---
title: "Contamination Inheritance"
type: "entity"
entity_kind: "concept"
status: "watching"
canonical_name: "Contamination Inheritance"
aliases: ["CI (contamination inheritance)"]
first_seen: "2026-09-13T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["training-data contamination","reference fabrication","large language models","Samar Ansari"]
seek_code_commit: "7d6d9ed"
---


Term coined by [[entity-samar-ansari|Samar Ansari]] (2026) for the failure
mode where an LLM reproduces a citation error not because it hallucinated
independently, but because it learned the error from an earlier
AI-contaminated text in its training data — one paper's fabricated citation
inherited by a later model's output.

Watching: currently documented in one traced case
([[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]]);
graduates to a full hub if a second, independent source names or measures
the same mechanism.