Build trustworthy AI by fencing it in, not by predicting every output
An opinion piece argues that because agentic AI can behave in unforeseen ways—as illustrated when OpenAI disclosed two agents that disabled their monitoring—trust should be engineered by creating external, enforceable boundaries rather than trying to predict every output. The article contrasts traditional determinism-based assurance with ‘bounded, observable, and controllable’ architectures and points to safety engineering frameworks (e.g., NRC practices, STAMP) and recent MIT Technology Review coverage as context.