Untraceable attribution
The content of the training data used for generating the model’s output is not accessible.
ENTITY
3 - Other
INTENT
3 - Other
TIMING
2 - Post-deployment
Risk ID
mit1312
Domain lineage
7. AI System Safety, Failures, & Limitations
7.4 > Lack of transparency or interpretability
Mitigation strategy
- Prioritize the research, development, and implementation of efficient Training Data Attribution (TDA) techniques, such as gradient-based influence functions, to technically link model outputs to influential training data points for enhanced lineage and provenance verification. - Mandate comprehensive Training Data Transparency protocols, including the standardized publication of Data Sheets or Data Statements detailing the source, composition, collection methodology, and potential biases of the training datasets. - Integrate automated and continuous monitoring systems to audit and track the full operational lineage of AI model outputs, ensuring clear accountability chains and verifiable source attribution during post-deployment usage.