AI concepts
LLM hallucinations
A language model producing answers that are formally correct in syntax but factually untrue. A hallucination is not an implementation bug or a defect, it is an architectural trait of probabilistic models. The hallucination rate in production enterprise applications ranges from 3 to 27 percent depending on the type of query and the benchmark.
Primary source: Stanford CRFM HELM, Vectara Hallucination Leaderboard 2025
The term is linguistically misleading: the model does not hallucinate in a psychological sense, it generates the statistically most probable sequence of tokens for a given prompt. If the truth is less probable than a plausible fiction, the model chooses the fiction. This applies to every LLM, not only OpenAI.
Measurement
The Vectara Hallucination Leaderboard, updated quarterly, measures the hallucination rate in a summarisation task. As of Q1 2026:
- GPT-4o: 1.5 percent
- Claude 3.7 Sonnet: 0.8 percent
- Gemini 2.5 Pro: 2.4 percent
- Llama 3.3 70B: 4.1 percent
These figures apply to a controlled task with provided context. In open tasks (Q&A without RAG) the rates are five to ten times higher.
Types of hallucination
Stanford CRFM distinguishes three classes. First, factual errors: the model states a specific piece of data (a year, a name, an amount) and gets it wrong. Second, source fabrication: the model cites publications that do not exist. Third, internal inconsistency: within a single answer the model contradicts itself.
In enterprise applications the second class is the most costly. A legal client receives case law from a chatbot that is not in the database. A financial client receives a quote from a KNF report that was never published.
Mitigants
Four technical approaches used in production. RAG (retrieval-augmented generation) with verified sources lowers the hallucination rate to 0.5 to 3 percent for informational tasks. Function calling with external validators (for example, SQL through a database instead of generating numbers) eliminates the factual class. Constrained decoding restricts the output to an allowed grammar. Multi-model verification, in which a second model verifies the output of the first, lowers the rate by around 40 percent at the cost of latency.
A structural limitation
Hallucinations cannot be eliminated entirely without changing the model's architecture. This means that any use of AI in high-stakes decision processes (medicine, law, finance) requires a human-in-the-loop. Not as a safety net just in case, but as an architectural element without which the solution is not ready to deploy.
This topic is part of the risk analysis in the AI Readiness Audit.