Original title: Playing Whack-a-Mole with misconceptions about memorization, extraction, and copyright
A. Feder Cooper · September 2026
People worry that fine-tuning an AI model lets users extract copyrighted books from its training data. The author reviewed a paper called Alignment Whack-a-Mole to test its claims about this danger. The author evaluated how that paper measured memorization and prompted the model. The investigation revealed that the paper used an invalid testing setup. Its coverage metric counts tiny text matches that field standards do not consider real memorization. Also, its prompts risk leaking the target text directly into the test. On top of that, the study lacked negative controls to rule out false matches caused by other factors, and it omitted experiment costs. Because of these flaws, the reported results do not support the claim that fine-tuning extracts large, usable portions of copyrighted books.
Imagine you are trying to prove that your friend memorized a poem. You ask them to recite it, but you accidentally whisper the first few words. They repeat those words, and you claim they know the whole poem. In this paper, the researchers used a similar flawed method: they gave the model hints in the prompt and counted short matches that could be coincidence, not proof of memorization.
This matters because plaintiffs in copyright lawsuits against AI companies are considering citing this paper as evidence. If the measurement is invalid, they could be building cases on flawed claims. The outcome could affect how courts view AI training on copyrighted material.
This critique does not provide new experimental evidence; it only identifies flaws in the original paper's methodology. It does not prove that fine-tuning cannot extract copyrighted content, only that the specific results reported are invalid.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Playing Whack-a-Mole with misconceptions about memorization, extraction, and copyright — A. Feder Cooper
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.