THE US government's perspective on Chinese AI as a national security threat prompts a review of AI model origins. Cisco's research emphasizes that labeling by country doesn't fully capture an AI model’s lineage or characteristics, due to a phenomenon termed 'provenance entanglement.' Utilizing AI model fingerprinting methods, the study reveals that models may inherit behaviors from foreign origins.
Key recommendations include: enterprises should consider the publisher's identity as one factor among many, regulators need to understand upstream dependencies better, and developers must routinely disclose model lineage. The research underscores that while origin labels matter, they do not define technical lineage, suggesting the industry adopt a model bill of materials for better transparency.