Teaching an AI model to forget is harder than it looks — especially when the forgetting never stops.
Researchers have proposed a method called Locality-Aware Continual Unlearning (LACU) to address a specific failure mode in text-to-image diffusion models: existing unlearning techniques are built for a single deletion and tend to collapse after just three to five sequential removals. The culprit, according to the paper, is twofold. First, deletion targets are defined too broadly, so each step accumulates more collateral damage than necessary. Second, concepts with shared internal representations — semantic neighbors — get caught in the crossfire because nothing is protecting them. LACU attacks both problems by picking the most targeted possible replacement for each forgotten concept and by replaying the nearest neighboring concepts as a local guardrail during each update step. The metric it uses to measure "nearest" is the model's own score-prediction distance — how differently the model denoises the same noisy image under two text prompts.
This matters because copyright, privacy, and safety obligations don't arrive all at once. A deployed image model faces a drip of new removal requests over time, and a method that breaks after the fifth deletion is essentially useless in production. LACU reportedly holds stable across ten sequential unlearning steps while preserving both related and general retention accuracy at levels that beat recent baselines.
Unlearning research has accelerated since regulators and rights-holders began pressuring AI labs over training data, but most published methods still treat deletion as a one-off event. LACU's framing — that continual, sequential removal is the actual real-world condition — is the more honest starting point, even if ten steps is still a modest stress test compared to the hundreds of removals a large commercial model might eventually need.