- AI hallucinations produce plausible but false content from statistical patterns.
- Hallucination rates dropped 96% since 2021 but persist in specialized domains.
- OpenAI research shows hallucinations are mathematically inevitable under current training.
An AI hallucination occurs when a language model generates content that sounds plausible but is factually wrong, fabricated, or unsupported by its training data. The model does not perceive or imagine anything. It produces statistically likely sequences of words that happen to be false.
Why It Matters
Key figure
96%
reduction in hallucination rates, 2021-2025
The reliability of every AI-assisted task depends on whether the output is true. In science, law, medicine, and journalism, a confident-sounding falsehood can pass unchallenged if the reader trusts the source.
AI hallucinations exploit that trust by producing text that reads like verified information but isn't.
The scale of the problem has shifted since ChatGPT's launch in late 2022. On summarization benchmarks tracked by the Vectara Hallucination Leaderboard, best-performing large language models went from a 21.8% hallucination rate in 2021 to 0.7% in 2025.
That 96% reduction sounds reassuring until you check domain-specific numbers: on certain legal queries, hallucination rates still reach 69% to 88%.
In 2023, New York attorneys Steven Schwartz and Peter LoDuca were sanctioned after submitting AI-hallucinated legal citations in Mata v. Avianca. In 2025, Deloitte submitted government consulting reports in Australia and Canada containing fabricated academic sources.
These are not edge cases. They are predictable consequences of a system that generates plausible text without verifying whether it is true.
OpenAI's own researchers published a paper showing that hallucinations are mathematically inevitable under standard training procedures that reward guessing over admitting uncertainty. The problem is structural, not a bug waiting to be fixed.
How It Works
Large language models generate text by predicting the next most probable token given everything that came before. They do not retrieve facts from a database. They do not understand what they are saying.
They produce sequences that match patterns learned during training.
When the training data contains gaps, contradictions, or insufficient coverage, the model fills the space with plausible-sounding completions. This is closer to confabulation than to hallucination in its psychiatric sense.
Confabulation is the neurological term for gap-filling fabrication. Hallucination involves false sensory perception. The AI system has no perceptions to distort.
The term itself predates the current debate. Eric Mjolsness used "hallucination" in his 1985 thesis on neural networks and pattern recognition.
Baker and Kanade's 2000 paper "Hallucinating Faces" applied it to computer vision models generating plausible facial details not present in the input. The word migrated to natural language processing after 2018, when Google researchers used it for machine translation errors.
Several mitigation strategies now exist. Retrieval-augmented generation (RAG) grounds model responses in retrieved documents. Chain-of-thought prompting forces models to show reasoning steps.
Constitutional AI applies rule-based constraints. None eliminates hallucination entirely, but each reduces its frequency in specific contexts.
Key Context
MIT researchers found in January 2025 that hallucinating models use 34% more confident language than when they produce correct answers. The pattern is counterintuitive: the model sounds most authoritative precisely when it is most wrong.
When AI generates false content that is then indexed and cited by other AI systems, the errors compound. The downstream effect is a feedback loop of fabricated information circulating without human verification.
That problem grows as AI-generated text becomes a larger share of the web.
FAQ
Can AI hallucinations be completely eliminated?
No. OpenAI's own research demonstrates that hallucinations are mathematically inevitable under current training and evaluation methods that penalize silence and reward guessing. Reducing their frequency is possible; eliminating them entirely would require a fundamentally different architecture.
What is the difference between AI hallucination and confabulation?
"Hallucination" borrows from psychiatry, where it means a false sensory perception. "Confabulation" comes from neurology and refers to filling memory gaps with fabricated content. Since language models have no perceptions to distort but do fill gaps in learned patterns, confabulation is the more precise analogy. The AI research community continues to use "hallucination" because it became established first. Cambridge Dictionary named "hallucinate" its word of the year for 2023.
How can you tell if an AI response is hallucinated?
There is no reliable internal signal. MIT's 2025 research showed that models use more confident language when hallucinating, not less. The only dependable check is external verification: compare the AI's claims against primary sources, check that cited papers and statistics actually exist, and treat any unsourced factual claim as unverified.
Why do AI models sound more confident when they are wrong?
Models are trained on text where confident language correlates with authoritative sources. When generating in low-certainty territory, the model defaults to the patterns associated with authority. The result is that statistical uncertainty in the model's predictions produces linguistically certain outputs.
Related Reading
Sources
- What Are AI Hallucinations? (IBM)
- Addressing AI Hallucinations and Bias (MIT Sloan)
- Beware of AI Hallucinations or Should We Call Confabulation? (PMC/NIH)
- Why Language Models Hallucinate (OpenAI)
- Hallucination Leaderboard (Vectara)
- Why 'Hallucination'? (LA Review of Books)
Fact Check: Claim-by-Claim Verification Verified
All core claims verified. Two factual corrections applied during editorial check: Mata v. Avianca date (2024→2023) and Google DeepMind attribution corrected to Google.
OpenAI primary source.
Editor note: "Google DeepMind" corrected to "Google" — the 2018 paper was by Google researchers, not specifically the DeepMind division.
Sources used for verification
Academic/Peer-reviewed:
Other reliable sources:
- What Are AI Hallucinations? - IBM
- Why Language Models Hallucinate - OpenAI
- Hallucination Leaderboard - Vectara
- Why 'Hallucination'? - LA Review of Books
- Addressing AI Hallucinations and Bias - MIT Sloan
