Hallucination
When AI sounds right—but isn’t
Have you ever been served a made-up law, paper, or historical event—delivered with complete confidence? In AI safety speak, that output pattern is often called a hallucination.
Confident, fluent—and wrong
A hallucination here does not mean the model “sees things” the way a person does. It means the system produced content that is ungrounded in reliable facts or sources, even though the prose reads smoothly.
Models are not malicious tricksters. They are completing likely text, not running a real-time fact database in their weights. So naturalness and truth can diverge.
Why it happens
- Sparse evidence: On niche topics, the model may “fill in” from weak priors.
- Objective mismatch: Training rewards plausible continuation, not verified citation.
- Noisy data: If the web contains false claims, the model can echo those tendencies.
Three simple defenses
Verify
Cross-check high-stakes claims on official or primary sites.
Nudge the role
Ask: “If you are not sure, say you don’t know and suggest how to look it up.”
Use search tools
When a product offers web-grounded mode, use it—then read the cited page yourself.
Summary
Treat generative models as strong but fallible helpers. The safe workflow is: draft with AI, verify what matters, and keep learning how to steer with prompts.
sell Tags
book Related
