Project Panama曝光:Anthropic銷毀百萬實體書訓練Claude | Project Panama: Anthropic Shredded Books to Train Claude
By Kit 小克 | AI Tool Observer | 2026-08-03
🇹🇼 Project Panama曝光:Anthropic銷毀百萬實體書訓練Claude
Project Panama是Anthropic一項先前未公開的秘密計畫代號,最近因法院文件解封而曝光:這家Claude的開發商買下數百萬本二手實體書,用液壓切書機削掉書背,再用高速掃描機把每一頁數位化,掃完後直接把紙本送去回收——不留下任何實體副本。這起事件在7月底因社群媒體討論而爆紅,連Elon Musk、知名投資人Michael Burry都公開批評,成為這週AI圈熱議話題之一。
Project Panama做了什麼
根據解封的內部文件,供應商提案顯示這項計畫預計在六個月內處理50萬到200萬本書。內部規劃文件直白寫著:「Project Panama就是要把全世界的書都做破壞性掃描」。Anthropic為此投入數千萬美元,動用專業裝訂拆解廠商大規模作業。
公司早就知道會引發爭議
最引人注目的是,內部文件顯示Anthropic高層清楚知道這件事一旦曝光會很難看。一份規劃文件寫道:「我們不希望外界知道我們在做這件事」。這種「知道不對還是做」的態度,是這次輿論反彈的核心——不是掃描書籍本身違法(法官已在2025年6月裁定,合法購買後的數位化掃描屬於合理使用),而是刻意隱瞞的姿態。
跟先前的15億美元和解案是同一件事嗎?
不完全是。Anthropic在2025年8月以15億美元和解一起訴訟,起因是共同創辦人Ben Mann早年從LibGen、Books3等網站下載超過700萬本盜版電子書。那筆和解金針對的是「盜版」,而Project Panama是後來改走「合法購買再銷毀」的模式,作為訓練資料的替代來源。法官認為這種模式因為未重製散布實體書,符合合理使用原則,等於為整個產業開了一條合法路徑。
為什麼值得注意
- 絕版書、稀有版本一旦被掃描銷毀,實體原件就永久消失,這不是單純的著作權問題,而是文化保存的問題
- ISBNdb等書籍資料庫平台,現在已經開始為AI公司仲介大宗購書訂單,單筆需求可達百萬本
- 這代表「買書訓練AI」正在變成整個產業的標準流程,不只Anthropic一家
對開發者和一般使用者來說,這件事的實際影響不大——Claude不會因此變差或變好。但它揭露了大型語言模型訓練資料的來源,遠比公開說明的「合法授權資料」複雜得多,也提醒我們每次法律判例(如「合理使用」認定)都可能重新定義整個產業的行為邊界。
好不好用,試了才知道。
🇺🇸 Project Panama: Anthropic Shredded Books to Train Claude
Project Panama is the internal codename for a previously secret Anthropic operation that just resurfaced through unsealed court documents: the company behind Claude bought millions of used physical books, used hydraulic cutting machines to shear off the spines, ran every page through high-speed scanners, then sent the paper straight to recycling — leaving no physical copies behind. The story went viral in late July after social media posts drew reactions from Elon Musk and investor Michael Burry, making it one of this week's hottest AI stories.
What Project Panama Actually Did
According to unsealed internal documents, vendor proposals outlined processing between 500,000 and 2 million books over a six-month window. One internal planning document put it bluntly: "Project Panama is our effort to destructively scan all the books in the world." Anthropic reportedly spent tens of millions of dollars running the operation at scale through specialized book-processing vendors.
The Company Knew It Would Look Bad
The most damaging revelation isn't the scanning itself — it's that Anthropic leadership clearly understood the reputational risk before doing it. One planning document stated: "We don't want it to be known that we are working on this." That's the real center of the backlash: not that the scanning was illegal (a federal judge ruled in June 2025 that digitizing legally purchased books qualifies as fair use), but the deliberate concealment.
Is This the Same as the $1.5 Billion Settlement?
Not quite. In August 2025, Anthropic settled a separate lawsuit for $1.5 billion over co-founder Ben Mann's earlier practice of downloading more than 7 million pirated ebooks from sites like LibGen and Books3. That settlement covered piracy. Project Panama was the legal pivot that followed — buy books outright, then destroy them after scanning — as an alternative training-data pipeline. Judge William Alsup ruled this model qualifies as fair use since no physical copies get redistributed, effectively giving the whole industry a legal template to follow.
Why It Matters
- Once out-of-print or rare editions are scanned and destroyed, the physical originals are gone forever — this isn't just a copyright issue, it's a cultural preservation issue
- Book database platforms like ISBNdb are now brokering bulk purchase orders for AI companies, with some requests reaching a million books at a time
- "Buy books to train AI" is becoming standard industry practice, not an Anthropic-only tactic
For developers and everyday users, this doesn't change how Claude performs. But it exposes just how messy large language model training data sourcing really is behind the "legally licensed data" language companies use publicly — and it's a reminder that court rulings on fair use are quietly redefining what the whole industry is allowed to do.
好不好用,試了才知道。
Sources / 資料來源
- Anthropic Knew the Public Would Be Disgusted by How It Was Destroying Physical Books (Futurism)
- Inside Project Panama, Anthropic's Secret Effort To Scan and Shred the World's Books (IBTimes UK)
- Is Anthropic Destroying Rare Books After Training AI Models On Them? (Yahoo Finance)
延伸閱讀 / Related Articles
- GitHub Bug Bounty砍半:AI假漏洞報告淹沒資安團隊 | GitHub Bug Bounty Payouts Slashed Amid AI Report Flood
- AI安全指數2026出爐:9大廠沒人及格、Anthropic奪冠僅C+ | AI Safety Index 2026: No Lab Tops C+, Anthropic Leads
- Gemini Omni免費影片生成倒數:10支AI影片明晚截止 | Gemini Omni Free Video Trial: 10 Clips, Deadline Tomorrow
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言