How to Read the Claim That AI Caused an Outage
"AI did it" arrived as an explanation for this month's failures faster than any investigation could have finished - from officials and executives alike. The claim can be literally true, badly wrong, or an evasion, and the wording usually tells you which.
When something large breaks, an explanation always arrives before an investigation could possibly have finished. This month the explanation has been AI: senior officials have linked the recent run of outages to it, executives have told legislators that automated patching was involved, and at least one company has said its own failure came down to an old bug instead. Those accounts cannot all be describing the same thing, which is the first clue about how to read them.
It helps to separate four distinct claims that all get spoken as "AI caused it."
The first is that a model wrote code or a configuration that shipped and was wrong. This is ordinary and plausible. Bad changes have always been the leading cause of outages; a model is a new way to author one, not a new category of failure. The fix is review policy.
The second is that automation approved the change - that the pipeline from authored to deployed ran with no human competent to stop it. This is a much more serious admission than the first, because it is about governance rather than authorship. The fix is gating, staged rollout and the ability to roll back in one step.
The third is that AI is the workload, not the author: inference and training traffic changed the load profile of a system sized for something else, and capacity gave out. This has nothing to do with generated code and everything to do with planning.
The fourth is AI as shorthand for "we now move faster than we can check." That is sometimes the truest description available, and it is also the one that names nobody and requires nothing in particular to change - which is precisely why it is attractive from a podium.
So apply a test to any explanation offered in the first week. Does it name a mechanism, a time and a change? "A configuration push at 14:12 UTC failed open in one region and the health check did not catch it" is a cause. "AI-driven patches" is a theme. Causes come with blast radius, duration and a specific remediation. Themes come with an adjective.
Then ask who the explanation serves. A vendor saying AI did it has traded a specific accusation of negligence for a diffuse one about the industry. An official saying it has a reason to want authority over something. A critic saying it has a case to make that predates the outage. None of this makes any of them wrong; it tells you what to verify rather than what to believe.
What is actually worth waiting for is the postmortem. It will be published days or weeks late, in deliberately flat prose, and it will name the change, the guardrail that was missing, and the time to detection. It is boring, specific and checkable. If the postmortem, when it lands, does say a model authored an unreviewed change that went to production, then the headline was right - and you will have the detail that makes it mean something, which the headline never had.
The same discipline applies elsewhere: separating what was reported from what was inferred is most of the work of reading technology news at all.
- There is a real difference between a model writing a bad change, an automated pipeline approving one, and AI being the workload that broke capacity; the three have different fixes.
- A cause named before the postmortem is a hypothesis, and often a convenient one.
- A credible explanation names a mechanism, a time and a specific change. "AI-driven" names none of those.