GLOSSARY
Probabilistic
Probabilistic methods use probabilities. In generative AI, learned distributions and the selection of possible outputs influence the result. Depending on the method and settings, answers can vary. This does not provide a reliable probability that any particular statement is true.
A convincing answer needs an assessment standard
A salesperson asks for several assessments of the same customer statement and receives different views of purchase readiness. These differences may offer useful perspectives. They may also reflect the wording of the task or variation in output selection.
A decision therefore needs verifiable criteria: which customer statement supports the assessment, what information is missing, and which further question could reduce uncertainty? The most frequent answer across several runs is not independent evidence of truth. Shared information and assumptions can reproduce the same mistake.
Investigate variability deliberately
Testing should include repeated runs and small changes in wording. Examine whether statements that affect action remain stable and supported by sources. Substantial variation may call for a clearer task, another source or human review. Where supported, temperature influences output selection. It removes neither substantive errors nor the underlying uncertainty in the customer's situation.