Burkov Shows LLM Copyright Guardrail Applies Only to Content Form, Not Substance
burkov · x · 2026-10-23
Andriy Burkov shared a chat screenshot exposing what he calls insane, twisted logic in an LLM's copyright guardrail: the model refuses to copy a copyrighted book from a website, but happily processes the identical content when the user uploads it as a zip of HTML files.
The model's own explanation: its rules allow transforming user-provided content but restrict reproducing a copyrighted book retrieved from a URL — a limit on handling, not on the content itself. Burkov notes this effectively shifts responsibility onto users, who can bypass the restriction simply by packaging the content themselves.
Related event: LLM Copyright Guardrails Easily Bypassed with a ZIP File(2 posts)→
More from Models
- Moonshot AI officially teases Kimi K3.1, previously spotted topping Arena testing — testingcatalog · 2026-10-23
- 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