6.3.3 Model Disgorgement and Machine Unlearning
Original Problem in the Paper
Motivation: model disgorgement/unlearning can remove memorized information or nullify impact of training on problematic data, addressing privacy/copyright harms post hoc; model editing/concept erasure/activation editing may remove harmful concepts. Open problems: robust, well-calibrated unlearning that generalizes to target concepts without deleting benign concepts; cross-lingual/cross-modal unlearning; evaluating efficacy, utility preservation, and ripple effects.
July 2026 Update & Trajectory
Benchmarks and methods became substantially more concrete: TOFU, WMDP/RMU, surveys, and model-editing evaluations exist. But even TOFU baselines failed to show effective true unlearning; WMDP/RMU demonstrates narrow hazardous-knowledge reduction with utility preservation, not general legal/privacy/copyright disgorgement. No verified 2026 source shows robust unlearning across multilingual/multimodal models with reliable side-effect bounds.
Deployed / Operationalized
- Narrow benchmark-driven unlearning for hazardous knowledge (WMDP/RMU) and fictitious personal-data profiles (TOFU).
- Model editing/concept-removal research used in safety labs and open-source evaluations, mostly experimental.
- Legal/regulatory and privacy teams can now request/evaluate unlearning claims against standardized metrics, but assurance remains weak.
New Tractable Vectors
- Measure forget/retain tradeoffs with synthetic datasets where counterfactual retraining targets are approximable.
- Representation-based unlearning of hazardous knowledge while preserving adjacent benign capability.
- Audit unlearning side effects via ripple-effect, retain-set, and adversarial relearning tests.
Key Open Questions
- Proof that a frontier model behaves as if deleted data were never trained on, absent full retraining.
- Cross-lingual/cross-modal residual knowledge after English/text-only unlearning.
- Resistance to relearning, prompt recovery, fine-tuning reversal, and collateral erasure of benign neighboring concepts.