Hallucination (artificial intelligence)

• 1. Etymology And Semantic Shifts

  ›

  In 1986, Eric Mjolsness used the word hallucination in his PhD thesis to describe a computer vision process that added detail to an image. This early usage carried a positive meaning about enhancing visual data rather than creating errors. By 1995, Stephen Thaler demonstrated how artificial neural networks could produce phantom experiences through random changes to their connection weights. The field of statistical machine translation began describing these failures as hallucinations during the 2000s. A semantic shift occurred in the 2010s when researchers started using the term for factually incorrect outputs generated by AI systems. Saurabh Gupta and Jitendra Malik identified hallucinations in visual semantic role labeling tasks in 2015. Andrej Karpathy wrote a blog post in 2015 describing his recurrent neural network language model generating an incorrect citation link. Google researchers applied the term to neural machine translation models in 2017 when they produced responses unrelated to source text. Computer vision experts used the word again in 2018 to describe instances where non-existent objects were detected due to adversarial attacks. Meta warned users about hallucinations in July 2021 when releasing BlenderBot 2, defining them as confident statements that are not true. OpenAI released ChatGPT in beta version on the 30th of November 2022, leading many news outlets to adopt the term for frequently incorrect or inconsistent responses. The Cambridge dictionary updated its definition of hallucination in 2023 to include this new sense specific to artificial intelligence.

  10121539

• 2. Technical Definitions And Classifications

  ›

  OpenAI defined hallucinations in May 2023 as a tendency to invent facts in moments of uncertainty. CNBC described the phenomenon in May 2023 as fabricating information entirely while behaving as if spouting facts. The Verge stated in February 2023 that models simply make up information without grounding it in reality. Researchers categorize hallucinations based on whether output contradicts the source or cannot be verified from the source. Intrinsic hallucinations occur when generated content appears factual but is ungrounded relative to provided input data. Extrinsic hallucinations arise when outputs contradict the prompt itself rather than external sources. Closed-domain systems restrict these errors to specific topics while open-domain models face broader challenges. Amabile and Pratt define human creativity as producing novel and useful ideas, which helps explain why machine creativity can lead to original but inaccurate responses. GPT-3 generates each next word based on a sequence of previous words including those it has already generated during the same conversation. This process causes a cascade of possible hallucinations as the response grows longer. Decoders can attend to the wrong part of encoded input leading to erroneous generation. Top-k sampling improves generation diversity but correlates positively with increased hallucination rates. Anthropic identified internal circuits in Claude in 2025 that cause the model to decline answering questions unless it knows the answer.

  68343545

• 3. Scientific Controversy And Criticism

  ›

  Usama Fayyad criticized the term hallucination for misleadingly personifying large language models and being vague. Mary Shaw called the current fashion for calling generative AI's errors hallucinations appalling because it anthropomorphizes software. Gary N. Smith argued in Salon that LLMs do not understand what words mean so the term unreasonably anthropomorphizes machines. Some researchers view AI outputs not as illusory but as prospective having some chance of being true similar to early-stage scientific conjectures. The term also faces criticism for its association with psychedelic drug experiences. Benj Edwards wrote in Ars Technica that while controversial, some form of metaphor remains necessary suggesting confabulation as an analogy for creative gap-filling processes. Hicks, Humphries, and Slater published an article in Ethics and Information Technology arguing that LLM output is bullshit under Harry Frankfurt's definition. They claimed models are indifferent to truth with true statements only accidentally true and false ones accidentally false. A White House report on fostering public trust in AI research mentioned hallucinations only in context of reducing them. The Nobel committee avoided using the term entirely when acknowledging David Baker's work with AI-generated proteins instead referring to imaginative protein creation.

  2733364594

• 4. Legal And Ethical Consequences

  ›

  On the 15th of November 2022, Meta AI published Galactica designed to store combine and reason about scientific knowledge. One case involved Galactica citing a fictitious paper from a real author who works in the relevant area before Meta withdrew it on the 17th of November due to offensiveness and inaccuracy. Stephen Schwartz submitted six fake case precedents generated by ChatGPT in his brief to the Southern District of New York on the 15th of May 2023. Judge P. Kevin Castel dismissed the Mata case on June 23 and issued a $5,000 fine to Schwartz and another lawyer for bad faith conduct. Castel described one cited opinion as gibberish bordering on nonsensical. Mark Walters sued OpenAI in Georgia state court in June 2023 after ChatGPT mischaracterized a legal complaint in a manner alleged to be defamatory against him. Judge Tracie Cason ruled in favor of OpenAI in May 2025 stating the plaintiff had not shown he was defamed. Air Canada was ordered by the Civil Resolution Tribunal in February 2024 to pay damages honoring a bereavement fare policy hallucinated by a support chatbot. The tribunal rejected Air Canada's defense that the chatbot was a separate legal entity responsible for its own actions. Deloitte submitted an A$440,000 report to the Australian government in July containing non-existent academic sources and a fake quote from a federal court judgement.

  6061626364

• 5. Paradoxical Scientific Utility

  ›

  David Baker's lab at the University of Washington used AI hallucinations to design ten million brand-new proteins that do not occur in nature. This work led to roughly 100 patents and the founding of over 20 biotech companies contributing to Baker receiving the 2024 Nobel Prize in Chemistry. Researchers at California Institute of Technology used hallucinations to design a novel catheter geometry that significantly reduces bacterial contamination. The design features sawtooth-like spikes on inner walls preventing bacteria from gaining traction potentially addressing millions of urinary tract infections annually. Anima Anandkumar emphasized these models are taught physics and outputs must be validated through rigorous testing. Scientists use AI to generate thousands of subtle forecast variations helping identify unexpected factors influencing extreme weather events. Memorial Sloan Kettering Cancer Center applied hallucinatory techniques to enhance blurry medical images while University of Texas at Austin utilized them to improve robot navigation systems. These applications demonstrate how hallucinations when properly constrained by scientific methodology can accelerate discovery from years to days or even minutes. Luther described instances in 2025 where generative AI tools incorrectly identified individuals or fabricated historical matches when analyzing archival military images.

  2646878891

• 6. Educational Integrity Challenges

  ›

  A 2024 study at the University of Mississippi found many student-submitted citations were partially or completely fabricated. Forty-seven percent of these sources had incorrect titles dates authors or combinations of all three errors. Zoë Teel noted in a 2023 paper that universities may need to implement their own citation auditing to track fictitious references. The Journal of Cranio-Maxillofacial Surgery mentioned academic publishers have acknowledged the issue with some journals like JAMA changing policies to discourage AI-generated citations. Turnitin is used for plagiarism checking but sometimes flags papers without any AI assistance. OpenAI shut down its own AI detection software due to lack of accuracy. A group of researchers at Northwestern University generated 50 abstracts based on existing reports and analyzed originality. Plagiarism detectors gave the generated articles an originality score of 100% meaning information appeared completely original. Research scientists identified these abstracts at a rate of 68% showing similar error rates to human judgment. Some say rather than hallucinations these events are more akin to fabrications and falsifications presenting risks to field integrity.

  96979899100

• 7. Mitigation Strategies And Future Research

  ›

  Researchers from OpenAI wrote that hallucinations occur because training and evaluation reward guessing over acknowledging uncertainty. Ji et al divided common mitigation methods into data-related approaches and modeling inference techniques. Data-related methods include building faithful datasets cleaning data automatically and augmenting inputs with external information. Model and inference methods involve changes in architecture using reinforcement learning or post-processing corrections. Researchers proposed getting different chatbots to debate one another until reaching consensus on answers. Neuro-symbolic architectures reason in formal logic while other approaches actively validate correctness using web search results. Nvidia Guardrails launched in 2023 can be configured to hard-code certain responses via script instead of leaving them to the LLM. Tools like SelfCheckGPT Trustworthy Language Model and Aimon emerged to aid detection in offline experimentation and real-time production scenarios. Evaluating multiple possible replies before answering queries by assigning confidence scores could mitigate problems but multiplies computational costs. Active learning would further increase these costs making it economically viable only for high-stakes domains like chip design supply chain logistics and medical diagnostics.

  3436454859