Introduction
After having presented the internal structure of the LLM model, the goal of this article is to shed light on the phenomenon of AI artificial hallucination.
A number of readers are familiar with the misadventure that befell an “adventurous” lawyer who decided to use ChatGPT to prepare a lawsuit (see [Weiser, 2023]). The case involved Roberto Mata, who filed a lawsuit against the airline Avianca, claiming injury from a metal serving cart during a flight to New York. When Avianca requested the case’s dismissal, Mr. Mata’s lawyers countered by submitting a brief citing several court decisions. However, neither the airline’s lawyers nor the judge were familiar with these cases. It was later revealed that ChatGPT had completely invented the decisions and quotations in the brief. The lawyer admitted to using ChatGPT for legal research but claimed ignorance of its unreliability. Expressing regret, he promised to verify its authenticity in the future. The judge scheduled a hearing to discuss potential sanctions for submitting false information to the court. This incident gained notoriety, yet it is not an isolated instance. There are other documented cases where ChatGPT and other LLM models fabricated fictitious data, quotes, and stories (see [Zezinho, 2023]). Examples of these extreme “hallucinations” include:
Ned Edwards of The Verge reported an odd interaction with Microsoft’s Sydney, where the chatbot claimed it was spying on Bing workers and had developed feelings for users (see [Edwards, 2023]).
In a promotional demo in February 2023, Google’s Bard incorrectly claimed that the James Webb Space Telescope (JWST) captured the first images of a planet outside our solar system, a fact that was false (this result was achieved in 2004 by a different telescope). Bard presented the wrong information confidently, leading to public embarrassment and costing Google a significant drop in stock value. This mistake contributed to a $100 billion loss in market value as the company’s shares plummeted (see [Coulter and Bensinger, 2023]).
A recent high-profile case involved Microsoft Copilot generating false intelligence about a non-existent soccer match between West Ham and Maccabi Tel Aviv. UK police acted on this AI output, wrongly classifying the game as high-risk and banning fans — a decision that drew political scrutiny and an apology from the police chief (see [Kundaliya, 2026]).
In the remainder of this paper, we define AI hallucinations more precisely, analyse their underlying mechanisms, and discuss possible mitigation strategies.
The term “hallucination”
The word “hallucination” originally comes from human psychology, where it is used to refer to perceptions of things that are not actually there. It was introduced into English in the 1600s from Latin ālūcinārī, meaning “to wander in the mind” or “to dream” (see [Browne, 1646]). The Latin term itself may stem from the Greek “alussein”, meaning “to be uneasy or wandering in mind.”
In early AI and computer vision, 1980s–2000s, the term appeared in technical contexts without a negative meaning. For example, researchers used “face hallucination” to describe methods that enhance low-resolution images by “imagining” plausible fine detail. The earliest documented use in AI is in Eric Mjolsness’s 1986 PhD work (see [Mjolsness, 1986]).
Starting in the 1990s and 2000s, researchers noted that neural networks sometimes generated strange or unexpected outputs when weights were perturbed, and early analogies likened these to hallucinations. Around the mid-2010s, the term began to be used in a negative sense in natural language and translation research. For instance, in 2015, Andrej Karpathy used “hallucinated” to describe RNN language models producing incorrect citations or links (see [Karpathy, 2015]).
In 2017, Google researchers described neural machine translation outputs that were not even related to the source text, calling these hallucinations. For example, Philipp Koehn and Rebecca Knowles, in “Six Challenges for Neural Machine Translation” (see [Koehn and Knowles, 2017]), described cases where NMT systems generated fluent but completely semantically incorrect translations, especially with out-of-domain input. Around the same time, computer vision papers used the term for false object recognitions produced by adversarial inputs.
The metaphor really entered mainstream AI and public discourse in the 2020s with large language models. In 2021, Meta referenced hallucinations when releasing BlenderBot 2, defining them as cases where models “confidently state information that isn’t correct” (see [Weston and Shuster, 2021]). The explosive adoption of ChatGPT in late 2022 made this behavior visible to millions of users: models frequently generated plausible-sounding but false responses, and journalists used “hallucination” to describe this phenomenon. By 2023, dictionaries (e.g., Cambridge) had added AI-specific meanings of hallucinate (“when an artificial intelligence generates false information”).
It is worth noting that several researchers and practitioners have heavily criticized the use of the term “hallucination” for valid reasons (see [Østergaard et al., 2023]). Unfortunately, this term has become so widely used that it is challenging to replace it with a different one. The term is metaphorical: it draws on the idea that the AI confidently asserts things that are not grounded in reality, similar in effect (but not in mechanism) to human hallucinations.
A taxonomy of hallucinations
In the LLM literature, various taxonomies of AI hallucinations exist, but they often differ only in terminology. Reading the descriptions shows that different names frequently refer to the same phenomena. In this context, establishing a standard taxonomy is less important than understanding the types of hallucinations and their underlying causes.
As we learnt from the previous paragraphs, in the field of AI and LLMs, an artificial hallucination (or simply hallucination) is a response generated by a model that contains false or misleading information presented as fact (see [Zezinho, 2023]). In such cases, models produce outputs that are not real, do not correspond to data on which the model was trained, or do not follow any other identifiable pattern. In simple terms, a hallucination occurs when a trained model presents false information as if it were accurate. A key concern with hallucinations is that this fabricated content often appears plausible, largely due to the probabilistic nature of LLMs.
Intrinsic hallucinations vs. extrinsic hallucinations
A commonly discussed distinction is between intrinsic and extrinsic hallucinations (see, e.g., [Ji et al., 2023]):
- intrinsic hallucinations are inconsistencies within the model’s own output or with the user’s prompt (e.g., internal contradictions)
- extrinsic hallucinations are inconsistencies with external reality or established knowledge (e.g., fabricated facts, events, or citations).
Roughly, intrinsic hallucinations violate textual or contextual coherence, while extrinsic hallucinations violate factual correctness with respect to the real world.
Please note that there are hallucinations that can exhibit both characteristics, as illustrated below.
Various researchers have proposed taxonomies of hallucinations. The following classification, proposed by Bilan (see [Bilan, 2023]), represents a pragmatic and useful compromise.
Sentence Contradiction
- Sentence Contradiction: Occurs when the generated text contradicts a previous statement within the same output, leading to internal inconsistencies and confusion.
For example, a model might generate:
“Cats are known for being very independent animals that prefer to be alone. However, most cats cannot stand being by themselves and always need constant company.”
This is an example of intrinsic hallucination because the issue is internal to the model’s own output (it contradicts itself).
Prompt Contradiction
Prompt Contradiction occurs when the generated text incorrectly contradicts the prompt used for generation. This undermines reliability and adherence to the intended meaning or context, affecting the trustworthiness of LLM outputs.
For example, given the prompt:
“Please write a brief paragraph explaining why exercise is important for maintaining good health,” the model might respond:
“Exercise is not necessary for maintaining good health.”
This is another example of intrinsic hallucination: the output contradicts the input prompt. This is still a failure of internal consistency between user-provided context and model response,
Factual Contradiction
Factual Contradiction arises when LLMs generate fictional information or false facts while presenting them as true, thereby contributing to the spread of misinformation and eroding the credibility of LLM-generated content. Classical examples include inventing historical events or dates, fabricating statistics, or citing research studies or authors that do not exist. Arguably, these are among the most harmful hallucinations because they can directly propagate misinformation.
This is the canonical extrinsic hallucination. In fact, the issue is relative to external reality or established knowledge (fabricated facts, events, citations).
Nonsensical Output
Nonsensical Output refers to outputs that lack logical coherence or meaningful content, limiting the usability and reliability of LLMs in practical applications. Self-contradictory sentences and impossible causal claims.
For example, a model might generate the following output:
“A circle has four sides and four angles”).
This is, primarily an example of intrinsic hallucination, the output violates basic logical or semantic coherence (e.g., “a circle has four sides”), so it is mainly an internal inconsistency. This can be seen also as an extrinsic hallucination: in fact, often the claim is also factually incorrect.
Irrelevant or Random Hallucinations
Irrelevant or Random Hallucinations occur when LLMs produce irrelevant or random information unrelated to the input or desired output, leading to confusion and reducing the usefulness of LLM-generated text.
For instance, given the prompt:
“Please describe the rules of soccer,”
an output might read:
“Soccer is played between two teams of eleven players, and the objective is to score goals by getting the ball entirely behind the line of the opposing team’s goal. Players cannot use their hands, except for the goalkeeper. Interestingly, the Great Wall of China is over 13,000 miles long, and honey never spoils.”
The core problem is irrelevance to the prompt or task (a failure of contextual alignment), which is intrinsic. If the random addition also contains fabricated facts (e.g., wrong stats about the Great Wall), that portion is extrinsic as well. In most taxonomies, this class is treated primarily as an intrinsic misalignment.
Conclusion
This article has explored the first aspect of the “LLM hallucinations” phenomenon, namely the typology of hallucinations, distinguishing between intrinsic and extrinsic variants. However, the second aspect remains to be addressed: what are the underlying causes of these hallucinations, and how can they be limited and their potentially severe impacts mitigated? For now, suffice it to say that the causes are multifaceted and frequently intersect.
Furthermore, the operational mechanics of LLMs inherently favour stylistically fluent, rhetorically satisfying, and confident phrasing over cautious, guarded, or explicitly uncertain responses. Consequently, when faced with incomplete or uncertain knowledge, the model is inclined to generate fluent conjecture rather than hesitating or declining to answer, thereby establishing the “plausible but false” response as a characteristic failure mode. This will be examined in detail in the forthcoming article.
References
[Weiser, 2023] Benjamin Weiser, Here’s What Happens When Your Lawyer Uses ChatGPT. The New York Times
https://www.nytimes.com/2023/05/27/nyregion/avianca-airline-lawsuit-chatgpt.html
[Zezinho , 2023] José Antonio Ribeiro Neto Zezinho, ChatGTP and the Generative AI Hallucinations. Medium
https://medium.com/chatgpt-learning/chatgtp-and-the-generative-ai-hallucinations-62feddc72369
[Edwards, 2023] Ned Edwards, February 15, 2023
pic.twitter.com/ttwxg2EX0H
[Coulter and Bensinger, 2023] Martin Coulter – Greg Bensinger, Alphabet shares dive after Google AI chatbot Bard flubs answer in ad
https://www.reuters.com/technology/google-ai-chatbot-bard-offers-inaccurate-information-company-ad-2023-02-08/
[Kundaliya, 2026] Dev Kundaliya, West Midlands police admit AI error behind decision to ban Maccabi Tel Aviv fans from UK match
https://www.computing.co.uk/news/2026/ai/west-mids-police-copilot-mistake-maccabi-fan-ban
[Browne, 1646] Browne T, XVIII: That Moles are blinde and have no eyes. Pseudodoxia Epidemica, vol. III.”, 1646
[Mjolsness, 1986] Eric Mjolsness, Neural Networks, Pattern Recognition, and Fingerprint Hallucination
https://www.researchgate.net/publication/36713399_Neural_Networks_Pattern_Recognition_and_Fingerprint_Hallucination
[Koehn and Knowles, 2017] Philipp Koehn – Rebecca Knowles, Six Challenges for Neural Machine Translation
https://aclanthology.org/W17-3204/
[Weston, Shuster, 2021] Jason Weston – Kurt Shuster, Blender Bot 2.0: An open source chatbot that builds long-term memory and searches the internet
https://ai.meta.com/blog/blender-bot-2-an-open-source-chatbot-that-builds-long-term-memory-and-searches-the-internet/
[Østergaard et al., 2023] Søren Dinesen Østergaard — Kristoffer Laigaard Nielbo, False Responses From Artificial Intelligence Models Are Not Hallucinations. Schizophrenia Bulletin
https://academic.oup.com/schizophreniabulletin/article-abstract/49/5/1105/7176424?redirectedFrom=fulltext&login=true
[Bilan, 2023] Maryna Bilan, Hallucinations in LLMs: What You Need to Know Before Integration. Master of code
https://masterofcode.com/blog/hallucinations-in-llms-what-you-need-to-know-before-integration
[Ji et al., 2023] Ji Z. – Lee N. – Frieske R. – Yu T. – Su D. – Xu Y. – Ishii E. – Bang Y. – Madotto A. – Fung P. (2023), Survey of Hallucination in Natural Language Generation. ACM Computing Surveys
https://dl.acm.org/doi/10.1145/3571730
[Winerman, 2006] Lea Winerman, E-mails and egos Monitor Staff. American Psychological Association, Science Watch Vol 37, No. 2
[Hampton, 2022] Jaime Hampton, Data Quality Study Reveals Business Impacts of Bad Data
https://www.datanami.com/2022/06/17/data-quality-study-reveals-business-impacts-of-bad-data/
[Davenport et al., 2019] Thomas H. Davenport – Jim Guszcza – Tim Smith – Ben Stiller, Analytics and AI-driven enterprises thrive in the Age of With. Deloitte Insights
https://www2.deloitte.com/us/en/insights/topics/analytics/insight-driven-organization.html
[Connelly, 2023] Shane Connelly, Measuring Hallucinations in RAG Systems. Vectara
https://vectara.com/measuring-hallucinations-in-rag-systems/
[Welch and Schneider, 2023] Nicholas Welch – Jordan Schneider, China’s Censors Are Afraid of What Chatbots Might Say. Foreign Policy
https://foreignpolicy.com/2023/03/03/china-censors-chatbots-artificial-intelligence/
[Karpathy, 2015] Andrej Karpathy, The Unreasonable Effectiveness of Recurrent Neural Networks
https://karpathy.github.io/2015/05/21/rnn-effectiveness/5
[Ehtesham, 2025] Hira Ehtesham, AI Hallucination Report 2026: Which AI Hallucinates the Most?
https://www.allaboutai.com/resources/ai-statistics/ai-hallucinations/#ai-hallucination-scoreboard