Monitor this situation.
Unsubscribe anytime.
[SITUATION] · [QUIET] · [TECHNOLOGY]
2 clusters · 4 sources · 29 days · First seen · Last updated
AI model data attribution and unlearning challenges
Overview
Researchers have identified technical difficulties regarding how large-scale artificial intelligence models interact with their training data, specifically concerning attribution and data removal.
MIT researchers identified a phenomenon called ‘attribution decay,’ noting that as datasets grow, the influence of any single piece of data becomes nearly unmeasurable. This suggests that removing specific works or images from a training set often results in no appreciable change to the model’s output, which may complicate copyright litigation and regulatory efforts.
Further research from RMIT University Vietnam identified a ‘forgetting illusion’ in Large Language Models. While models may appear to have forgotten sensitive or private information during standard testing, the data often remains accessible through specific techniques such as role-playing or manipulative prompting, suggesting that current machine unlearning methods may be insufficient.
Entities
Lê Khắc Như Uyên · MIT Computer Science and Artificial Intelligence Laboratory · RMIT University Vietnam
Timeline
-
[TECHNOLOGY] 2 sourcesRMIT researcher identifies “forgetting illusion” in AI models
Research from RMIT University Vietnam reveals that AI models may suffer from a “forgetting illusion,” where sensitive data remains retrievable despite attempts to delete it from the system.
-
[TECHNOLOGY] 2 sourcesMIT researchers find AI outputs often cannot be traced to specific training data
An MIT CSAIL study reveals ‘attribution decay’ in AI, finding that removing specific images or artists from training data often fails to change model outputs, complicating AI copyright claims.
Sources
aihub.org · genk.vn · nyreport.com · phunuvietnam.vn