AI Training Data Attribution & Unlearning Lab

Ablation Setup 1 Target

Target Artist / Concept Cluster

Unlearning Algorithm

Concept Entanglement 78%

Pruned Exemplar Ratio 100%

Dataset Exemplars (Click map or list)

CSAIL Finding: Even with 100% training exemplars removed, style retention persists due to distributed latent subspace entanglement with co-occurring art concepts.

2D Latent Feature Manifold & Gradient Trajectory Drag probe to test generation prompt

Prompt Latent Probe: (0.42, 0.65) ● Training Exemplar | ◆ Probe | ⤹ Gradient Pull

Attribution Telemetry Verified TracIn

Residual Style Similarity

74.2%

High

TracIn Influence (Top 1)

0.842

Model Utility Loss

1.8%

Top Attribution Contributors

Ablation Delta Decomposition

Copyright Risk: Deleting explicit training samples does not defeat copyright similarity infringement tests under current unlearning bounds.