In June 2023, a federal judge sanctioned two lawyers for filing a brief full of case citations that did not exist. A chatbot had invented them: six confident, correctly formatted, entirely fictional cases, complete with docket numbers and quotable passages. When one lawyer asked whether the cases were real, it assured him they were, citing "reputable legal databases such as LexisNexis and Westlaw."[1]
That's a hallucination: output that is fluent, plausible, delivered with total confidence, and wrong, a structural consequence of what a language model is. It is trained to continue text plausibly, not to report facts truthfully, and that distinction drives everything in this explorable.
Why this matters now
Language models draft emails, summarize meetings, answer medical questions, and write code. Their failures are uniquely treacherous because they fail fluently: a search engine that finds nothing shows you nothing; a language model that knows nothing shows you a well-written paragraph. And they fail with confidence: the invented answer arrives in the same calm, authoritative voice as the correct one, with no hesitation or hedge to signal that anything is wrong.
How to read this explorable
The pages are sequential, moving from how models work, to why they confabulate, to what to do about it. Jump to any part:
- Part IFoundationsWhat language models actually do, and why that matters for truth.
- Part IIAnatomy of a HallucinationWhat hallucinations are, what they look like, and how to classify them.
- Part IIIWhy Models HallucinateThe statistical, architectural, and incentive-driven roots of confabulation.
- Part IVMeasuring & DetectingBenchmarks, uncertainty, and catching hallucinations in the act.
- Part VMitigation & EngineeringGrounding, retrieval, abstention, and building systems that fail safely.
- Part VILiving With Confident MachinesPractical judgment for users, builders, and society.
Most pages contain an interactive widget: click, drag, or run it, and watch what actually changes. A few use a static figure or table instead, where a diagram makes the point more clearly than something clickable would.