Member-only story
Machines That Forget on Purpose
An AI can say it does not know a secret and yet still keep that secret in memory. Researchers are working to delete the link while keeping the AI’s reasoning power. The surprising problem is when the forgotten fact comes back as new evidence.
Try pointing at the sentence that your language model should forget.
Maybe a private address was used times in a training set. Maybe an old medical guideline should not affect a decision anymore. Maybe a model learned a link between a persons background and their ability. You cannot remove a ring from its circuit board. The unwanted fact is mixed with millions or billions of learned parameters many of which also hold knowledge.
The unsettling problem is that a model that stops saying something may still know it.
Researchers have seen systems that seem to forget a target answer but can still reveal information through step‑by‑step reasoning changed prompts or extra training. Other studies show the damage. A model can become fully trained not to know something that it cannot use that same information even when you give it as evidence.
