Notes from the Shore · 36
Is AI behind all these outages—or just taking the blame?
Yesterday afternoon I was boarding a United flight from Seattle to O’Hare when my phone started alerting me to an outage affecting the airline’s weight-and-balance system. Around me, people were complaining that it was probably AI. They had reached a diagnosis before we had left the gate, which is an increasingly normal experience of using technology in public.
I understood the suspicion. I suspect AI is already contributing to operational failures through decisions about how companies develop software and staff the teams that keep it running. But I cannot tell you that it caused this outage, and neither could anyone around me. That distinction matters if we’re going to get beyond passengers guessing and companies reassuring us.
The reporting on United’s outage describes pilots unable to obtain final aircraft-weight information needed for departure. United said flights already in the air and its United Express regional network were unaffected, and that there was “no indication that this was a cybersecurity-related issue.” The problem was resolved about an hour and a half after pilots began reporting it. Holding departures when crews cannot obtain the information they need is a safety precaution, not evidence that AI was flying the plane.
What does “not AI-related” rule out?
The suspicion extends well beyond my flight. Chris Hayes voiced a similar doubt on Bluesky, linking the Microsoft, FAA, and banking disruptions in a skeptical account of the official explanations. Those events didn’t all happen Monday: Microsoft’s disruption began Thursday, the FAA ground stop followed Friday, and banking problems and United’s outage occurred Monday. Their proximity doesn’t establish a common cause.
The explanations weren’t identical, either. Microsoft said its incident was “not related to cyber or AI / agents” and traced it to network changes. At Friday’s briefing, however, Homeland Security Secretary Markwayne Mullin declined to rule out AI involvement in the FAA failure. Treating all of this as one official, settled denial obscures what each organization has actually established.
Even if every outage had a different immediate cause, an airline can run conventional software while its developers use AI to write changes, its operations staff rely on AI to interpret alerts, or its managers leave jobs unfilled because they expect AI to make the remaining people more productive. Those are possible routes by which AI could affect reliability; I have no evidence that any of them explains United’s failure.
They are also different questions from whether someone attacked the network. I don’t think a cyberattack is the most plausible explanation here. I am more concerned about ordinary management decisions made with too much confidence in the tools. A company could accurately report that its failed application contains no AI while leaving those decisions entirely unexamined.
A system does not have to contain AI for confidence in AI to make it less reliable.
The staffing expectation is explicit elsewhere in industry. In his message to Amazon employees, Andy Jassy said he expected extensive AI use to reduce the company’s total corporate workforce over the following few years. That doesn’t prove understaffing or explain another company’s outage. It does establish that expectations about AI are entering workforce plans. I don’t want companies to count those savings before they know what the affected teams can safely do without.
There is research on the reliability question.
Google Cloud’s 2025 DORA report, State of AI-assisted Software Development, draws on a survey of nearly 5,000 technology professionals and more than 100 hours of qualitative research. Its authors reported that AI adoption was associated with better software delivery throughput and product performance, but continued to have a negative relationship with delivery stability. This is industry-backed research from a company selling AI, and an association rather than a finding that AI caused these outages. The combination is still worth taking seriously: producing more software and operating more reliably are not interchangeable achievements.
I use these tools and want their benefits. I also think confidence in AI-assisted development and reliability monitoring can outrun the evidence. If management increases the volume of changes while reducing experienced review, or treats an automated summary as an adequate investigation, a failure can have a perfectly conventional technical explanation and still involve a bad bet on AI.
Who gets to check the explanation?
That brings me to Garance Burke’s AP reporting on the federal AI-safety center, published November 8. The Center for Advancing Innovation and Standards for Super Intelligence, within the Commerce Department, lacked a permanent leader and had stopped publishing its research. The head of its parent institute was serving as acting director. AP also reported, citing two former officials, that staff had been asked to report informally to a White House office, raising concerns about the independence of its findings.
The distinction matters here too. The office wasn’t literally without leadership, and unpublished research isn’t necessarily research that has stopped. Nor is this center an airline-outage investigator. Its problems do not establish anything about United. What AP describes is a weakening public check on AI claims just as those claims are becoming harder to evaluate outside the companies making them.
I am skeptical that a company, in this economy, would volunteer the broadest possible account of how enthusiasm for AI affected a failure. Admitting a broken router is one thing; examining whether a staffing plan or a development target made the organization less resilient is considerably less comfortable. That incentive is a reason to require evidence, not a reason to assume a cover-up.
I want incident reviews to examine AI’s role in development and operations, with enough detail for independent investigators to test the explanation. I want government evaluators with stable leadership and findings the public can read. Companies deserve the chance to show that AI wasn’t involved, and passengers deserve something better than having to take their word for it. Until we can examine how these decisions affect reliability, I expect to keep hearing the same diagnosis at the gate.
The novel this note lives behind.
Into Alignment — a novel about an AI company, the agents inside it, and the humans deciding what they were for. Available in paperback, hardcover, and ebook.
Letters from the Shore
Get the next note by email.
Notes land on the blog first; the letters go out by email, a little longer and a little slower. Occasional; unsubscribe any time.