How to deliberately bait copyrighted songs and visuals out of video models
nptacek · x · 2026-09-14
A thread details how copyrighted material can be deliberately extracted from video generation models: packing prompts with high-density semantics from a music video (BPM, key, specific instruments) plus aesthetic cues (e.g., blue color grading) can pull popular copyrighted songs or visuals — often not the intended one. This happens because the model's text filter detects the intent and rewrites the input, but residual semantics leak through. A quoted reply adds the inverse lesson: generic prompts like "1970s retro sci-fi" keep regenerating xenomorphs and the Weyland-Yutani logo; building your world's visual rules from the ground up is the only escape from training-data style contamination.
More from Models
- DeepSeek V4.1 Flash takes first on Artificial Analysis' new v4.3 benchmark, dethroning Astra — Randomdotmath · 2026-09-14
- Musk says 2.5T-parameter Grok 4.8, trained on new C++ stack, finishes training this week — inductionheads · 2026-09-14
- Ethan Mollick: data-sovereignty small models struggle to keep up with frontier AI — emollick · 2026-09-14
- Attention paper author Bahdanau: paying for Gemini, 'forgot AI can be this bad' — DBahdanau · 2026-09-14
- Users complain OpenAI quietly tightened limits: '$200 subscription feels like a free trial' — CtrlAltDwayne · 2026-09-14
- User grants GPT-6 agent full access in one click to manage his emails, bills and schedule — soumitrashukla9 · 2026-09-14