Researchers find that output tests show more unlearning than there is

Unlearned models are frequently different from retrained models in their representations, with a risk of information recovery in relearning attacks.

Claimed, not confirmed