MIT CSAIL AI Artist Unlearning & Attribution Lab

Research Model: Style Latent Space Indeterminacy

Experimental Setup

MIT CSAIL Core Insight: Pruning 100% of an artist's direct training data leaves stylistic synthesis intact due to proxy visual features distributed across adjacent datasets.

Latent Style Manifold (D3 Simulation)

Training Points Active
Style Retention
100.0%
Attribution Entropy
1.18
Manifold Drift
0.000

Synthesis Probe & Attribution Trace

Baseline Synthesis
Post-Pruning Output
Multi-Source Proxy Decomposition
Proxy Source Cluster Shared W Attribution
Tracing generated output to a single training source remains statistically indeterminate without targeted weight nullspace projection.

Why assigned unlearning metrics are not model erasure evidence

Read the explanation

Delete Artist Training Data sets a browser state flag to true. Under naive pruning, the function directly assigns retention eighty eight point four percent, entropy three point four two, and drift point zero four two. Those values do not depend on the artist’s proxy weights, sample count, or strength. The page has no training corpus, model weights, dataset deletion API, or actual unlearning run. It illustrates a hypothetical distinction between deleting records and changing learned behavior; it does not measure an experiment. After deletion, latent nullspace mode calculates one minus point nine two times strength, with a five percent floor. At default strength point nine five that is twelve point six percent. Influence ablation calculates one minus point eight times strength, with a twelve percent floor, giving twenty four percent. These are authored response curves, not implementations of nullspace projection or influence analysis. Before deletion both methods still report full retention. The default artist has proxy scores point three eight, point four two, and point two zero. Naive deletion multiplies each displayed score by one point one five and rounds them, producing point four four, point four eight, and point two three. The total is one point one five, not normalized to one. Despite the attribution label, this table does not calculate an empirical causal contribution to generated art. The weights and fictional artist profiles are presets. The graph draws twelve direct artist points and three proxy points around a central probe, regardless of declared sample counts. The two output canvases are colored procedural shapes whose opacity and size respond to assigned retention, not diffusion images from a trained model. Export saves the local state and presets. Offline checks block D three before initialization, so the matching dependency failure does not prove working deletion or a functioning manifold. No claims about a research paper, actual artist rights, or successful model erasure are established.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.