An AI hallucination is when an AI system states something false as if it were true: a figure that isn't in the data, a clause that isn't in the contract, a source that doesn't exist. A national standards body's generative AI profile prefers the term confabulation, which it describes as "a phenomenon in which GAI systems generate and confidently present erroneous or false content in response to prompts".
Why it happens
The same profile says "Confabulations are a natural result of the way generative models are designed". It explains that such models "generate outputs that approximate the statistical distribution of their training data; for example, LLMs predict the next token or word in a sentence or phrase". When the right answer isn't in front of the model, a plausible wrong one can come out.
That is what makes hallucinations hard to spot. They look like answers. A 2026 benchmark of AI reading handwritten documents found that "63-91% of model errors are prior-driven versus only 49% for humans". People made more cautious partial readings. Overall, though, the gap between the best model and people was narrow: 71.85 against 77.09 on the benchmark's composite score.
Open-ended and grounded use
Asking a model a question from memory invites confabulation. Asking it to answer from a document or dataset it has been given lowers the risk. In our experience it doesn't remove it, which is why the checks below still matter.
In a timber business
A plausible wrong volume passes a glance. So does a plausible wrong contract term, which can reach a customer.
How to guard against it
In our view, the dependable checks sit outside the model. Every figure should be traceable to the record it came from, and totals should be calculated by tools that calculate exactly. Answers about documents should quote the passage they rely on.
What it isn't
A hallucination isn't a lie, since the model has no intent. Nor is it the same as an out-of-date answer, which reflects old data rather than invention.
How Quarri works explains the platform as a layer over existing systems, not a migration.
Sources
- NIST, AI 600-1, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", July 2024: nvlpubs.nist.gov
- Zhang and others, "WildHandBench", arXiv 2608.22959, 24 August 2026: arxiv.org
Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.