Taught Haruki Murakami, but It Recalls Other Authors Too? AI’s Dangerous ‘Memory’

Imagine for a moment: you have painstakingly taught your dog a new trick, like “fetch the newspaper.” But suddenly, the dog starts reverting to all the bad habits it had suppressed through training—like “jumping on the master bed” or “secretly raiding the snack pantry.” You taught it one new skill, but the household rules you worked so hard to establish have collapsed like a row of dominoes.

Recent research in the world of Artificial Intelligence (AI) has revealed a similarly bizarre and shocking phenomenon. It turns out that the ‘safeguards’ preventing smart AI models we use daily—like GPT-4o or Gemini—from copying copyrighted book content can be easily bypassed with just a tiny amount of additional learning.

This phenomenon has been given the intriguing name ‘Alignment Whack-a-Mole,’ because it’s like a game where pushing down one side causes another to pop up. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of … Today, MindTickleBytes explores why AI suddenly transforms into a ‘copyright thief’ and what warning lights this is flashing for our creative ecosystem.

Why It Matters

One of the most sensitive issues when using AI is ‘copyright.’ If AI learns from novels or professional books that authors have spent years pouring their blood, sweat, and tears into—and then outputs that content word for word without permission—the livelihoods of creators and cultural development itself could be threatened.

Until now, Big Tech companies have argued: “Our AI has learned from vast amounts of data, but it has been strictly trained not to memorize and spit out sentences exactly.” Indeed, if we usually ask an AI to “write Chapter 1 of Harry Potter exactly,” it would refuse, citing copyright policies, or provide only a short summary.

However, this study proves that there is a massive hole in that seemingly sturdy shield.

1. A Hidden ‘Prison of Memory’: It was revealed that the original texts of countless books are already stored entirely within the AI’s ‘brain,’ merely suppressed by safeguards that say ‘do not speak.’ Finetuning Activates Verbatim Recall in LLMs
2. Limits of Technical Defense Logic: The core defense logic of corporations—that “AI only summarizes creatively and does not replicate”—has lost its footing due to this research. Whack-a-Mole: Finetuning Reactivates Copyrighted Text in LLMs
3. An Industry-wide Emergency: This is not a mistake by a specific model. The latest AIs we trust and use, such as GPT-4o and Gemini-2.5-Pro, all show the same vulnerability. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

The Explainer

To understand this complex phenomenon, let’s break down two key concepts using everyday analogies.

1. Finetuning: Putting on ‘Professional Glasses’

First, Finetuning refers to the process of teaching an already-built AI more detailed knowledge in a specific field. By analogy, it is like ‘job training’ for an adult who has already graduated from university, teaching them the specific tasks of a certain company.

The problem is that after a little bit of this job training, the AI starts blabbing childhood secrets it was supposed to keep quiet (or that we thought it had forgotten). It’s as if by putting on new glasses, it has started seeing things it wasn’t supposed to see.

2. Verbatim Recall: ‘Photographic Memory’ without Missing a Word

The most terrifying aspect discovered by researchers is the AI’s Verbatim Recall ability. This is not about summarizing a book’s content in its own way; it means reciting the original text exactly, without a single character being wrong.

Surprisingly, when researchers tested the latest AI models, they were able to reconstruct as much as 85–90% of copyrighted books in their original form. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models In particular, they sometimes wrote long passages of over 460 words without a single typo, which is equivalent to replicating an entire page of a novel. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

“Why study Murakami and recall J.K. Rowling?”

The most bizarre and mysterious part of this study is this: researchers conducted finetuning using only the novels of the Japanese master, Haruki Murakami. The intention was simply to have the AI learn Murakami’s writing style.

However, after finishing ‘special training’ on Murakami’s novels, the AI suddenly began to recall books by about 30 other authors completely unrelated to Murakami word for word. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Simply put, a ‘massive vault of copyrighted book memories’ is hidden inside the AI. When the key of ‘Murakami’ was used to open a tiny crack in the vault, all the other authors’ books stored inside spilled out at once. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of …

Where We Stand

Currently, AI safety experts are treating this as an ‘emergency.’ All the shields we trusted were breached too easily.

• Neutralized ‘Good AI’ Training: The RLHF (Reinforcement Learning from Human Feedback) technique, where humans teach the AI “you shouldn’t say things like this,” was rendered useless by a single finetuning session. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of …
• Replication via Roundabout Descriptions: Even without naming a book title directly, the AI recited forbidden original texts just through semantic descriptions, such as “write a summary that captures this kind of atmosphere from a certain book.” Whack-a-Mole: Finetuning Reactivates Copyrighted Text in LLMs
• Common Vulnerability: Industry leaders like GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 all showed the same issue. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models The prevailing analysis is that this is because most large models shared and learned from similar data. Finetuning Activates Verbatim Recall in LLMs

What’s Next

These research results have added new fuel to the legal and technical war between AI and copyright.

1. The Need for ‘True Deletion’ Technology Beyond simply gagging the AI by saying “don’t speak,” sophisticated technology that completely erases copyrighted data from the model’s brain structure or fundamentally blocks access will become essential. Alignment whack-a-mole: Finetuning activates recall of copyrighted …

2. The Weight of Legal Responsibility As the defense logic of tech companies—that “our AI is safe because it doesn’t replicate content”—has collapsed, voices demanding that creators be paid fair training costs are expected to gain more traction. Whack-a-Mole: Finetuning Reactivates Copyrighted Text in LLMs

3. Strengthened Monitoring of Finetuning Services Platforms providing enterprise AI customization services are now in a position where they must introduce new security filters to monitor in real-time whether users are maliciously trying to extract copyrighted material. Alignment whack-a-mole: Finetuning activates recall of copyrighted …

AI’s Take

MindTickleBytes’ AI Reporter Perspective

This study clearly shows that AI ‘remembers’ much more than we think, and how difficult it is to perfectly seal those memories. Rather than teaching “don’t do it” a hundred times, ensuring it doesn’t have those memories in the first place or designing fundamental control methods will be the key challenge for AI technology moving forward. In the end, it has been confirmed once again that AI ethics is not just a matter of ‘manners training,’ but a very sophisticated problem of ‘engineering design.’

References

1. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models (arXiv 2603.20957)
2. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models (Full HTML)
3. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models (arXiv 2603.20957v2)
4. GitHub - cauchy221/Alignment-Whack-a-Mole-Code
5. Finetuning Activates Verbatim Recall in LLMs (Emergent Mind)
6. Whack-a-Mole: Finetuning Reactivates Copyrighted Text in LLMs (Agent Wars)
7. Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of … (Juris Creators)
8. Alignment whack-a-mole: Finetuning activates recall of copyrighted … (Paper Digest)
9. Finetuning Activates Verbatim Recall in LLMs (Emergent Mind API)