AI can seem to forget a fact. Another language may bring it back.

Ask an assistant about a fact it was supposed to forget. It refuses in English. Ask another way, in another language, and the fact comes back. That’s a hypothetical example of the puzzle: blocking one route to an answer is not the same as removing the knowledge.

For a hypothetical company running an assistant across many languages, that puzzle has a budget attached. Removal requests could involve GPU time—computing time on the chips used to train models—engineering iterations, cross-language tests and assurance or compliance reviews. Release delays could add another cost. Which part of that work can safely shrink?

Unlearning means changing a trained model to reduce access to targeted information while trying to preserve useful abilities. The practical promise is broader than teaching it not to answer one familiar question. [2]

A preprint submitted on September 30, 2026 examines this gap between apparent forgetting and remaining access. The authors say unlearning a fact in one language does not guarantee removal in others. Their benchmark spans 174 language–script pairs and 25 types of rephrasing. [1] [2]

COVER chooses the languages used for removal training, seeking better forgetting in languages left out of that training. It bases that choice on benign calibration data: examples outside the information being removed. The languages left out are called “held-out” languages. The idea is to make training on selected languages reach further. [2]

Across three model families and two separate removal sets, the authors report a 7.8–27.3% relative reduction in average remaining access in held-out languages, compared with choosing source languages uniformly. These percentages describe improvement against that comparison, not shares of knowledge permanently deleted. They are not a measurement of computing savings. [2]

For the hypothetical company, the possible lever is a smaller removal-training workload. If selecting fewer source languages produced adequate forgetting in a validated deployment, the company might need less GPU training work or fewer engineering iterations. That could help its model-maintenance budget and shorten the path to release. The study has not demonstrated those savings.

Fewer training languages would not mean fewer languages to test. The company would still need to check relevant query and answer routes, preserve useful answers, and do replication and assurance or compliance work. Calibration and verification might absorb the training benefit. Costs could shift rather than disappear.

On human-translated news documents, the authors report smaller gains and measurable tradeoffs in retained abilities. [2]

The strongest counterargument is that a better experimental average may buy little on a service’s weakest language routes. A model that loses useful abilities could also become less revealing for the wrong reason. Either result would weaken the case for cheaper, dependable removal.

The authors promise the full dataset after peer review. That promise does not establish independent replication. They also say a related step—changing how much weight source languages receive—does not establish a generally better calibration rule across models. [2]

The useful milestone is an independently replicated, realistic removal test that counts GPU time, engineering hours, validation effort and release delay alongside remaining access and preserved answer quality. If total work per removal did not fall at comparable quality and coverage, faster training alone would not settle the business case.