This Week in "Putting AI to Work" (4/10/26)
Growing gaps between the US and Europe, a VC's view of the state of enterprise AI, a Manhattan Project for... the equivalent of supermarket scanner data?, and other interesting stuff
Who’s Using AI, and Is That AI Making a Difference?
Two new pieces of research here. First is a report on U.S. versus European AI use. As the graph above shows, this topic is important. European productivity growth has lagged badly in the 21st century, and AI might be able to help close that gap. As European competitiveness czar Mario Draghi hopefully put it in 2024, “has opened [a window] for Europe to redress its failings in innovation and productivity”
But the new report shows that U.S. workers use AI far more than their European counterparts: “5.2% of work hours in the US are spent using AI, while the share in Germany, France, and Italy is less than one-third of this.”
The researchers dug in on a key question and a puzzling one: “why a technology that is available at similar prices across countries and requires only minor, if any, upfront investments is used at such different rates[?]” They found that about half of the gap could be explained by the fact that U.S. workers are on average younger and better educated (and hence more likely to use AI).
But management practices matter a lot too. This is always the case, as excellent previous research has established. Look how strong the correlation is between good management practices and AI adoption:
The authors also found in both the U.S. and Europe, sectors that used AI more saw higher productivity growth between 2019 and 2025.1 But they didn’t find those higher productivity growth sectors needed fewer workers: “For both Europe and the US, we find no robust evidence that AI adoption is associated with employment expansion or contraction at the industry level in recent years.”
Second, Andreessen Horowitz partner Kimberly Tan published a report on “Where Enterprise AI is Actually Working.” She found that.
29% of the Fortune 500 and ~19% of the Global 2000 are live, paying customers of a leading AI startup…
To qualify for this statistic, these enterprises had to have signed a top-down contract with an AI startup, successfully converted a pilot, and have gone live with the product in their organization.
This level of penetration in such a short period of time is remarkable since Fortune 500 enterprises are not known to be early adopters of technology.
To estimate where the Fortune 500 is putting AI to work internally, the report takes the interesting approach of looking at what kinds of AI startups they’re working with. Some startups focus on coding; others on customer service, and so on. The conclusion is that “On the revenue momentum, enterprise adoption of AI is dominated by a clear set of use cases and industries. Coding, support, and search represent the lion’s share of use cases by far (with coding being an order-of-magnitude outlier even among this set), while the tech, legal, and healthcare sectors have been the industries most eager to adopt AI.”
The a16z report also stresses that “model capabilities are improving fast. There are several domains that have shown dramatic improvements in the last 4 months — with accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement. We expect these jumps to yield compelling new products and companies in their relevant domains.”
AI Gets Scientists Out of Their Rut
There is debate at present about whether AI leads to innovation and exploration or standardization and route automation. A new paper makes an important contribution to this debate and an encouraging one.2 The researchers found that Google DeepMind’s AlphaFold, which accurately predicts the structures of proteins, reversed a declining trend in research on novel proteins — ones whose structures had not been precisely mapped:
The authors find deep and broad impacts of AlphaFold:
we show that a long-running decline in the study of novel proteins halted after AlphaFold2’s release, with the shift concentrated among studies citing AlphaFold2 and targets with high-confidence predictions. This pattern extends to 248,191 downstream papers that consume structural knowledge, where engagement with genes lacking experimental structures and with understudied human genes increased since 2021.
They also weigh in on the “does AI lead to novelty or perpetuation of the status quo?” debate:
Amid rising concern that AI may reinforce scientific canons7–10, our findings offer an early field-level case where AI predictions expand scientific frontiers… These results suggest AI can complement human knowledge and redirect collective attention in science, with broad implications for emerging AI for science models.
Things are Not Going to Calm Down Any Time Soon
Over at Marginal Revolution, Alex Tabarrok highlights two mirror-image features of powerful agentic AI like Anthropic’s new Mythos: they can do a great deal of damage automatically, and they can be easily damaged by diverse stratagems. Tabarrok:
Two new papers/initiatives indicate severe risks from AI, interestingly in opposite directions. The first is that the most advanced frontier models are now capable of finding and exploiting software in ways that could be used to crash or control pretty much all the world’s major systems…
The second paper on AI risks is AI Agent Traps from Google DeepMind. They point out that AI agents on the web are vulnerable to all kinds of attacks from things like text in html never read by humans, hidden commands in pdfs, commands encoded in the pixels of images using steganography and so forth.
I like Alex’s conclusion:
Putting this together we have the worrying combination that very powerful AI’s are very vulnerable. Will AI solve the problems of AI? Eventually the software will be made secure but weird things happen in arms races and its going to be a bump ride.
It is in fact going to be a bump[y] ride.
AI Job-Pocalypse Watch
Massive AI-caused technological unemployment continues to not show up in the employment statistics, even for the most AI-exposed jobs like writing code. Business Insider finds that "Tech job openings rebounded sharply in 2026, challenging popular narrative that AI is wiping out engineering roles...more than 67,000 software eng job openings, highest level in 3 years. Listings have doubled since a trough in mid-2023."
On X, New York Times economics reporter Ben Casselman demonstrated how easily it is to draw conclusions about AI’s employment impact that are both evidence-based and wrong. He posted a series of charts that show substantial employment declines in lots of sectors beginning right around the release of ChatGPT:
Then he revealed that he sampled on the dependent variable: He went looking for charts with that shape. And that shape could appear for all kinds of reasons; As Casselman put it, “some of these are almost certainly examples of industries that were hit by rate hikes, which began at more or less the same time. Basically everything housing-related turns down starting in 2022.”
Casselman then posted another series of employment charts from “some of the industries that have gotten talked about the most in the AI conversation.” These do not show a downward kink in the ChatGPT era:
As Casselman summarizes, “Bottom line: Be skeptical of drawing conclusions about the effect of AI (or anything else!) based simply on time series charts. And be especially skeptical when the examples are cherry picked, not selected in some more rigorous way.”
Economist Alex Imas has a good idea for how to get better clarity on AI’s impact on jobs and wages: let’s start systematically collecting exactly the data we need. As James O’Donnell explains in Technology Review, we lack good data on
the most pressing economic question of our time: the specifics of price elasticity, or how much demand for something changes when its price changes. And this is the second part of what Imas emphasized last week: We don’t currently have this data across the economy. But we could.
We do have the numbers for grocery items like cereal and milk, Imas says, because the University of Chicago partners with supermarkets to get data from their price scanners. But we don’t have such figures for tutors or web developers or dietitians (all jobs found to have “exposure” to AI, by the way). Or at least not in a way that’s been widely compiled or made accessible to researchers; sometimes it’s scattered across private companies or consultancies…
“We need, like, a Manhattan Project to collect this,” Imas says. And we don’t need it just for jobs that could obviously be affected by AI now: “Fields that are not exposed now will become exposed in the future, so you just want to track these statistics across the entire economy.”
Getting all this information would take time and money, but Imas makes the case that it’s worth it; it would give economists the first realistic look at how our AI-enabled future could unfold and give policymakers a shot at making a plan for it.
Oh My God
Let’s give the final word of this week to developer Udi Wertheimer, who makes a reassuring observation about (the lack of) runaway AI to date:
anthropic has been saying for years that their models “scare them”, try to escape, exhibit self-awareness
we now have open source uncensored opus-4.5-level models and none of them are self aware, trying to escape, or stealing nuclear codes
but yeah i’m sure this time it’s real
The responses to his post included this cartoon, which I must share:
And Now a Word from Our Sponsor
This weekly roundup is brought to you by Workhelix, the startup I cofounded to help organizations know and grow the ROI of their AI. If that topic is top of mind for you, please get in touch.
They also stress, though, that they’re only observing correlation between AI use and productivity growth, not causation
Encouraging if you like innovation, that is. And maybe you don’t. Maybe you just want to pave the cowpaths, not cure Alzheimer’s, not unlock the limitless free energy of nuclear fusion, and so on.