If AI Trains Mostly on AI Text, Where Does New Knowledge Come From?

As AI writes more of the internet, training data becomes self-referential and loses genuine novelty. The fix is to detect and preserve new ideas, then turn live, validated real-world context into the new engine of learning. MCP can be understood as “AI’s senses” for real-world validation and discovery. Using novelty-specialist models, curator systems, and reality-testing loops via MCP and audit logs, we can harness entropy productively.