— Ch. 1 · Etymology And Semantic Shifts —
Hallucination (artificial intelligence).
~7 min read · Ch. 1 of 7
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.
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.