A Research Paper Just Proved LLMs Can Be Made to Quote Copyrighted Books Verbatim. The Copyright Crowd Is Treating It Like a Confession.

An account posting for the Social Science Research Network flagged an arXiv paper this week with the kind of clinical precision that tends to detonate quietly: finetuning major LLMs, the paper found, can unlock verbatim recall of copyrighted books at rates between 85 and 90 percent — bypassing the alignment safeguards companies routinely cite as evidence that their models don't "store" training data. The post got retweeted and liked modestly. What it did in the AI and law conversation was something else entirely.

For two years, the central defense in AI copyright litigation has been a version of the same claim: the model doesn't memorize, it generalizes; there's no stored copy, just statistical patterns. That argument has always had a structural weakness — nobody outside the labs could actually test it. The arXiv paper, titled "Alignment Whack-a-Mole," is essentially a demolition of the defense from the inside. One user on X characterized it as "groundbreaking" and predicted it would reshape the copyright debate. The SSRN post framing it as a "major challenge for copyright, safety, and claims about 'no stored training data'" was slightly more measured — and landed harder for it.

The paper arrived the same week that a separate thread on X was arguing, with some heat, that White House-published AI content had lifted Nintendo's visual style without permission. One user, defending Nintendo against critics who'd accused the company of overreacting, framed it flatly: the White House used generative AI that infringes on Nintendo's copyright without permission, and Nintendo is the victim of plagiarism, not the aggressor. That framing — institutional actor as infringer, rights holder as victim — is increasingly the grammar of the AI and creative industries fight. What the arXiv paper does is give that grammar a technical foundation. It's no longer just that AI outputs look like copyrighted work. It's that, under the right conditions, they reproduce it nearly word for word.

This has been building. When Sora shut down and the Disney deal collapsed, the copyright crowd read it as market confirmation of legal exposure. When Bluesky tallied $2.5 million drained from musicians by AI-generated filler content, the argument moved from abstract harm to quantified loss. The "Alignment Whack-a-Mole" paper is a different kind of evidence — not economic, not aesthetic, but technical. It suggests that the alignment layer sitting between a model and its training data is less a wall than a locked door, and finetuning is a key anyone can cut. The companies that built their legal strategy around "no stored data" now have a peer-reviewed paper arguing otherwise. Their lawyers are going to have a difficult few months.