AI companies are buying old books to avoid training on AI-generated slop
CackleRooster · reddit · 2026-07-21
A post highlights a claim that AI companies are buying lots of old printed books because they are free of AI-generated text.
The argument is that web data increasingly contains synthetic text, which may contaminate future training sets and contribute to “model collapse.” Books published before 2022 are presented as a cleaner source because they predate the recent flood of AI slop.
Related event: AI Companies Buy Old Books to Avoid Synthetic Data Pollution(2 posts)→
More from Research
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27