Great article, but the headline, "Anecdotes Everywhere, Evidence Almost Nowhere" is slightly flawed. Anecdotes are evidence, and they are particularly important in situations like this, where there is rapid change. You've correctly made the point at the end of the post that we can't wait for the evidence to be clear.
The Covid comparison is good in this regards. My background is in risk, and I found the "no evidence" mindset that was so common in the public discourse to be extremely stupid. I know you're not actually making that mistake; I just want to push back on the normalisation of the idea that anecdotes aren't evidence.
I realise that the whole tone of this article is 'remain uncertain, evidence is patchy' but I think you should take care relegating things like the vagueness of the 'economic growth' study/studies to footnotes! It's practically misinformation at this point.
Another uncertainty I have is around how fungible AI investment is, counterfactually. It seems like many growth stocks went down alongside AI stocks going up, presumably because growth investors rotated out. Same might go for other kinds of investment. So the connotation that AI investment is counterfactually driving growth isn't necessarily supported. I haven't seen anyone else making that point. It seems compelling to me. But I'm not an economist!
I have occasionally seen references to the idea that AI investment may be crowding out other forms of investment. It's a good point that this should be noted alongside the numbers regarding "contribution to GDP growth".
"The US unemployment rate has been drifting up for the last few years, but only modestly."
Any look at broad AI impacts on US employment might potentially benefit from taking into account the US labor participation rate, which has dropped from around 67% at the cusp of the 2020 pandemic to around 62% today. That's due to a variety of factors, including net retirement and more workers on disability:
The top-line US employment rate metric, U-3 – the rate that's recently been in the 4%+ range – uses the labor force as its base. If jobless people stop actively searching for work, for any reason, they leave the labor force. This drops the participation rate and shrinks the U-3 denominator. That paradoxically makes the unemployment rate lower (and hence look 'better'), because some workers have temporarily or permanently left the labor force and are no longer counted as "unemployed."
For that reason, it may be worth looking at the U-6 Total Unemployed Rate as an additional, broader measure of US employment trends that might also potentially surface some AI impacts.
Like U-3, the U-6 rate includes standard unemployed workers, but it also adds in "marginally attached workers." Those are people who want to work, but haven't looked for a job within the past four weeks: whether out of despair at finding work (on the part of "discouraged workers") or due to life circumstances like schooling, health, or family responsibilities. As well, U-6 also adds in "people working part-time for economic reasons (those who want full-time work but accept part-time hours due to slack business or cutbacks)":
The U-6 rate drifted up from 6.7% to 8.7% during the period from around mid-2023 to late 2025, until – after a stats discontinuity break due to a government shutdown – it's since fallen in recent months. That may or may not in part reflect AI-related impacts, yet it does paint a somewhat different picture than U-3, which moved from 3.5% to 4.4% over that same recent 1-1/2 year period.
There's a FRED page listing U-3, U-6, and other "alternative methods of labor underutilization." That page also offers a feature for generating custom graphs:
Great article, but the headline, "Anecdotes Everywhere, Evidence Almost Nowhere" is slightly flawed. Anecdotes are evidence, and they are particularly important in situations like this, where there is rapid change. You've correctly made the point at the end of the post that we can't wait for the evidence to be clear.
The Covid comparison is good in this regards. My background is in risk, and I found the "no evidence" mindset that was so common in the public discourse to be extremely stupid. I know you're not actually making that mistake; I just want to push back on the normalisation of the idea that anecdotes aren't evidence.
I realise that the whole tone of this article is 'remain uncertain, evidence is patchy' but I think you should take care relegating things like the vagueness of the 'economic growth' study/studies to footnotes! It's practically misinformation at this point.
Another uncertainty I have is around how fungible AI investment is, counterfactually. It seems like many growth stocks went down alongside AI stocks going up, presumably because growth investors rotated out. Same might go for other kinds of investment. So the connotation that AI investment is counterfactually driving growth isn't necessarily supported. I haven't seen anyone else making that point. It seems compelling to me. But I'm not an economist!
I have occasionally seen references to the idea that AI investment may be crowding out other forms of investment. It's a good point that this should be noted alongside the numbers regarding "contribution to GDP growth".
"The US unemployment rate has been drifting up for the last few years, but only modestly."
Any look at broad AI impacts on US employment might potentially benefit from taking into account the US labor participation rate, which has dropped from around 67% at the cusp of the 2020 pandemic to around 62% today. That's due to a variety of factors, including net retirement and more workers on disability:
https://fred.stlouisfed.org/series/CIVPART
The top-line US employment rate metric, U-3 – the rate that's recently been in the 4%+ range – uses the labor force as its base. If jobless people stop actively searching for work, for any reason, they leave the labor force. This drops the participation rate and shrinks the U-3 denominator. That paradoxically makes the unemployment rate lower (and hence look 'better'), because some workers have temporarily or permanently left the labor force and are no longer counted as "unemployed."
For that reason, it may be worth looking at the U-6 Total Unemployed Rate as an additional, broader measure of US employment trends that might also potentially surface some AI impacts.
Like U-3, the U-6 rate includes standard unemployed workers, but it also adds in "marginally attached workers." Those are people who want to work, but haven't looked for a job within the past four weeks: whether out of despair at finding work (on the part of "discouraged workers") or due to life circumstances like schooling, health, or family responsibilities. As well, U-6 also adds in "people working part-time for economic reasons (those who want full-time work but accept part-time hours due to slack business or cutbacks)":
https://fred.stlouisfed.org/series/U6RATE
The U-6 rate drifted up from 6.7% to 8.7% during the period from around mid-2023 to late 2025, until – after a stats discontinuity break due to a government shutdown – it's since fallen in recent months. That may or may not in part reflect AI-related impacts, yet it does paint a somewhat different picture than U-3, which moved from 3.5% to 4.4% over that same recent 1-1/2 year period.
There's a FRED page listing U-3, U-6, and other "alternative methods of labor underutilization." That page also offers a feature for generating custom graphs:
https://fred.stlouisfed.org/release/tables?eid=4773&rid=50