang et al.; Pumarola et al.; Baker and Kanade). 

Only later did “hallucination” begin to be used when describing errors, adapting the term for its analogous usage in language models. In a 2008 study on statistical machine translation, “hallucination” was used to describe a failure mode (Deng and Byrne). Later, in 2015, the term was used in a blog post by Andrej Karpathy—a founding member of OpenAI—to describe his recurrent neural network language model that was generating a false Yahoo URL and algebraic geometry (Karpathy). In 2017, Google researchers used the term to describe the responses generated by neural machine translation models that were “quite fluent but entirely unrelated to the input” (Koehn and Knowles). By 2018, usage of the term has begun to pick up in the domains of Natural Language Generation and LLMs (Maleki et al.).  

Since the proliferation of LLM-based chatbots, the term “hallucination” has gained wider recognition and usage beyond the discipline of computer science. After OpenAI released the beta version of ChatGPT in November 2022, users noticed that the program would often generate confident and plausible-sounding fabrications (Weise and Metz). In 2023, “hallucinate” was the Cambridge Dictionary’s 2023 Word of the Year, and the dictionary updated its definition of hallucination to include the AI-specific meaning of the term (Creamer). Dictionary.com also awarded “hallucination” as its 2023 word of the year, noting a 46% increase in dictionary lookups over the previous year (Dictionary.com’s 2023 Word of the Year Is… | Dictionary.com).

The use of the term “hallucination” as a metaphor, however, remains controversial and has been described as a “misnomer” since AI lacks sensory perceptions, and its errors arise from data and prompts rather than the absence of any external stimuli (Maleki et al.). It has been criticized for stigmatizing mental illnesses—particularly schizophrenia—and for further anthropomorphizing AI models by attributing human characteristics to them, likening AI “hallucinations” to the psychiatric phenomenon of false sensory perceptions that feel real but occur without any actual external stimulus (Østergaard and Nielbo). Using the term “hallucination” reinforces the metaphor of already treating AI systems and computers as if they were humans with brains and subjectivity, alongside ideas of intentionality, authorship, beliefs, and cognition.  

The metaphor of “hallucination” also invokes similarities with psychedelics, particularly LSD and psilocybin. In 2015, Google engineers developed a computer vision program called DeepDream that uses a convolutional neural network designed to find and then enhance patterns or faces in images, connecting it back to the original definition of “hallucinations”, adding details to images (Mordvintsev et al.). This program was used by the developers to falsely detect and enhance patterns and images via “algorithmic pareidolia”, thus creating deliberately overprocessed images that resembled psychedelic hallucinations (LaFrance). This additional use of “hallucination” once again suggests a resemblance between these models and the workings of the human brain. This can speculatively be connected to the Silicon Valley’s history of counterculture and recreational psychedelic drug usage (Tvorun-Dunn).  

When considering the criteria by which one judges AI “hallucinations”, if one is looking for accurate information, “hallucinations” are explicitly seen as errors. However, when used for more creative pursuits, their usage opens up new potentials. Going back to the original, more positive connotations of “hallucinations”, their ungroundedness can be used to create more details used for surreal visual storytelling, or “sympoietic creativity” (Ge; Halperin and Lukin).

See Also

• Artificial Intelligence (AI) - Simulation of human intelligence processes by computer systems, to create or interpret content in innovative and sometimes literary ways
• Chatbot - Software application designed to simulate conversation with human users, often used in digital narratives to provide interactive storytelling experiences
• Unreliable Narrator - A narrator who speaks and narrates in a fashion that is not in accordance with the implied author’s norms, often understood as a narrator who, intentionally or not, lies and/or deceives

Works Referenced

Baker, Simon, and Takeo Kanade. “Hallucinating Faces.” Proceedings Fourth IEEE International Conference on Automatic Face and Gesture Recognition (Cat. No. PR00580), IEEE Computer Society, 2000, pp. 83–88. https://doi.org/10.1109/AFGR.2000.840616. 

Creamer, Ella. “‘Hallucinate’ Chosen as Cambridge Dictionary’s Word of the Year.” The Guardian, 15 Nov. 2023, https://www.theguardian.com/books/2023/nov/15/hallucinate-cambridge-dictionary-word-of-the-year. 

Deng, Yonggang, and W. Byrne. “HMM Word and Phrase Alignment for Statistical Machine Translation.” IEEE Transactions on Audio, Speech, and Language Processing, vol. 16, no. 3, Mar. 2008, pp. 494–507. https://doi.org/10.1109/TASL.2008.916056 

“Dictionary.com’s 2023 Word of the Year Is…” Dictionary.com, 12 Dec. 2023, https://www.dictionary.com/articles/word-of-the-year-2023. 

Ge, Liang. “Spectral Imaginings and Sympoietic Creativity: AI Hallucinations and the Ethics of Posthuman Creativity.” Big Data & Society, vol. 12, no. 4, Dec. 2025, article 20539517251400696. https://doi.org/10.1177/20539517251400696. 

Halperin, Brett A., and Stephanie M. Lukin. “Artificial Dreams: Surreal Visual Storytelling as Inquiry Into AI ‘Hallucination.’” Proceedings of the 2024 ACM Designing Interactive Systems Conference, 2024, pp. 619–637. https://doi.org/10.1145/3643834.3660685. 

Karpathy, Andrej. “The Unreasonable Effectiveness of Recurrent Neural Networks.” Andrej Karpathy Blog, 21 May 2015, https://karpathy.github.io/2015/05/21/rnn-effectiveness/. 

Koehn, Philipp, and Rebecca Knowles. “Six Challenges for Neural Machine Translation.” arXiv, 12 June 2017, https://doi.org/10.48550/arXiv.1706.03872. 

LaFrance, Adrienne. “When Robots Hallucinate.” The Atlantic, 3 Sept. 2015, https://www.theatlantic.com/technology/archive/2015/09/robots-hallucinate-dream/403498/. 

Maleki, Negar, et al. “AI Hallucinations: A Misnomer Worth Clarifying.” 2024 IEEE Conference on Artificial Intelligence (CAI), 2024, pp. 133–138. https://doi.org/10.1109/CAI59869.2024.00033. 

Mjolsness, Eric. Neural Networks, Pattern Recognition, and Fingerprint Hallucination. California Institute of Technology, 1985. https://doi.org/10.7907/XKTBJ-10524. 

Mordvintsev, Alexander, et al. “Inceptionism: Going Deeper into Neural Networks.” Google Research Blog, 18 June 2015, https://research.google/blog/inceptionism-going-deeper-into-neural-networks/. 

Østergaard, Søren Dinesen, and Kristoffer Laigaard Nielbo. “False Responses From Artificial Intelligence Models Are Not Hallucinations.” Schizophrenia Bulletin, vol. 49, no. 5, May 2023, pp. 1105–1107. https://doi.org/10.1093/schbul/sbad068. 

Pumarola, Albert, et al. “Unsupervised Person Image Synthesis in Arbitrary Poses.” arXiv, 27 Sept. 2018, https://doi.org/10.48550/arXiv.1809.10280. 

Tvorun-Dunn, Maxim. “Acid Liberalism: Silicon Valley’s Enlightened Technocrats, and the Legalization of Psychedelics.” International Journal of Drug Policy, vol. 110, Dec. 2022, article 103890. https://doi.org/10.1016/j.drugpo.2022.103890 

Weise, Karen, and Cade Metz. “When A.I. Chatbots Hallucinate.” The New York Times, 1 May 2023, www.nytimes.com/2023/05/01/business/ai-chatbots-hallucination.html. 

Xiang, Hanyu, et al. “Deep Learning for Image Inpainting: A Survey.” Pattern Recognition, vol. 134, Feb. 2023, article 109046. https://doi.org/10.1016/j.patcog.2022.109046.

Further Reading

Dumit, Joseph, and Andreas Roepstorff. “AI Hallucinations Are a Feature of LLM Design, Not a Bug.” Nature, vol. 639, no. 8053, 2025, p. 38. https://doi.org/10.1038/d41586-025-00662-7. 

Klein, Naomi. “AI Machines Aren’t ‘Hallucinating’. But Their Makers Are.” The Guardian, 8 May 2023, https://www.theguardian.com/commentisfree/2023/may/08/ai-machines-hallucinating-naomi-klein. 

Turner, Fred. From Counterculture to Cyberculture: Stewart Brand, the Whole Earth Network, and the Rise of Digital Utopianism. University of Chicago Press, 2008.

Text is available under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International