Read any AI governance framework published this year and you'll meet a set of principles presented as hard-won new wisdom. A human must be able to monitor the system and intervene. The system's authority should be bounded in advance. Its decisions must be logged and reconstructable. Dangerous actions should be prevented, not just detected. And it should be tested beyond the happy path before anyone trusts it.
I agree with every word. I also recognize every word, because none of it is new. Those are the working rules of automation safety, written across other industries decades before anyone prompted a model. Every autopilot ever certified ships with a disconnect the pilot can always reach. Least privilege, the idea that a component gets the minimum access its task requires, was formalized in a security paper in 1975. Commercial aviation has required flight data recorders since the late 1950s, and process plants have kept historians for as long as they've had control rooms. Fail-safe design, relief valves, interlocks, watchdog timers, is the reason industrial accidents are the exception. And no chip I ever validated shipped on a happy-path demo. You qualified at the corners, worst case, or you didn't qualify.
So when AI governance "discovers" human oversight, bounded autonomy, audit trails, preventive guardrails, and range testing, it isn't inventing. It's rediscovering. Even the standards bodies quietly concede this: NIST is governing AI agents by writing overlays on its existing SP 800-53 control catalog rather than starting fresh, and Europe's aviation regulator is extending its existing development-assurance practices to AI instead of replacing them. The people with the deepest safety heritage are inheriting, not inventing.
Why old AI governance principles are good news
Here's why this matters beyond point-scoring, and why I'd want every leader who feels overwhelmed by AI governance to internalize it. If the principles are old, three things follow.
They're proven. You don't have to win a philosophical argument about whether bounded authority or preventive guardrails "work." A century of aviation and half a century of industrial control settled it. The burden of proof flips: the person who wants to skip the discipline is the one arguing against the record.
There's a playbook. Interlock design, fault trees, worst-case qualification, engineer-of-record accountability: these come with literatures, standards, and people who've spent careers applying them. You are not the first team to oversee an automated system that acts faster than a human can watch. You're the newest one.
And, most usefully, the genuinely new parts get sharp. Because some of this problem is new, and pretending it's all old is as lazy as pretending it's all new. Two things have no real prior art. First, the behavior isn't specified anymore, it's learned: nobody wrote the rule the model follows, so you can't enumerate its states the way you could a controller's, and testing can sample the space but never close it. Second, the system argues. Every automated system before this one failed silently or loudly. This one fails fluently, in confident prose that invites you to skip the check. I've written whole issues on that trap, and it is genuinely without precedent. Those two, the unspecifiable state space and the persuasive failure, are where the real invention budget belongs. Not on rediscovering the disconnect switch.
So the move for a leader is inheritance first. When someone proposes an AI control, ask what field solved this shape of problem already, and start from their answer. When a framework lands on your desk, read it as a translation of automation safety into new vocabulary, because mostly that's what it is, and translations are easy to audit for what got lost. And reserve your skepticism, your research, and your genuine invention for the two problems that have none of this heritage behind them.
Fields that keep rediscovering their own rules pay for them twice, once in the original accidents and once in the new ones. The discipline exists. Inherit it, and spend the invention where the problem is actually new.