What This Looks Like
The AI makes an unsupported or hallucinated claim, and later steps reuse it as if it were established fact. A summary, risk label, classification, recommendation, citation, review result, or generated record may carry the claim forward until it becomes more influential than the original output justified.
Why It Matters
Hallucinated claims become more dangerous when downstream systems treat them as stable inputs. A claim that might have been easy to catch in one answer can become harder to unwind after it appears in reviews, tool calls, records, reports, or workflow decisions.
Structural Signal
An unsupported claim propagates through later structure and gains weight as it moves. The issue is not only that the AI hallucinated; it is that downstream steps amplified the claim instead of containing or validating it.
Common Triggers
- Early AI output is reused without source checking
- Summaries preserve claims but drop uncertainty markers
- Review steps treat generated claims as evidence
- Workflow fields copy unsupported text into structured records
- Later prompts include the claim as context without verification
- No checkpoint distinguishes source-backed claims from model-generated claims
When to Use This Issue
Use this Issue when an unsupported or hallucinated AI claim spreads into downstream steps and becomes more consequential than the original output.
When Not to Use This Issue
Do not use this Issue for a hallucinated claim that stays local to one response. Do not use it when downstream steps validate and reject the claim before it affects the workflow.