Writing · AI and human responsibility
Outages Aren't a Conspiracy. They're the Bill Coming Due.
By Richard K. Marshall ·
The last couple of weeks produced a familiar pattern. Microsoft 365 went down. The FAA grounded flights for hours. A major bank had problems. Within days the usual voices were calling it a coordinated test, a kill-switch preview, or the start of something larger. Most of that is noise. The more useful explanation is simpler and less comforting: we built a lot of critical systems on a set of assumptions that no longer hold, and we are now running them in an environment those assumptions never contemplated.
For a long time the dominant rule in infrastructure was "if it ain't broke, don't fix it." That wasn't laziness. It was risk management. When a change could take down air traffic, payments, or core communications, the safe move was to leave working systems alone. Downtime was scheduled. People were warned. Testing was heavy. The NASA culture of the Apollo era is the clearest example: exhaustive review, redundancy, and a deep bias against surprises. That approach kept complex systems running, but it also left a lot of technical debt and hidden fragility in place. The systems worked well enough, so the underlying risks stayed invisible.
That world is ending. The operating model that has produced the most progress in the last fifteen years is closer to SpaceX than to NASA. Ship, break, learn, fix, repeat. Accept temporary failure as the price of speed. The results in rocketry and software have been dramatic. The problem is that much of the world's critical infrastructure was never designed for that cadence. A lot of it is still running on older codebases, older operational habits, and incomplete documentation. When you start changing those systems at modern speed, the seams show.
AI accelerates both the exposure and the repair. It is very good at surfacing patterns and failure modes that used to be too expensive or too obscure to examine. It is also very good at introducing new dependencies and new ways for systems to interact in ways no single person fully understands. The combination means more risks become visible at the same time that the rate of change is rising. Some of those risks were always there. We simply lacked the tools, the incentives, or the willingness to look. Ignorance was manageable when change was slow. It is not manageable when change is continuous.
The recent outages fit this picture better than they fit a master plot. A router upgrade. A capacity change. A software bug. Each one is ordinary on its own. What feels new is the frequency and the visibility. We have moved from a world in which most failures could be contained and scheduled to a world in which failures are more likely to be public and simultaneous. That does not require coordination. It requires tightly coupled systems, higher change velocity, and better detection. The same tools that let us find problems faster also make the problems harder to hide.
None of this is an argument against speed. The old model left too much risk unexamined and too many improvements unmade. The new model will continue to produce real progress. It will also continue to produce public failures until the people who operate critical systems get better at limiting blast radius and recovering quickly. That is an engineering and operational problem, not a secret one. The organizations that treat reliability as a design constraint rather than an after-the-fact apology will pull ahead. Everyone else will keep supplying fresh material for the next round of theories.
The practical point is straightforward. A lot of what used to be invisible is now visible because the conditions that kept it invisible have changed. That is uncomfortable. It is also useful. The systems were never as solid as the lack of outages made them appear. We are simply no longer in a position to pretend otherwise.