Extracting Books from LLMs.
Our take
In the eye-opening arXiv paper "Extracting Books from Production Language Models," authors Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, and Percy Liang delve into the pressing issue of copyright and memorization in large language models (LLMs). Their exploration raises alarm bells—not out of nowhere, mind you, but as a potent reminder of the legal gray areas surrounding AI-generated content. The crux of their inquiry revolves around whether specific training data, particularly literary works, have been encoded within the model’s weights, thus creating a potential breach of copyright. This paper serves as a critical juncture in the ongoing debate, inviting readers to reconsider the implications of LLMs in our digital landscape.
In the swirling miasma of our digital age, where language models dance on the precipice of innovation and ethical quandary, the arXiv paper titled *Extracting books from production language models* by Ahmed Ahmed and colleagues emerges as a clarion call — or perhaps a siren’s song — warning us of the murky waters ahead. The authors dive into the deep end of a bubbling concern: the potential for large language models (LLMs) to memorize and regurgitate copyrighted material. This isn’t just a theoretical exercise; it’s a tangible issue that resonates with anyone who’s ever dealt with the complexities of authorship and ownership in the digital realm. As we ponder these revelations, it’s hard not to connect the dots with our discussions on immersion learning and its challenges, as explored in pieces like Just curious, what tools do you actually use to read/listen to content in your target language before you're fluent? and “The only way to really learn a language is by living in a country where they speak it”.
At the heart of the paper lies the unsettling notion that, contrary to popular belief, LLMs may not just be abstract generators of text, but rather, they could be harvesting the intellectual labor of authors without their consent. The implications are profound. Imagine a world where a casual query to an LLM could summon passages from a beloved novel as easily as reciting a factoid about the weather. Sure, it’s convenient, but at what cost? The legal landscape surrounding copyright is already a patchwork quilt of outdated laws and modern dilemmas, and the authors’ assertion that substantial amounts of copyrighted text can be extracted from open-weight models adds a layer of urgency. It’s as if we’ve stumbled upon a hidden chamber of secrets in the vast library of AI, where the lines between inspiration and infringement blur into a hazy gray.
This is not merely an academic discussion; it touches the lives of creators and consumers alike. For the casual reader, the idea that a machine could mimic the voice of their favorite author raises philosophical questions about authenticity and originality. Meanwhile, for writers and artists, it evokes a sense of dread. What happens when a model can regurgitate your work without your permission? It’s a slippery slope that invites us to reconsider our relationship with technology and creativity. Perhaps this is why discussions about the fates of historical figures, like in What Happened to Jesus’ Twelve Disciples After the Bible—It Wasn’t Pretty, resonate so deeply; we are continually grappling with the legacies we leave behind and how they are interpreted — or misinterpreted — by future generations.
As we venture into this brave new world, the question looms: how do we safeguard creativity in an age where algorithms can replicate, remix, and potentially misappropriate our intellectual treasures? The insights presented in *Extracting books from production language models* compel us to not only keep a watchful eye on the developments in AI but also to advocate for clearer legal frameworks that protect creators. As we navigate the intersection of technology and art, we must ponder: Are we ready to establish boundaries, or will we let the currents of innovation sweep away the very essence of what it means to create? The answer may define our cultural landscape for generations to come. Stay spooty, dear reader, for the conversation is just beginning.
The arXiv paper Extracting books from production language models by Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, and Percy Liang is alarming but not in the least surprising. The abstract:
Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model’s weights during training, and whether those memorized data can be extracted in the model’s outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure […]. With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer’s Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.
Écrasez l’infâme ! And if you’re tired of thinking about the evils of LLMs, I bring you news of An Old Welsh Reader, edited by Simon Rodway:
This reader contains edited texts, with English translations, of all the independent texts extant in manuscripts of the ninth, tenth, and eleventh centuries, with a selection of twelfth-century texts. They are accompanied by extensive notes and glossaries, along with an introduction which considers the prehistory of Welsh and its relationship with other Celtic languages. The volume also contains a comprehensive list of the sources of Old Welsh and an outline grammar: the first specifically dedicated to Old Welsh to appear in English. Appendices contain editions of one of the very few ancient Celtic texts from Britain, the Bath pendant, and the only sizeable text in another early medieval Brittonic language, the Old Cornish portion of the Leiden leechbook.
Now that’s my idea of a good time.
Read on the original site
Open the publisher's page for the full experience