Why AI Isn’t Actually Boosting Productivity | Trumponomics

Why aren’t we seeing the productivity boom the artificial intelligence industry has promised? Stephanie Flanders is joined by Oxford University professor and How Progress Ends author Carl Benedikt Frey to explore why rapid advances in AI haven’t yet translated into stronger economic growth. Together they examine the productivity struggle, the race between the US and China for AI leadership and what history teaches us about technological revolutions.

TL;DR

A ~29-minute episode of Bloomberg’s Trumponomics on the Bloomberg Podcasts channel, published 5 August 2026 — Stephanie Flanders (head of government and economics at Bloomberg) interviewing Carl Benedikt Frey, associate professor at Oxford and author of How Progress Ends: Technology, Innovation and the Fate of Nations. Load-bearing claims:

1. The book’s thesis: institutions have to change phase, and that is what makes progress hard to sustain. Growth by adopting and scaling technology invented elsewhere works for a long time — the Soviet Union did it successfully for four decades, taking advantage of the Ford Motor Company’s open-door policy to build its own vehicles industry. “In the age of mass production, the Soviet system actually worked fairly well, because when technology is mature, when production is fairly standardized, then you can hold factory managers accountable essentially just by benchmarking performance.” When mass production’s returns petered out globally in the 1970s, that system stopped working: novelty cannot be benchmarked, and the computer revolution — “to which Soviet contributions were essentially none” — required decentralised exploration. His illustration of the difference: a Soviet engineer refused by the Red Army had two or three other options, and then “your idea would die with you”; whereas Bessemer Venture Partners declining to invest in Google in 1999 did not end Google, because others stepped in — and the decline itself “illustrates that Google was not a safe bet at the time. AltaVista and Yahoo, they were dominating search.” The pattern: exploration needs decentralisation, scaling needs consolidation, and the transitions between the two are where progress breaks.
2. Progress is unnatural, not inevitable. “If progress was inevitable, it would not have taken 200,000 years to have an industrial revolution. If progress was inevitable, most of the world would be rich and prosperous. If progress was inevitable, Britain, the country where I live, would not have suffered two decades of productivity stagnation.”
3. Inventive output is up; transformational output is down. “If you look at patenting, if you look at scientific publications, any measure of inventive output — all those indicators are up. Yet if you look at the economy, that’s pointing in a very different direction, and even measures of research productivity and breakthrough innovation are down. So we’re getting more in terms of scientific and inventive output, but it seems that we’re getting less transformational output.”
4. The episode’s most transferable idea — the verification tax. Frey expects AI to show up in the productivity statistics eventually, but argues it will deliver less of a boost than the computer revolution did, for a structural reason: “with AI you still need verification. AI automates a lot of knowledge work, but at the end of that you need to verify the output. And so it’s the time saving minus the time for verification.” The contrast he draws with the prior wave is precise: “I could sit around and wait for hours, even weeks, for new material to arrive for me to start my research, and the internet gave me access to that instantaneously. And so it automated downtime. AI is not automating that downtime. It’s automating the production, and you still need verification as well.”
5. The historical control on the optimistic case. “In many ways the computer revolution was more transformative than the AI we have today. The computer and the internet connected the best scientists and inventors around the world. It streamlined the research process enormously. It gave us access to the world’s store of knowledge essentially in our pockets. And what do we get out of that? Basically a decade-long productivity upsurge mostly confined to the United States.”
6. China is less centralised than assumed, and its industrial policy can be pro-competitive. Unlike the Soviet Union, where every industry was managed centrally from Moscow, “the Chinese economy is much more decentralized, and provincial governors and mayors have much greater autonomy,” competing against each other on growth targets for promotion inside the one-party system — “essentially a tournament of political competition that is very hard to replicate anywhere else.” The comparative point: “where Europe does industrial policy, it often tends to be anti-competitive — German rearmament essentially means plowing funds into Rheinmetall. In China, industrial policy can often be pro-competitive because you have provinces seeding new firms that are competing against each other.” He also notes that China’s innovation leaders are mostly startups, privately and often foreign funded — “in that sense China is not that different from either Europe and the United States. What is different is that it doesn’t have the rule of law,” so political connections matter more. And the convergence has run the other way from what was expected: “what people believed in the 2000s [was] that China would become more like the United States; the opposite seems to be happening, and the United States today looks more like the political capitalism that you have in China” — his example being OpenAI’s openness to the government taking a 5% stake.
7. Vested interests as the recurring brake — and the reason business dynamism is falling. “The leaders in bicycles didn’t become leaders in automotive, although they tried. Legacy media companies did not lead the social media revolution. The legacy car companies did not lead in electric vehicles. The legacy retailers did not lead in e-commerce… and old industries have every incentive to prevent that sort of competition.” The named mechanisms: killer acquisitions (“incumbents buy up promising startups just to shut them down”) and a revolving door between the US Patent and Trademark Office and incumbents, “whereby patent examiners grant them low-quality patents and then take on jobs for these firms in return.” The result is the paradox at the centre of his argument: “it helps explain why we’ve seen a decline in business dynamism despite the fact that every technology — from the personal computer, the internet, the cloud, and now AI — will have made it much cheaper to set up a company and operate a firm. And yet we’re seeing less entry.” In China the same barriers operate, plus a shift in CCP priorities over ~15 years from economic targets toward self-sufficiency, national security and common prosperity, which “creates a greater reliance on state-owned enterprises… and by any measure, state-owned enterprises in China have been less innovative and less productive.”
8. Against scale-is-all — the AlphaGo reversal and the data-efficiency argument. Flanders puts the strongest counter-case: scale seems to matter more than anything, so gains accrue to whoever has the most energy and chips, and new entrants cannot afford the fixed costs. Frey’s answer: “if the world was just a static distribution of events, you could probably brute-force things… But the world is not just a static distribution of events. It’s changing all the time.” His evidence is a fact he says few people know: AlphaGo beat Lee Sedol in 2016, “yet just a couple of years ago human amateurs using standard computers beat the best available Go programs quite easily, by exposing them to new positions, new concepts that they would not have encountered in training. And so that raises a fundamental question — even in cases where we achieve superhuman performance, we cannot be sure if that’s actually going to be true tomorrow when circumstances change.” Humans “are capable of learning from just a few examples; we are very, very data efficient,” and getting there “will need some innovation.”
9. “AI is still waiting for its separate condenser moment.” The path forward is genuinely open — “it may be that large language models is the future of AI. It may be that small language models. It might be world models. It might be something entirely different.” The analogy: “during the first industrial revolution, early on, steam engines were tremendously energy inefficient. They were basically just used to drain coal mines. It took the separate condenser to make them energy efficient, for them to be applied to transportation later on — railroads and steamships. And I think AI is still waiting for that moment, and that’s a question of further innovation, not just scaling.”
10. Adoption for what — the question that matters more than who is three months ahead. Unless a firm “gets onto a curve where it really just pulls away from the rest,” Frey does not think it matters whether the US or China, or OpenAI or Anthropic or Google, is three months ahead. “And then there’s a question of adoption. And obviously at the end of the day, the use of a technology is what drives productivity. But there also the question is adoption for what? If people adopt it for email, it’s not going to drive growth in any meaningful way — but it may look good in the sense that the adoption rate is high. But if people adopt it for scientific research in ways that develop new products and new technologies, that’s obviously a different matter.”
11. Drill deeper versus drill more holes — a mechanism for the stagnation. “When you get a new powerful productivity tool, you can do one of two things. You can either use it to drill deeper, or you could use it to drill more holes. And if you do more projects — just drill more holes — in the end of the day your attention is going to be more thinly spread across multiple projects, and as a result of that you’re actually less likely to make a breakthrough at any given time. We see that in the data. AI seems to have had the same effect. AI means that we can do more things, but people seem to be using it to do more things rather than digging deeper.” The driver is incentives: “in academia, the incentive is publish or perish. And so we shouldn’t be surprised if people use it to produce more output rather than spending the next 10 years maybe producing nothing but coming out with a real breakthrough.”
12. Policy for countries behind the frontier — and why the old playbook is now constrained. Historically the best move for followers was to adopt technology invented elsewhere; the US made that deliberate policy through Marshall aid, sharing technology with allies, “and that contributed I think to a meaningful degree to the postwar miracle in Europe, also in Japan and Korea.” The reason most of the world is not at the computing frontier is not that computer technology is unavailable but that “there are some institutional or perhaps cultural constraints that prevent adoption.” AI may break the pattern because of national-security concerns — “you saw that recently with the Trump administration imposing restrictions on foreign use of Anthropic’s latest model. We can expect to see similar things happening going forward, perhaps at greater scale.” His conclusion: “you cannot really be dependent on the technology leader. You have to try to grow some domestic capacity, and with large language models the easiest way of doing that is through open source or open weights — and that’s how China has closed the gap. It really embraced an open-weight ecosystem, in large part because of export controls on chips, which essentially forced it to go there.” He expects many countries to pivot toward Chinese or European technology “if they feel that America is an unreliable trading partner in technology,” or to attempt a domestic open-weight ecosystem — “although that is going to be a harder approach for most places.”

What was actually ingested

The full auto-generated (ASR) English caption track (662 segments, consistent with duration: 28:55 / length_seconds: 1735). No chapter markers. Bloomberg’s standard open, mid-roll music breaks and production credits are excluded from the substantive summary. Speaker attribution is unambiguous throughout (two speakers, long turns).

Dynamic-capabilities tagging

• contextual/external-triggers — the episode’s operative external trigger for firms and states alike is the politicisation of frontier-model access: “the Trump administration imposing restrictions on foreign use of Anthropic’s latest model. We can expect to see similar things happening going forward, perhaps at greater scale,” which converts model dependence into a strategic exposure and forces the domestic-capacity question. Export controls on chips are named as the trigger that pushed China to an open-weight ecosystem in the first place.
• digital-sensing/digital-scouting — Frey’s prescription for every organisation and country behind the frontier is a scouting prescription: adopt technology invented elsewhere (the Marshall-aid precedent), while recognising that the binding constraints on doing so are institutional and cultural rather than availability-based. His adoption for what distinction is the quality criterion attached to it — scanning that ends in email assistance is not the same capability as scanning that ends in new products and technologies.

Linked entities and concepts

• Bloomberg Podcasts — publisher; entity created in this ingest on the second-source promotion rule (first appearance was Erginbilgiç on Bloomberg Leaders).
• Anthropic, OpenAI — named in the export-restriction and government-stake examples respectively.
• RaboResearch — the same task-versus-aggregate gap, with Frey supplying the verification-tax mechanism the chart does not.
• Collison — declining business dynamism against Stripe’s ~2× business-formation jump; see this source’s relationships: for what the disagreement turns on.
• Hugging Face — open weights as the route to capacity, and China’s open-model position read as a consequence of export controls.
• BBC AI Decoded — the same question four days apart at firm and macro altitude.
• micro-productivity-trap — the verification tax as a macro-level mechanism for why task gains do not aggregate, and adoption for what as the quality-over-rate correction to adoption metrics.
• ai-benchmarks — the AlphaGo reversal as a case where superhuman benchmark performance did not survive a distribution shift.
• open-source-ai — open weights as the sovereignty route for countries behind the frontier, and export controls as the cause of China’s open-weight position.
• enterprise-ai-adoption — the adoption-rate-versus-adoption-quality distinction.
• ai-employment-effects — the business-dynamism and firm-entry side of the AI economy.
• strategic-foresight — the exploration-versus-scaling institutional cycle, and the separate-condenser framing of where the next discontinuity would have to come from.
• dynamic-capabilities — the two cells tagged above.

Dangling (single-source mention, deferred per author-entity promotion): Carl Benedikt Frey, Stephanie Flanders. Frey is the wiki’s first appearance of an economist whose earlier work (Frey & Osborne) sits behind much of the displacement literature the ai-employment-effects page tracks; a second source would trigger promotion.

Source quality note

Auto-generated (ASR) transcript; proper nouns corrected at acquire time (Carl Benedikt Frey, Bessemer Venture Partners, Rheinmetall, AlphaGo, Lee Sedol, Anthropic, open-weight, publish-or-perish, export controls on chips — see the raw file’s notes:). This is a book-promotion interview: Frey is discussing his own recently published argument, and the episode is structured to let him make it rather than to test it. Flanders does put the strongest counter-case (scale-is-all, AI as a different kind of technology) and Frey engages it directly, which is more adversarial than the format usually allows. Empirical claims are asserted from his research without citation in the audio — the patenting-up/breakthrough-down divergence, the declining-entry statistics, the drill-more-holes finding (“we see that in the data”) and the amateur-beats-Go-engine result would each need tracing to the underlying papers before being treated as settled. The AlphaGo reversal in particular is a real and well-documented result (adversarial policies against KataGo) but “human amateurs using standard computers… quite easily” compresses it considerably.