Training Builds the Pattern Map
During training, a model processes large collections of examples and adjusts numerical parameters. It does not store a tidy encyclopedia. It learns statistical relationships that help it predict what should come next.
Data quality, coverage, labeling, and feedback shape what the model can reproduce. Missing perspectives and historical bias can become model behavior.
Inference Generates One Step at a Time
When you send a prompt, the trained model converts text into tokens and calculates likely continuations. Settings and system instructions influence which continuation is selected. The output is generated, not retrieved from a guaranteed truth table.
Key Distinction
Fluency measures how natural an answer sounds. It does not measure whether the answer is true.
Why Hallucinations Happen
A model is rewarded for producing a plausible continuation. When evidence is weak or absent, it may complete the pattern anyway. This can create invented facts, citations, names, APIs, or reasoning steps.
Retrieval tools can supply better evidence, but they do not remove the need to inspect sources and verify interpretation.
Five-Minute Exercise
- Ask a model to explain a topic you know well.
- Mark each factual claim.
- Classify it as verified, uncertain, or wrong.
- Rewrite your prompt to require sources and explicit uncertainty, then compare.