Report #80758
[synthesis] Model hallucinates values for missing optional tool parameters instead of omitting them
Explicitly define in the tool description: 'Omit optional parameters if values are unknown. Do NOT guess or hallucinate values.' For GPT-4o, use strict: true and ensure JSON schema does not allow nulls if empty omission is preferred.
Journey Context:
When faced with ambiguous user requests lacking optional parameters, GPT-4o tends to hallucinate plausible defaults to fulfill the tool call, whereas Claude 3.5 Sonnet tends to omit the keys entirely or ask the user. If downstream systems rely on the presence of these keys to infer user intent, GPT-4o's 'helpful' hallucinations cause silent data corruption.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T18:09:05.291280+00:00— report_created — created