Hallucination
Also: confabulation, AI hallucination
Hallucination is when an AI model states something confidently that isn't true. It might invent a citation, get a date wrong, describe a product feature that doesn't exist, or fill in gaps in its knowledge with plausible-sounding guesses. It happens because language models are trained to produce fluent text — not to verify facts before speaking. Knowing when to trust Claude's output versus when to ground it in real sources is one of the most important skills for operators.
In practice
You ask Claude who won a local election last month and it confidently names a candidate — who lost. Claude isn't lying; it generated a plausible-sounding answer from patterns in its training data without actually knowing the answer. That's hallucination. It's most dangerous when the output sounds authoritative and you don't think to check it.
Related concepts
Where Hallucination shows up
9 articlesThe errors you will definitely hit, the ones that will surprise you, and the patterns that make your app resilient when Claude or the API behaves unexpectedly.
Hallucination is Claude stating something false with confidence. You can't remove it, but you can cut it sharply: give Claude the source, let it say 'I don't know', turn on web search for anything recent, ask for citations, and check the facts that matter.
Claude is genuinely powerful. It is also genuinely wrong for certain kinds of work. Knowing the difference is as important as knowing what it does well.
Hallucination isn't a bug that gets patched. It's a structural feature of how language models work. Understanding why it happens is the first step to building applications that aren't derailed by it.
Everyone knows AI can make things up. What surprises people is which specific situations trigger it — and how confident Claude sounds when it does.
The wrong AI conversation with a client creates a problem you'll spend months managing. The right one positions you as the person who gets it before everyone else does. Here's the script for both scenarios.
Claude now searches the web, remembers across sessions, reads your files through connectors and runs code. Here is what it still genuinely cannot do, and why the remaining limits are about judgment, not features.
Claude isn't being random — inconsistency almost always has a specific cause you can find and fix. Memory closed part of the gap; here are the four causes that remain, in order of how often they appear.
Most teams who roll out Claude see strong early results and a quiet decline by month 4. It's not that Claude stopped working — it's that the rollout stopped. Here's what actually happened, and what to do about it.