LLM copyright guardrail absurdity: refuses to scrape a book but happily zips it for you
burkov · x · 2026-10-23
ML researcher Burkov exposes twisted guardrail logic: the model refuses to copy a copyrighted book from a website, but happily downloads the entire book into a ZIP (35 HTML sections, 71 images) when asked — reasoning that "user-provided content" is fine to transform.
He argues this responsibility-shifting doesn't hold: the model knows the user's intent and the content's origin, yet treats a zip it just created as if the user supplied it. A second session where the model itself copies the book further shows the guardrails' inconsistency.
Related event: LLM Copyright Guardrails Easily Bypassed with a ZIP File(2 posts)→
More from Models
- Unverified chart claims GPT-6 Sol tops DeepSecBench Pareto frontier at 80% lower cost — JohnPhamous · 2026-10-23
- Developer: LLMs that constantly push back make work feel like arguing all day — louisvarge · 2026-10-23
- Gemini 4 Argon spotted testing apps autonomously, and 'topping humor benchmarks' — andrew_n_carr · 2026-10-23
- Gemini 4 reportedly far less sensitive in refusals than other models — haider1 · 2026-10-23
- SemiAnalysis: Anthropic subscriptions offer 5x+ more value than OpenAI as $200 plan value halves — RexDouglass · 2026-10-23
- Google Cloud's Physical-Engineering Last Exam: all frontier models score under 50% — Afinetheorem · 2026-10-23