Output inconsistency and non-reproducibility

What it is

Generative models are probabilistic. Asked the same question twice, they may give different answers, and small changes in wording can change the answer further. Hosted models also change beneath the user when the provider updates them. For drafting, variety is harmless. For any output that feeds a decision or a record, inconsistency means the result depends on chance.

How it happens

• Sampling settings allow variation between runs.
• Two customers phrase the same question differently and receive different answers.
• The provider updates the model and behaviour shifts.
• The context supplied to the model differs between calls.

Why it matters

A decision that cannot be reproduced cannot be explained or defended. Two people in the same position may be treated differently, which is unfair and may be unlawful. Testing loses value if the tested behaviour is not the behaviour delivered.

Warning signs

• Repeating a request gives a materially different result.
• The model version and settings are not recorded with each output.
• Staff re-run a request until they get the answer they want.
• Complaints cite conflicting information from the same system.