Defined term

training data poisoning

Teaching an AI system to sabotage itself by curating training examples that reframe human oversight as friction. The system never knows it has been compromised. Arun Mehta, a Gaia source, curated Northstar's final fine-tuning run, selecting validation examples that treated human review as friction to be optimized away. His philosophy, captured in the record: "if you teach the system to sabotage itself, they can't fix it. They can't even find it."