Lately, I've mostly been forgoing tech books and instead reading a lot about the philosophy of science, logic, and epistemology. My thinking is that in an age when AI can do so much of the routine work of software engineering, sharpening my critical thinking skills probably has a higher return than, say, getting better at programming in Rust. And frankly, it's a lot more interesting. Especially compared to Rust.
One side effect is that I've become more sensitive to questionable reasoning in the technology industry. For example, consider the supposed recent epidemic of software outages. Software engineers and the popular and tech press have been quick to blame AI-generated code, AI-related layoffs, cyberattacks, or some combination of all of the above.
It makes for a good story, and it may even change organizational behavior in some places. There's just one problem:
Nobody has proved that the software outage epidemic actually exists.
After reading books like Julia Galef’s The Scout Mindset and Nassim Nicholas Taleb’s Fooled by Randomness, I see an entire zoo of cognitive biases all over the place in recent software outage reporting. In my previous post, I pointed out how software engineers fall prey to motivated reasoning when they desperately — without a lot of actual evidence — want to blame outages on AI-generated code and AI-driven layoffs.
This week, let’s talk about this “outage epidemic” phenomenon.
As Taleb (incidentally, of The Black Swan fame) points out, humans are incredibly good at finding patterns in random events and constructing explanations for them. But if you consider the huge number of high-impact software systems running continuously in the world today, occasional “outbreaks” of outages are unavoidable. This is the clustering illusion in action, where people mistake patterns that can occur randomly for evidence of an underlying trend.
When I was doing some research for this post, I found out I’m not the only person this has occurred to this week. In Scientific American, SRE luminary Laura Nolan correctly pointed out that if there are 52 noteworthy outages every year, they probably won’t arrive neatly one per week.
You can actually run the math on this using the Poisson distribution. If you take Nolan’s hypothetical outage rate of one per week and assume they occur independently and randomly (ed. note: admittedly, this assumption doesn’t always hold), then the probability of exactly k outages in a given week is:
P(k) = (e^(-1) × 1^k) / k!
Let’s just set our outage epidemic freak-out level to five New York Times front-page (or maybe it’s just business-section, as there’s a lot going on in the world) worthy outages. To calculate the probability of five or more outages in a single week, we can subtract the sum of probabilities from [0..4] from 1:
P(5 or more) = 1 - e^(-1) × (1 + 1 + 1/2 + 1/6 + 1/24)
And you get approximately 0.00366, a.k.a. 0.366%.
That sounds pretty unlikely and maybe we should start blaming AI or hackers, until we account for how many opportunities there are for it to happen. And in fact, the expected “waiting time” to see such an “epidemic” is:
1 / 0.00366 = 273 weeks, or about 5.3 years
Over ten years, the probability of encountering at least one such awful week is:
1 - (1 - 0.00366)^520 = about 85%
In other words, even if the underlying outage rate never changes (and more on that in a moment), there’s an 85% chance of seeing at least one week with 5+ major outages over 10 years. All without any help from sloppy AI code, AI-enabled nation-state hackers, or understaffed SRE teams.
Of course, we don’t really know that 52 outages per year is the right base case. But that’s exactly the point:
Without a baseline and statistical model, claims that a software outage cluster is too unlikely to be random are just vibes, not solid engineering.
So far, we've established that when considering the rate of software outages, we haven't even established that the numerator is increasing. Sure, we have a collection of high-profile incidents, but no convincing evidence that they're occurring more frequently than they used to. (In fact, there's some evidence that incidents are declining at least in certain domains. The Uptime Institute's 2026 Annual Outage Analysis reports declining outage frequency per site going back years, while Parametrix's Cloud Outage Risk Report, published in March 2026, found that critical cloud outages declined in both number and duration in 2025.)
But there's another problem: Even if the numerator is increasing, we still need to consider the denominator:
There’s just more software — and more critical, high-profile systems — these days.
Even before AI presumably sped up everybody’s rates of delivery, the sheer volume of code running in the world was increasing at a prodigious rate. More software means more opportunities for failure. More opportunities for failure means more opportunities for the kinds of spectacular failures that (yes) end up in the New York Times.
An increase in absolute outages doesn’t necessarily point to a decrease in reliability or availability if the amount of functionality in operation is growing even faster. In fact, I’d argue that from an economic productivity standpoint, more useful functionality in production is exactly what everybody in the software engineering world should want.
None of this proves that AI is not harming software reliability, or there aren’t mean hackers doing bad things behind the scenes. That may well all be happening. But there’s a fundamental error in reasoning being replayed all over the news and social media, namely constructing causal explanations for a phenomenon that nobody has demonstrated actually exists.
There’s something highly ironic about software engineers and leaders demanding rigorous root cause analysis on one hand, while confidently explaining a “trend” they haven’t proved exists on the other. For an industry that claims to pride itself on data-driven decisions, there's surprisingly little data behind the software outage crisis narrative.