When the economy grows but AI is just one of many moving parts, how do you prove AI actually caused it?

This explores how we can tell whether growth in output is really AI's doing, as opposed to other things moving the economy at the same time. To be upfront: the corpus doesn't hold a methods toolkit for this, like natural experiments or comparing firms that adopted AI with firms that didn't. What it does hold is something arguably more useful for a newcomer. It maps the specific ways that attempts to pin down AI's contribution go wrong, and those failure points show where any clean measurement would have to look.

Start with the most public attempt. Brynjolfsson reads 2025's combination of slower job growth and continued GDP expansion as the signature of an AI productivity surge, the moment the long-predicted 'J-curve' finally turns upward Is AI productivity finally showing up in economic data?. The logic is simple: if output rises while labor input flattens, something is making each worker more productive. The weakness is just as simple. That pattern has many possible causes, and other economists point out that AI leaves no clear fingerprint in employment, productivity, or earnings data outside a handful of tech leaders. The isolation problem in a nutshell is that aggregate numbers can fit an AI story without proving one.

The next problem is that the evidence on the ground is itself distorted, in two opposite directions. Executives report AI productivity gains that are larger than what shows up in measured results, partly because operational improvements take time to reach revenue Do AI productivity gains feel larger than they actually measure?. Meanwhile, most workers who say AI saves them time actively hide or downplay their use, out of stigma and fear for their jobs Why do workers hide productivity gains from AI use?. Put those together and you get an unexpected result: you can't reliably attribute growth to AI when the people closest to the work are both overstating and concealing how much AI is involved. Survey-based attribution has a built-in bias problem before the economics even starts.

A third complication is the level you measure at. Narayanan and Kapoor argue that AI compresses the middle 'execute' layer of knowledge work, the drafting and producing, while the 'decide' and 'deliver' layers stay the same or grow Does AI really compress all layers of knowledge work equally?. Translation and legal work show stable or rising employment despite real AI gains. So a task-level speedup can be entirely genuine and still disappear into occupation-level statistics, because the freed-up time moves into other work instead of showing up as fewer workers or more measurable output. Anthropic's scenario modeling adds a distributional twist: even when AI-driven growth is real, it may show up mostly as rising capital share rather than broadly higher wages Does AI growth inevitably shift wealth away from workers?. Where you look for AI's contribution therefore shapes whether you find it.

The less obvious lesson comes from the corpus's more philosophical notes: 'output' may be the wrong thing to count. If AI produces context-specific tokens at the point of use rather than countable goods, quantity can rise while the value of each unit falls Is AI fundamentally changing how value gets produced?. Taken further, AI can produce knowledge faster than anyone can check it, inflating volume while reliability drops Can AI generate knowledge faster than humans can evaluate it?. On this view, isolating AI's contribution to output growth means first deciding whether more output counts as growth at all. If you want a model of when AI's contribution would stop being subtle and become unmistakable, the feedback-loop growth model shows the conditions under which automation's effects compound and stop being marginal When do AI feedback loops trigger explosive growth?.