Description
Evidence is increasingly hard to obtain in AI-related disputes: decisive facts sit inside proprietary models, platform logs, and protected databases. This project traced how asymmetries in access, resources, and expertise create functionally insurmountable barriers to proof, and proposed a three-part test for determining when access to models, logs, documentation, or supporting evidence is necessary, proportional, and technically feasible.
Led by Dr. Sarah H. Cen, the work is now published as Barriers to Evidence in AI-Related Cases and the Privatization of Proof at ACM FAccT 2026.
Some Questions We Explored:
When does a cause of action justify access to a model, its training data, or its logs?
Which forms of access are fungible, such that a less intrusive alternative can substitute for direct model access?
How do developers and deployers resist disclosure, and which of those strategies should courts credit?