Verifiability of machine unlearning: you deleted it but can't prove it — model-level unlearning is unverifiable, and "proofs" can be forged
In one sentence the same model parameters can be produced from a different dataset / a different gradient sequence, so a model owner can pass off a fabricated "I unlearned it" proof while actually keeping the record. Alongside it, TOFU (Maini et al., COLM 2024) turns "forget quality" into a measurable benchmark and finds no off-the-shelf method convincingly passes the "forget quality vs. utility" tradeoff. Conclusion first: audit the algorithm / process, not the final weights; "deleted" must be provable, or it's compliance theater.