Term: Hallucination
~5 min read
Estimated time: ~5 min read — for the in-app brief plus opening the primary source.
What this is
A hallucination is a confident but false or fabricated output — invented citations, numbers, product claims, or case names.
Everyday example
Claude, ChatGPT, or Grok invents a court case name or a statistic that sounds real. The sentence is fluent; the fact is false.
A hallucination is a fluent, plausible invention — a citation, clause, or number that is not true.
- It is a known failure mode of generative models, not a rare bug.
- Confidence in the tone is not evidence.
- Grounding and verification cut the risk; they do not remove it.
- Deloitte’s 2025 government-report refund is the board-level example.
Next action: On the next AI-drafted brief, check every citation and number before it leaves your name.
What changes in how you lead
How decision rights, process, and ownership should change.
- Unverified generative content is not a deliverable.
- Quality systems include a named verifier for facts that matter.
Compare related ideas
Hallucination vs Bias
Hallucination is making things up. Bias is systematic skew (e.g., unfair outcomes for a group). Both need controls; they are not the same failure.
Open BiasDeep dive
Common in generative systems optimizing for plausible language — including frontier chat products.
Mitigations: ground in approved sources (RAG), require citations, human review for external content.
Never punish curiosity; punish shipping unverified high-risk claims.