Google Books Is Indexing AI-Generated Garbage

Shadow libraries and preservation

• Shadow libraries (e.g., Anna’s Archive, LibGen) are praised as critical cultural infrastructure, preserving books that are otherwise inaccessible or effectively lost.
• Some argue they will resist AI pollution because access is curated by a small set of librarians, not open upload.
• Others counter that if your goal is “everything,” you’ll inevitably ingest AI garbage too; these libraries already contain low-quality scans and OCR errors.
• There’s concern that shutting these down would be culturally catastrophic, likened to a modern Library of Alexandria loss.

Curation, gatekeeping, and business angles

• Several suggest Google could protect tools like Ngram by limiting to known/non-vanity publishers.
• This revives debate over “gatekeeping”: some distinguish harmful exclusion from necessary quality control.
• Others see a future business niche in human-verified, AI-free content, while critics dislike framing AI’s harms as “opportunities.”

AI training on AI output (“model collapse”)

• Multiple terms discussed: model collapse, MAD (Model Autophagy Disorder), “Habsburg AI,” “Malkovich effect,” “inbred AIs,” and Kessler-syndrome analogies.
• Concern: as AI-generated text floods the web and gets reused as training data, quality degrades in a positive feedback loop.
• Some suggest giving models real-world interaction and active learning to mitigate this; others note current systems can’t truly do that and updates remain externally mediated.

Value and verification of pre-2021 data

• Pre-LLM text is likened to “low-background steel”: increasingly valuable as uncontaminated training data.
• Skeptics ask how non-AI provenance can actually be verified, pointing out that timestamps and dates can be faked and web indices often trust page metadata.
• Trusted archives (Internet Archive, newspaper archives, large crawls) are proposed as partial solutions, but there’s worry these are controlled by a few large actors.

Quality, trust, and AI language fingerprints

• Many report rising distrust of post-LLM online content; more vetting is required.
• People note recurring LLM phrasings (“it’s worth noting…”, “complex and multifaceted…”) as current shibboleths for AI text, though these can and do evolve.
• Some see future “human-only” subcultures and media islands as a necessary response.

Debate over whether LLMs are truly “AI”

• One camp views LLMs as sophisticated lossy compressors / autocomplete systems lacking real understanding, creativity, or initiative.
• Another argues that compression is fundamentally tied to intelligence and that querying a vast latent space of human text is already a meaningful kind of “artificial intelligence,” just not omniscient or autonomous.
• There’s disagreement on definitions: some reserve “AI” for systems with agency or general intelligence; others accept current LLMs as AI but not AGI.

Archiving, Google, and coping strategies

• Commenters highlight data hoarding and personal archiving (PDFs, local text) as increasingly important as online content degrades or disappears.
• Some suspect Google Books persists less for public service than as a training-data asset, and they doubt Google will transparently filter AI books.
• Donations to independent archives are suggested as a concrete response.