Hallucination (artificial intelligence)

From LIMSWiki
Jump to navigationJump to search

An AI Overviews result incorrectly stating that Joaquín Correa is the brother of Ángel Correa; the two are unrelated.[1]

In artificial intelligence (AI), a hallucination is generated content that is false, unsupported, or inconsistent with the information on which the output is supposed to be based.[2][3] The term is especially associated with large language models (LLMs), which can generate fluent and plausible statements, quotations, citations, or explanations that are not factually correct or are not supported by the supplied source material.

There is no single universally accepted definition of hallucination. Research commonly distinguishes factuality, concerning correspondence with independently verifiable facts, from faithfulness, concerning consistency with a supplied source, context, or instruction.[4][3] Thus, an output can be factually correct but unsupported by the source it purports to summarize, or faithful to an inaccurate source while remaining false in the external world.

Hallucinations are a reliability problem because incorrect material may be expressed in the same confident and fluent style as correct material. Documented examples include fabricated academic references, nonexistent judicial decisions, inaccurate summaries, invented biographical claims, and incorrect information supplied by customer-service chatbots.[5] The frequency of such errors varies with the model, task, prompting method, available context, sampling procedure, and definition used, so reported hallucination rates are not directly comparable across all systems and benchmarks.[3]

Terminology and scope

The use of hallucination in AI covers several related but distinct phenomena. In contemporary LLM research, one influential taxonomy separates factuality hallucinations from faithfulness hallucinations.[3] Factuality hallucinations contradict or cannot be verified against real-world information. Faithfulness hallucinations depart from information or constraints supplied to the model, and may include inconsistency with a source document, user instruction, or the model's own preceding output.

Earlier work on abstractive summarization distinguished intrinsic hallucinations, in which generated content contradicts the source, from extrinsic hallucinations, in which generated content cannot be verified from the source.[4] An extrinsic statement is not necessarily false; it can be true but unsupported by the material being summarized. This distinction contributed to later research treating factuality and source faithfulness as related but separate evaluation problems.

The term is also used outside language generation, sometimes with substantially different meanings. In computer vision, face hallucination has historically referred to synthesizing plausible high-resolution detail from a low-resolution image.[6] In vision-language models, however, the term more often describes textual claims that are unsupported by visual input.

Hallucination is not synonymous with every kind of AI error. Reasoning mistakes, failures to follow instructions, biased outputs, software failures, errors introduced by external tools, and incorrect information faithfully repeated from a source may overlap with hallucination depending on the definition being used, but are analytically distinct.[3]

History

The terminology predates modern generative AI. In computer vision, hallucination was used for techniques that infer or synthesize image information not directly present in a low-resolution input. A 2014 survey of image super-resolution traced use of the term to a 1986 doctoral thesis by Eric Mjolsness.[7] Simon Baker and Takeo Kanade later used hallucinating faces for a super-resolution technique that synthesized facial detail.[8]

The word subsequently acquired an error-related meaning in language processing. Hallucination was used in research on machine translation for fluent output that was poorly related or unrelated to the source sentence.[9] In 2015, Andrej Karpathy used hallucinated to describe a recurrent neural network generating a nonexistent URL in synthetic text.[10]

In July 2021, Meta described its BlenderBot 2 chatbot as prone to hallucinations, which it characterized as confident statements that were not true.[11]

The term entered widespread public use following the release of general-purpose generative chatbots including ChatGPT in late 2022. News organizations increasingly used hallucination for plausible but incorrect chatbot responses.[12] In 2023, the Cambridge Dictionary added the AI-related sense of the word and selected hallucinate as its word of the year.[13]

Types

Factuality hallucination

A factuality hallucination is an output that conflicts with, or cannot be supported by, verifiable information about the world.[3] Examples include invented events, incorrect relationships between people, nonexistent organizations, false numerical information, and fabricated descriptions of publications.

Fabricated citations are a prominent subtype. An LLM may produce a reference containing plausible combinations of real authors, journal names, titles, dates, or identifiers even though the cited work does not exist. It may also cite a real publication while inaccurately describing its contents.

Faithfulness hallucination

A faithfulness hallucination occurs when an output is inconsistent with the context or constraints on which it is supposed to be based.[3] This can occur in summarization, translation, question answering over documents, or other grounded generation.

For example, a summary may add information not present in its source. Such an addition can be factually true but still unfaithful to the source. Conversely, faithfully summarizing a false statement in a source does not make the underlying statement true.

The category has also been used more broadly for instruction inconsistency and internal logical inconsistency.[3]

Citation and attribution hallucination

Models can fabricate publications, quotations, judicial decisions, statutes, URLs, or other authorities. They can also attribute real material to the wrong person or cite an authentic source for a proposition that the source does not support.

Citation hallucination is particularly consequential because conventional citation formatting can make invented material appear independently verified.

Multimodal hallucination

In multimodal large language models and vision-language systems, hallucination generally refers to generated descriptions or claims that are unsupported by the image, video, audio, or other non-text input.

This usage differs from the older computer-vision meaning of hallucination in image reconstruction and super-resolution, where synthesizing absent visual detail may be the intended task.

Causes

Hallucination does not have a single cause. Similar-looking erroneous outputs can arise from different stages of model training and deployment.

Training data

Hallucinations can result from training data that are incomplete, inaccurate, contradictory, outdated, or unrepresentative.[14] Noisy training pairs can also teach models performing conditional-generation tasks to introduce information not present in their inputs.[2]

The presence of misinformation in training data is not the only cause. Models can produce false combinations of individually familiar facts even when the underlying data contain accurate information.

Training and evaluation objectives

Large language models are initially trained to predict tokens from context. That objective models statistical regularities in language but does not independently determine the truth of every proposition that can be generated.

A 2025 analysis by Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang argued that hallucinations during pretraining can arise from ordinary statistical errors and that they persist partly because standard evaluations tend to reward guessing rather than expressions of uncertainty.[15] Under an accuracy-based evaluation in which an unanswered question receives no credit, a model that is uncertain can sometimes improve its expected score by guessing.

The authors argued that reducing this incentive requires benchmarks to reward appropriate abstention or expressions of uncertainty, rather than treating every unanswered question as equivalent to an incorrect response.[15]

Failures of grounding and retrieval

Providing a model with external information can reduce its dependence on information represented in its parameters, but it does not guarantee a correct answer. In retrieval-augmented generation, failures can result when the relevant information is not retrieved, incorrect information is retrieved, or the model fails to use retrieved evidence faithfully.[3]

Grounded systems can also combine supplied evidence with unsupported information from model parameters. Evaluation of such systems therefore commonly distinguishes retrieval quality from the faithfulness of the generated answer.

Model-specific mechanisms

Mechanistic interpretability research has begun to investigate internal mechanisms associated with hallucination. In 2025, Anthropic researchers used attribution graphs to study Claude 3.5 Haiku. They reported evidence for features associated with unfamiliar entities and declining to answer, and for other features associated with entities or answers the model appeared to recognize.[16]

In several case studies, experimentally changing these features could make the model answer questions it would otherwise decline or become uncertain about questions it would otherwise answer. The researchers also examined a hallucinated academic citation in which recognizing a familiar person's name appeared to partially suppress mechanisms associated with uncertainty.[16]

The authors cautioned that their attribution graphs only imperfectly approximate the underlying model, that the reported examples were selected case studies, and that the identified mechanisms should not be assumed to explain hallucinations generally.[16]

Prompt and context effects

A model's likelihood of producing unsupported information may also be affected by false premises, ambiguous questions, contradictory context, adversarial inputs, and prompts that imply that a particular answer must exist. Models may accept an incorrect premise and elaborate on it rather than challenge it.

Detection and evaluation

Hallucination detection methods vary according to the kind of error being measured.

For factuality, systems can divide an output into individual factual claims and compare those claims with trusted external sources. Techniques include information retrieval, knowledge bases, automated fact-checking, and model-based verification.[3]

For faithfulness, evaluation can compare generated claims directly with the source material. Proposed methods include fact-overlap measures, natural language inference classifiers, question-answering systems, uncertainty estimates, and the use of other LLMs as evaluators.[3]

A separate family of approaches attempts to estimate whether a model is uncertain about the meaning of its own possible answers. In 2024, Sebastian Farquhar and colleagues introduced semantic entropy, which groups generated answers by meaning and measures uncertainty across those semantic alternatives rather than merely across token sequences.[17] The method was intended to detect a class of plausible but arbitrary incorrect outputs the authors called confabulations.

Automated hallucination detection is itself imperfect. Systems used as evaluators can make factual or interpretive errors, and factuality cannot always be determined solely from the model's prompt or internal confidence.

Mitigation

Methods for reducing hallucination operate at the data, training, inference, and application levels.

Ji et al. grouped mitigation methods into data-related methods and modeling or inference methods. Data-related approaches include improving the quality and faithfulness of training data and augmenting inputs with external information. Modeling and inference approaches include architectural changes, changes to training objectives, reinforcement learning, and post-processing or correction of generated outputs.[2]

Grounding methods, including retrieval-augmented generation and access to databases, search engines, or other tools, can provide models with information at inference time. Their effectiveness depends on retrieval quality and on whether the generated response remains faithful to the retrieved material.[3]

Verification methods can check generated claims against external evidence, compare multiple candidate responses, validate citations, or delegate structured tasks such as calculation to deterministic software.

Uncertainty and abstention methods attempt to make models decline to answer when evidence is inadequate. Research on evaluation incentives has argued that benchmark design should reward appropriate uncertainty rather than systematically encouraging models to guess.[15]

Researchers have also investigated neuro-symbolic architectures combining neural language models with formal knowledge bases or inference systems.[18]

Human verification remains important in applications where erroneous output can have legal, medical, scientific, financial, or other consequential effects. Human review is effective only if the underlying claims and sources are independently checked rather than accepted because the generated text appears plausible.

No mitigation technique eliminates all hallucinations, and methods that reduce one form of error may not address others.

Documented consequences

Generative AI has produced nonexistent cases, quotations, and statutory authorities in legal filings in multiple jurisdictions.

A prominent early incident occurred in Mata v. Avianca, Inc. in 2023. Attorney Stephen Schwartz used ChatGPT while researching a personal-injury case and submitted judicial authorities that did not exist. After the fabricated cases were discovered, the United States District Court for the Southern District of New York imposed a $5,000 sanction on Schwartz and another attorney for bad-faith conduct.[19]

Subsequent incidents involved both private lawyers and government attorneys. In 2026, a Georgia assistant district attorney acknowledged failing to verify AI-assisted legal research containing fictitious or inapposite authorities; the Supreme Court of Georgia vacated the resulting order and suspended her from practicing before the court for six months.[20]

In April 2026, Sullivan & Cromwell apologized to a federal bankruptcy court after hallucinated legal citations appeared in a filing submitted on behalf of a client.[21]

A database maintained by legal researcher Damien Charlotin tracks judicial decisions and filings involving suspected AI-generated fictitious citations and related errors.[22] Because the database records identified incidents rather than a representative sample of all legal filings, its totals do not establish the prevalence of hallucinations among AI-assisted legal work generally.

Similar cases have occurred outside the United States. Canadian courts and tribunals had identified numerous filings containing fictitious authorities by 2026,[23] while the High Court of Justice in England and Wales addressed hallucinated authorities in the conjoined 2025 cases Ayinde v Haringey and Al-Haroun v Qatar.[24]

Professional bodies including the Solicitors Regulation Authority, the Law Society of England and Wales, and the Bar Council have subsequently issued guidance concerning lawyers' use of generative AI.[25][26][27]

Scientific and academic research

Fabricated or inaccurate scholarly references are a documented failure mode of LLMs.

A 2023 study examining 178 references produced for proposed scientific writing found numerous incorrect or nonexistent digital object identifiers (DOIs), including 28 references for which the researchers could not locate the cited work or a valid DOI.[28]

Another 2023 study examined 115 references generated by ChatGPT-3.5 for medical writing. The authors classified 47% as fabricated, 46% as authentic references containing inaccurate information, and 7% as authentic and accurately represented.[29] The figures describe that particular model and experimental setup and are not a general error rate for later LLMs.

Research librarians have also reported student-submitted citations containing combinations of real and fabricated bibliographic information. A 2024 study at the University of Mississippi examined AI-associated citations found in student work and documented incorrect authors, titles, dates, and nonexistent publications.[30]

In 2025, Deloitte acknowledged AI use after nonexistent academic sources and an inaccurate judicial quotation were identified in an Australian government report; the company produced a revised report and agreed to refund part of its fee.[31]

Customer-facing systems

In Moffatt v. Air Canada, a customer relied on information from an Air Canada website chatbot stating that a bereavement fare could be requested retroactively. The airline's actual policy did not permit the discount after travel had been booked. In 2024, the British Columbia Civil Resolution Tribunal held Air Canada responsible for the erroneous information and awarded damages.[32]

The tribunal rejected Air Canada's argument that the chatbot was effectively a separate entity responsible for its own statements. The case became a frequently cited example of the distinction between the technical source of an AI error and the responsibility of an organization deploying an automated system.

Other modalities and uses of the term

Computer vision

The historical computer-vision meaning of hallucination differs from its contemporary use for factual errors in language models. In super-resolution, hallucination can mean generating plausible visual detail not recoverable directly from a low-resolution image.[6]

The word has also been used for false-positive detections by image-recognition systems and for failures induced by adversarial examples. Wired reported examples including computer-vision systems failing to recognize an adversarially altered stop sign and a Google Cloud Vision system assigning a high probability to an incorrect object category.[33]

Researchers have disputed whether some adversarial examples are best interpreted as hallucinations, noting that models may respond to statistical features that are genuinely predictive in their training distributions but imperceptible or irrelevant to humans.[34]

Generative images and video

A video of the Glenfinnan Viaduct generated by Sora, containing physical and historical inconsistencies including an additional track, right-running trains, inconsistent carriage geometry, and an altered depiction of The Jacobite.
The real Glenfinnan Viaduct with The Jacobite.

The term hallucination is sometimes extended to generative image and video systems when generated content conflicts with a prompt, reference image, physical constraints, or known facts. The boundaries of this usage are less standardized than in LLM research: an unexpected generated feature may alternatively be described as an artifact, generation error, or failure of prompt adherence.

Generative systems used for image analysis can also hallucinate identifications or historical matches. Kurt Luther described cases in which generative AI used for archival photo research incorrectly identified people or fabricated historical matches.[35]

Generative scientific design

The word hallucination has also occasionally been used positively for the generation of novel scientific structures or hypotheses. For example, reporting on David Baker's protein-design research has described computational generation of previously nonexistent protein structures as a form of productive hallucination.[36]

This usage is conceptually different from an LLM presenting a false factual proposition. In generative scientific design, novelty may be the intended output and candidate designs can subsequently be tested against physical or experimental constraints. The Nobel Committee for Chemistry, in describing Baker's work associated with the 2024 Nobel Prize in Chemistry, used the phrase "computational protein design" and described the creation of new proteins rather than treating the work as factual error.

Criticism of the term

The metaphor of hallucination has been criticized for anthropomorphizing AI systems and for grouping distinct kinds of error under language borrowed from human perception.

A 2024 paper by Negar Maleki, Balaji Padmanabhan and Kaushik Dutta argued that the term is inconsistently defined and can mislead discussion of AI failure modes.[37] Computer scientist Mary Shaw has similarly criticized the terminology for anthropomorphizing software and characterizing objective errors as though they were perceptual quirks.[38]

Murray Shanahan has argued more broadly that anthropomorphic descriptions of language models can encourage inappropriate inferences about mental processes in systems based on statistical computation.[39]

Michael Townsen Hicks, James Humphries and Joe Slater argued that LLM outputs are better characterized using Harry Frankfurt's concept of bullshit, because in their analysis the generation process is not inherently oriented toward the truth of the resulting assertions.[40]

Other proposed terminology includes confabulation, fabrication, factual error, mirage, and undesirable output.[37][41] These terms emphasize different properties and are not exact synonyms.

Susanne Förster and Yarden Skop have argued that use of the hallucination metaphor in industry communication can shift attention from design and deployment decisions toward apparently autonomous characteristics of a model.[42]

Despite these criticisms, hallucination remains widely used in AI research, industry, journalism, and public discussion. Technical literature increasingly defines the specific error being measured—such as factual contradiction, source unfaithfulness, or visual grounding failure—rather than relying on the umbrella term alone.[3]

See also

References

  1. ↑ "Joaquín Correa fue presentado en Botafogo para jugar el Mundial de Clubes y tuvo un insólito cruce con un periodista: 'No es mi hermano'". Clarín. 14 June 2025. Archived from the original on 2 July 2025. Retrieved 10 August 2025.
  2. 1 2 3 Ji, Ziwei; Lee, Nayeon; Frieske, Rita; Yu, Tiezheng; Su, Dan; Xu, Yan; Ishii, Etsuko; Bang, Ye Jin; Madotto, Andrea; Fung, Pascale (31 December 2023). "Survey of Hallucination in Natural Language Generation". ACM Computing Surveys. 55 (12): 1–38. arXiv:2202.03629. doi:10.1145/3571730.
  3. 1 2 3 4 5 6 7 8 9 10 11 12 13 Huang, Lei; Yu, Weijiang; Ma, Weitao; Zhong, Weihong; et al. (24 January 2025). "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions". ACM Transactions on Information Systems. 43 (2) 42: 1–55. doi:10.1145/3703155.
  4. 1 2 Maynez, Joshua; Narayan, Shashi; Bohnet, Bernd; McDonald, Ryan (2020). "On Faithfulness and Factuality in Abstractive Summarization". Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 1906–1919. doi:10.18653/v1/2020.acl-main.173.
  5. ↑ Metz, Cade (6 November 2023). "Chatbots May 'Hallucinate' More Often Than Many Realize". The New York Times. Archived from the original on 7 December 2023. Retrieved 6 November 2023.
  6. 1 2 Liu, Ce; Shum, Heung-Yeung; Freeman, William T. (18 July 2007). "Face Hallucination: Theory and Practice". International Journal of Computer Vision. 75 (1): 115–134. Bibcode:2007IJCV...75..115L. doi:10.1007/s11263-006-0029-5.
  7. ↑ Nasrollahi, Kamal; Moeslund, Thomas B. (2014). "Super-resolution: a comprehensive survey". Machine Vision and Applications. 25 (6): 1423–1468. doi:10.1007/s00138-014-0623-4. ISSN 0932-8092.
  8. ↑ "Hallucinating Faces". Robotics Institute Carnegie Mellon University. 1999. Archived from the original on 7 January 2026. Retrieved 29 November 2025.
  9. ↑ "Hallucinations in Neural Machine Translation". Google Research. Archived from the original on 2 April 2024. Retrieved 2 April 2024.
  10. ↑ Karpathy, Andrej (21 May 2015). "The Unreasonable Effectiveness of Recurrent Neural Networks". Andrej Karpathy. Archived from the original on 1 November 2020. Retrieved 24 December 2025.
  11. ↑ "Blender Bot 2.0: An open source chatbot that builds long-term memory and searches the internet". Meta AI. Archived from the original on 23 February 2024. Retrieved 2 March 2024.
  12. ↑ Weise, Karen; Metz, Cade (1 May 2023). "When A.I. Chatbots Hallucinate". The New York Times. Archived from the original on 4 April 2024. Retrieved 8 May 2023.
  13. ↑ Creamer, Ella (15 November 2023). "'Hallucinate' chosen as Cambridge dictionary's word of the year". The Guardian. Retrieved 7 June 2024.
  14. ↑ Aaronson, Susan Ariel (2024). Introduction: What Hath Generative Artificial Intelligence Wrought? (Report). Centre for International Governance Innovation. pp. 1–4. Archived from the original on 22 November 2025. Retrieved 26 October 2025.
  15. 1 2 3 Kalai, Adam Tauman; Nachum, Ofir; Vempala, Santosh S.; Zhang, Edwin (4 September 2025). "Why Language Models Hallucinate". arXiv:2509.04664 [cs.CL].
  16. 1 2 3 Lindsey, Jack; Gurnee, Wes; Ameisen, Emmanuel; et al. (27 March 2025). "On the Biology of a Large Language Model". Transformer Circuits Thread. Anthropic. Retrieved 18 September 2026.
  17. ↑ Farquhar, Sebastian; Kossen, Jannik; Kuhn, Lorenz; Gal, Yarin (19 June 2024). "Detecting hallucinations in large language models using semantic entropy". Nature. 630: 625–630. doi:10.1038/s41586-024-07421-0.
  18. ↑ Garcez, Artur (30 May 2025). Vass, Steven (ed.). "Neurosymbolic AI is the answer to large language models' inability to stop hallucinating". The Conversation. doi:10.64628/AB.5gpku36ct.
  19. ↑ Brodkin, Jon (23 June 2023). "Lawyers have real bad day in court after citing fake cases made up by ChatGPT". Ars Technica. Archived from the original on 26 January 2024. Retrieved 26 June 2023.
  20. ↑ Scarcella, Mike (5 May 2026). "AI errors in US murder case lead to discipline for Georgia prosecutor". Reuters. Retrieved 20 May 2026.
  21. ↑ Fairfield, Karen; Scarcella, Mike (21 April 2026). "Sullivan & Cromwell law firm apologizes for AI 'hallucinations' in court filing". Reuters. Retrieved 21 April 2026.
  22. ↑ "AI Hallucination Cases". damiencharlotin.com. Damien Charlotin. Retrieved 18 September 2026.
  23. ↑ Zheng, Tom Macintosh (14 April 2026). "The Rise of AI-Hallucinated Case Law in Canadian Courts and Tribunals". Canadian Legal Information Institute.
  24. ↑ Hyde, John. "Judge refers solicitors to SRA over 'inexcusable' submission of AI-generated authorities". Law Gazette.
  25. ↑ "Misuse of AI - Warning notice". Solicitors Regulation Authority. 17 August 2026. Retrieved 26 August 2026.
  26. ↑ Leslie, Mia. "How AI tools hallucinate – and why it matters in law". Law Society of England and Wales. Retrieved 26 August 2026.
  27. ↑ "New guidance on generative AI for the Bar". Bar Council. Retrieved 26 August 2026.
  28. ↑ Athaluri, Sai Anirudh; Manthena, Sandeep Varma; Kesapragada, V. S. R. Krishna Manoj; Yarlagadda, Vineel; Dave, Tirth; Duddumpudi, Rama Tulasi Siri (11 April 2023). "Exploring the Boundaries of Reality: Investigating the Phenomenon of Artificial Intelligence Hallucination in Scientific Writing Through ChatGPT References". Cureus. 15 (4) e37432. doi:10.7759/cureus.37432. PMC 10173677. PMID 37182055.
  29. ↑ Bhattacharyya, Mehul; Miller, Valerie M.; Bhattacharyya, Debjani; Miller, Larry E. (19 May 2023). "High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content". Cureus. 15 (5) e39238. doi:10.7759/cureus.39238. PMC 10277170. PMID 37337480.
  30. ↑ Watson, Alex P. (3 July 2024). "Hallucinated Citation Analysis: Delving into Student-Submitted AI-Generated Sources at the University of Mississippi". The Serials Librarian. 85 (5–6): 172–180. doi:10.1080/0361526X.2024.2433640. ISSN 0361-526X.
  31. ↑ Tadros, Edmund; Karp, Paul (5 October 2025). "Deloitte to refund government, admits using AI in $440k report". Australian Financial Review. Archived from the original on 11 October 2025. Retrieved 12 October 2025.
  32. ↑ Belanger, Ashley (16 February 2024). "Air Canada must honor refund policy invented by airline's chatbot". Ars Technica. Archived from the original on 22 April 2025. Retrieved 22 April 2025.
  33. ↑ Simonite, Tom (9 March 2018). "AI Has a Hallucination Problem That's Proving Tough to Fix". Wired. Condé Nast. Archived from the original on 5 April 2023. Retrieved 29 December 2022.
  34. ↑ Gilmer, Justin; Hendrycks, Dan (6 August 2019). "A Discussion of 'Adversarial Examples Are Not Bugs, They Are Features': Adversarial Example Researchers Need to Expand What is Meant by 'Robustness'". Distill. 4 (8). doi:10.23915/distill.00019.1.
  35. ↑ Luther, Kurt (2025). "A Guide to Exploring Photo Sleuthing and Generative AI". Military Images. 43 (4): 8–11. ISSN 1040-4961. JSTOR 27404321. Archived from the original on 29 January 2026. Retrieved 23 November 2025.
  36. ↑ Broad, William J. (23 December 2024). "How Hallucinatory A.I. Helps Science Dream Up Big Breakthroughs". The New York Times. Archived from the original on 24 December 2024. Retrieved 24 December 2024.
  37. 1 2 Maleki, Negar; Padmanabhan, Balaji; Dutta, Kaushik (2024). "AI Hallucinations: A Misnomer Worth Clarifying". 2024 IEEE Conference on Artificial Intelligence (CAI). pp. 133–138. arXiv:2401.06796. doi:10.1109/CAI59869.2024.00033. ISBN 979-8-3503-5409-6.
  38. ↑ Kang, Eunsuk; Shaw, Mary (2024). "tl;dr: Chill, y'all: AI Will Not Devour SE". Proceedings of the 2024 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software. pp. 303–315. arXiv:2409.00764. doi:10.1145/3689492.3689816. ISBN 979-8-4007-1215-9.
  39. ↑ Shanahan, Murray (2024). "Talking about Large Language Models". Communications of the ACM. 67 (2): 68–79. doi:10.1145/3624724.
  40. ↑ Hicks, Michael Townsen; Humphries, James; Slater, Joe (June 2024). "ChatGPT is bullshit" (PDF). Ethics and Information Technology. 26 (2) 38. doi:10.1007/s10676-024-09775-5. Archived (PDF) from the original on 13 May 2025. Retrieved 23 May 2025.
  41. ↑ Mills, Anna; Angell, Nate (2025). "Are We Tripping? The Mirage of AI Hallucinations". SSRN 5127162.
  42. ↑ Förster, Susanne; Skop, Yarden (2025). "Between fact and fairy: tracing the hallucination metaphor in AI discourse". AI & Society. doi:10.1007/s00146-025-02392-w.

Notes

This article is a direct transclusion of the Wikipedia article and therefore may not meet the same editing standards as LIMSwiki.

Retrieved from "https://www.limswiki.org/index.php?title=Hallucination_(artificial_intelligence)&oldid=62746"