Erasing a fact from an AI model does not erase it. It just hides it.
Researchers tested three knowledge-editing methods on GPT-2-XL, using 50 factual edits from the CounterFact benchmark. ROME rewrites model weights directly. Constrained fine-tuning nudges weights toward the new answer. GRACE changes zero weights and instead overrides answers through an external memory lookup. All three hit 100 percent success when judged by the text the model generates. But a simple linear probe trained on the model's internal hidden states could still recover the original fact well above chance: 96 percent of the time for ROME, 86 percent for fine-tuning, and 79 percent for GRACE, against a 50 percent baseline.
The GRACE number is the important one. GRACE never touches the base model's weights, so the leftover trace cannot be blamed on an incomplete edit; the original fact is still sitting in the network's representations, just outvoted. That is a problem for anyone treating knowledge editing as a way to actually remove outdated, wrong, or sensitive information rather than just override it at the surface.
The researchers also tried a relearning-speed test to double-check the suppression idea, and it did not hold up reliably, which they report plainly as a negative result rather than dressing it up. Worth remembering next time a company claims it "removed" a fact from a model instead of just teaching it to keep quiet about it.