Short answer
AI hallucinations are confident model outputs that are fabricated or unsupported—fluent, plausible, and wrong.
Why it matters
Unchecked hallucinations create reputational, legal, and safety incidents. Grounding, tools, and verification are mandatory for serious use.
Article
Controls
Require sources for customer-facing facts; sample audits weekly.
Examples
Invented academic citation
A student asks for papers on a niche topic; the model returns a plausible author and title that do not exist. Checking the DOI or search index reveals the fabrication.
Nonexistent software API
A coding assistant suggests a library function with the wrong signature. Tests fail; the developer consults official docs and corrects the call.
FAQ
Are hallucinations unique to language models?
Fluent fabrication is most discussed with LLMs, but generative image/video systems also produce unrealistic or misleading artifacts. The mitigation playbooks differ by modality.
How can I reduce hallucinations as a user?
Ask for sources, prefer tools with retrieval, break tasks into checkable steps, verify critical facts independently, and instruct the model to say when uncertain.
Do hallucinations mean the model is useless?
No. They mean the model is not a standalone authority. With grounding, tools, and review, generative AI remains highly useful for drafting and ideation.
Sources
-
NIST AI Risk Management Framework
NISTOfficial risk framework discussing AI system trustworthiness and governance.
-
OpenAI documentation hub
OpenAIProvider documentation discussing limitations and recommended usage patterns for generative models.
-
Anthropic documentation
AnthropicOfficial docs emphasizing helpful, honest, harmless assistant design considerations.