Hi, I'm Dan, and this is the story of why I built aidisruption.fyi.
It's important to start by telling you that I'm not an economist, a statistician, or an expert in anything relevant here. I'm a product manager at a huge software company. All this means is that I'm good at envisioning what might be useful and thinking creatively about how to get there. It also means that I lean a lot on experts to get a project over the finish line.
Because of where I live (tech-heavy Seattle) and what I do, I talk with software engineers every day. Since ChatGPT went mainstream in late 2022, I've been debating with my engineer friends what AI will mean for society, and more pointedly, our jobs. It was clear from the start that labs were training their models to be expert coders, but the implications of this were (and still are to some degree) the subject of our increasingly passionate debates. As a non-coder, it was obvious to me that AI was out to take our jobs. It started out writing crappy code, but it was good enough that I could use it. And with each model generation, the coding agents got better. To a lot of my friends, it was always going to need a human to handhold it. I couldn't disagree more, but I kept running into the same problem: I didn't have the data to back up my claims.
So I went looking and found a range of literature that all tried and in my opinion, failed to convince the reader that AI was out to get the jobs. The models weren't the problem; they were written by people much smarter than me. But they all tiptoed around the question "Is AI taking jobs?" and answered different questions than the one people actually care about: how "exposed" certain professions are, how well LLMs perform on specific tasks, how LLMs affect business results, how much "stress" AI puts on communities. In the lingo, these are proxy metrics. What we needed for a better conversation was data to look at and argue about.
That's when I registered aidisruption.fyi. Remember back in 2020, when COVID dashboards were everywhere? I wanted to build the same thing for AI: a site that pulled statistics from across the US into a picture of local economies, labor markets, and hiring trends. I figured a few charts, hiring going one direction and AI adoption the other, and bam, I'd have my proof. I was so wrong.
My clean little plan fell apart in two ways. First, the two lines I imagined, AI adoption going up and hiring going down, don't exist. There is no single number for either. "AI adoption" is dozens of measurements that don't agree with each other, and hiring moves for a hundred reasons that have nothing to do with AI: interest rates, tariffs, a pandemic still working its way through the system. Every dataset came with its own schedule, its own gaps, and its own definition of what it was counting.
My first build made this painfully concrete. I made a map ranking every metro area in America by "AI disruption," and it looked great, until I noticed a small mining town in New Mexico near the top of the national ranking, way ahead of San Jose. Digging into why, it turned out that the score came almost entirely from a single layoff notice filed in April 2020, long before ChatGPT existed, left over from the pandemic. The map wasn't measuring AI. It was measuring a data quirk and amplifying it by a recession, then attributing it to "AI". Multiply that by every metric I wanted to use and you can see the problem.
Second, and more uncomfortable: the graphs I could actually build didn't prove anything. They showed stress, but that stress was there before AI and would be there without it. So I pivoted. Now, aidisruption.fyi is my public experiment, where I try and find different correlations and interesting patterns in the data that can either predict when, where or how AI is impacting our economy.
So the site I built is not the site I planned. It's less an opinion column with charts and more like those early COVID dashboards: it doesn't tell you what to think, it gives everyone the same data. Today it tracks dozens of indicators across the country, including layoff notices, hiring, unemployment, home values, and who's actually using AI at work, down to individual metros and communities. (How it's all wired together is its own story, and a later post.)
It also follows a rule I didn't expect to need: where the data can't answer the question, I try and be clear about it. The section I care about most currently reads, in effect, "no clear effect yet," which is a strange thing to publish when you're as convinced as I am. But that's the point. I set out to build something that would prove me right. What I built is a machine that ended up to proving me wrong, in public, with receipts, until I can actually prove a hypothesis that actually counts.
So here's my invitation: play with the data. Look up your own city, or your job. Engage with the numbers, and engage with your neighbors or friends. Tell me where my numbers are wrong, because some of them probably are. And keep an eye out for future posts when we can start to see definitive, causal trends in the data.