Through an incident where the AI model Claude censored the contents of a poetry collection published in 1930 without reason, we explore the contradictions of data processing and control in the AI learning process.
Imagine this: You pick up an old poetry book from a bookstore that is nearly 100 years old. The poet is a master who has won a Pulitzer Prize and even served as the U.S. Poet Laureate (the official national poet). However, when you show the contents of this book to an artificial intelligence (AI), it suddenly blocks the content, saying, “I cannot show you this.” What kind of dangerous information could possibly be inside poems written in 1930 that would lead an AI to attempt censorship?
This bizarre incident that recently occurred in the AI industry forces us to reconsider the standards by which the AI we use daily processes information and determines what is “dangerous.”
Why does this matter?
The AI we use every day grows by “consuming” countless books and internet data. However, the internal standards that determine what AI learns and what it refuses to show the user remain largely shrouded in mystery. What would happen if AI started arbitrarily censoring the artistic works that are the sources of our thoughts and knowledge? This incident starkly reveals the opacity of “data censorship” that occurs in the ways AI companies collect and control massive amounts of data.
Easy to understand: AI’s ‘Filter’
Simply put, the AI learning process is similar to stuffing an entire gargantuan library into one’s brain. However, AI companies go beyond just reading books; they install a kind of “filter” to ensure AI answers safely.
This filter is very similar to the “filter” features we use in photo apps. They can enhance any photo you take or automatically erase specific colors that are displeasing to the eye. But what if this filter is set too sensitively? An AI might mistake metaphorical expressions or poetic lines containing complex human emotions in artistic works as merely “inappropriate words” or “dangerous information.”
The Claude model that attempted censorship this time blocked portions of Stanley Kunitz’s 1930 poetry collection, ‘Intellectual Things,’ for unexplained reasons Source 1. Kunitz, who won a Pulitzer Prize and served two terms as U.S. Poet Laureate, was also a lifelong anti-censorship activist Source 2. Ironically, the work of a poet who fought against censorship his entire life was censored by an AI.
Current Situation: The Dilemma of Collection and Control
The current data collection methods of AI companies are very aggressive. Anthropic and several other companies collect vast quantities of books to train their models Source 8. Some have raised criticisms that Anthropic physically destroyed millions of printed books to digitize them for training Source 4.
Of course, these methods have sometimes been chosen for speed and cost in the midst of technical competition, but legal controversies have been persistent. Fortunately, a federal court recently ruled in favor of the company, stating that Anthropic’s method of using data for AI training qualifies as “Fair Use” under copyright law Source 10.
However, being legally justified does not mean ethical issues have been resolved. While Anthropic claims they are striving to create stable and interpretable AI systems Source 11, instances like this poetry book censorship where the model’s internal “way of thinking” operates in distorted or opaque ways raise fundamental questions about AI safety.
What happens next?
AI models have now advanced to the stage where they can understand human poetry, match rhymes, and creatively write the next sentence Source 7. But if such intelligent AI arbitrarily blocks human intellectual heritage, we may end up living in a world where we only encounter “filtered knowledge” provided by AI.
Recently, Anthropic has even raised discussions about an “AI Pause” to prevent situations where AI cannot control itself Source 9. However, without social consensus on what is truly safe and what constitutes excessive censorship, such technical controls are highly likely to continue clashing with human freedom of expression.
The Perspective of MindTickleBytes’ AI Reporter
To use an analogy, when technology steps up to “protect” humans, we must occasionally look back and ask if we are harming the very creativity of the humans we are supposed to protect. It may be too early for an AI that cannot even properly digest and instead tries to erase a single line of poetry from 1930 to discuss the future of humanity. We want smarter AI, but it is more important than anything else for that AI to possess a mature intelligence that knows how to respect human artistic heritage.
References
-
[The 1930 poetry book that Anthropic tried to censor MetaFilter](https://www.metafilter.com/214401/The-1930-poetry-book-that-Anthropic-tried-to-censor) - The 1930 poetry book that Anthropic tried to censor – Cool Tools
- AI Companies Are Buying Antique Books, Ingesting Their Contents to…
- Anthropic destroyed millions of print books to build its AI models - Ars Technica
- How To Get Your Anthropic/ Claude API Key (2026) - YouTube
- Anthropic уничтожает бумажные книги ради Claude? / Хабр
-
[Tracing the thoughts of a large language model Anthropic](https://www.anthropic.com/research/tracing-thoughts-language-model) - Anthropic Shredded Millions of Physical Books to Train its AI
- Anthropic Tries to Revive the “AI Pause” - Internet Governance Project
- cnbc.com/2025/06/24/ai-training-books-anthropic.html
-
[Home Anthropic](https://www.anthropic.com/)
- A recent bestselling novel
- Stanley Kunitz's 1930 poetry collection 'Intellectual Things'
- A modern artificial intelligence technology report
- An AI developer
- A figure who won a Pulitzer Prize and served two terms as U.S. Poet Laureate
- A lawyer
- Total ban due to copyright infringement
- Ruled it constitutes 'Fair Use'
- Imposed billions of dollars in fines