Can AI solve Europe’s problems?
Baumol's disease, regulatory resistance, and the O-ring problem
“Economic growth may be constrained not by what we do well but rather by what is essential and yet hard to improve.” (Aghion, Jones and Jones, 2017)
Epoch AI’s models predict annual GDP growth above 20% once AI automates roughly 30% of tasks. They justify this by saying that “significant AI-driven growth accelerations happen more easily than we thought” and that historical brakes on economic growth are “weaker than anticipated.”
Europe desperately needs this growth. High debt levels, unsustainable welfare states, and a working-age population shrinking by 2 million a year from 2040 create enormous fiscal pressures. AI could reverse Europe's productivity malaise and restore its economic dynamism.
I am an AI optimist. Although I believe Epoch AI’s predictions are excessively optimistic, there are good reasons to believe AI does represent a genuine technological revolution that will lift productivity growth substantially. But Europe faces three specific constraints that may prevent it from capturing AI's full benefits: Baumol's cost disease will hit Europe harder than other regions, resistance to change is entrenched throughout European institutions, and O-ring constraints limit system-wide gains. Addressing these barriers will determine whether Europe actually thrives in the AI era.
Baumol
In the 1960s, William Baumol described how productivity gains in one part of the economy inevitably push up costs in another. He famously used the example of a string quartet: it takes the same number of musicians the same amount of time to play a piece today as it did 200 years ago. Their productivity is flat. Meanwhile, other sectors have become so productive that they can offer higher wages.
The sector that becomes more productive drives up wages for everyone. Sectors that are not being automated, such as nursing, therapy, or skilled crafts must match rising wages without the benefit of productivity gains. A hairdresser in 2024 earns far more than a hairdresser in 1950, not because hairdressing productivity increased dramatically, but because other sectors became so productive. As a result, low growth sectors become a larger and larger share of the economy.1
Applying this insight to artificial intelligence, economists Philippe Aghion, Benjamin Jones, and Charles I. Jones (2017) model AI as the latest wave in a 200-year process of automation. Their work suggests a crucial limit to AI-driven growth: an economy may be constrained not by the tasks AI excels at, but by those that are essential and simultaneously hard to improve. As AI makes some tasks hyper-efficient, the remaining human-intensive services become the bottlenecks. Consequently, the ultimate constraint on growth may not be the power of machine intelligence, but our ability to make progress in the most difficult-to-automate areas.
Europe's demographic and institutional structure concentrates economic activity precisely where AI may struggle to deliver gains. In particular, these are tourism, age-related care, and public sector works. Projected age-related public expenditures will rise from 25% today to 29% of GDP by 2070 in Europe. Tourism drives one tenth to one fifth of GDP in Spain, Italy, and Greece. Public administration employs nearly a quarter of the workforce in many European countries.
By 2030, one in four EU citizens will be over 65. An aging society is less able to switch and adopt new technologies. Healthcare, elder care, and social assistance demand human presence. Yet they will need to compete for workers in an economy where AI raises productivity and wages elsewhere. Healthcare costs will inevitably consume larger GDP shares as the rest of the economy becomes more productive.
Tourism shows the same constraint. AI may streamline bookings or provide information, but it does not replace what tourists seek in a host country: a vibrant hospitality industry. A family-run Italian vineyard or Alpine hiking guide will not be able to scale output like a software engineer or even a semiconductor factory.
Lastly, European governments employ roughly 20% of the workforce—teachers, police, civil servants, health workers. These roles have seen minimal productivity growth for decades, yet their wages rise with the general economy. AI might assist with paperwork or analysis, but the core functions—classroom teaching, community policing, public health—remain human-intensive (governments are also generally slower to adopt new technologies). As AI makes other sectors more productive, the relative role of public services will climb, forcing governments to choose between higher taxes, reduced services, or unsustainable deficits. Hence fiscal dynamics may actually worsen, rather than improve.
Resistance to change
In those sectors where AI can make a difference, European employment‑protection rules and other entry‑and‑exit barriers slow both the deployment of new technology and the reallocation of talent toward more productive firms. Some examples:
German codetermination law forces large companies (over 2,000 employees) to run every major capital‑expenditure decision—including automation—through a parity supervisory board that cannot act without worker consent. Germany's updated Works Constitution Act now explicitly mentions AI in three contexts: works councils can demand external AI expertise, employers must give advance notice of AI implementation in work processes, and councils must approve AI-assisted recruitment guidelines. Recent AI cases show projects being frozen for months until agreements are signed.
Italian and Spanish law make head‑count reductions so costly that many automation projects never reach break‑even. In Italy, every dismissal that would qualify the worker for NASpI unemployment benefits triggers the “ticket di licenziamento”—41% of the monthly unemployment benefit cap (1 550 € in 2024) per completed year of service in the past three years—plus doubled rates in collective layoffs. In Spain, any collective dismissal (ERE) obliges the firm to finance an external out‑placement plan of at least six months, in addition to statutory severance (20 days per year of service, up to 12 months).
French labour law goes further: before technology‑driven job changes the employer must negotiate a Plan de sauvegarde de l’emploi whose mandatory “actions de formation… ou de reconversion” cover every affected worker.
Nordic collective‑bargaining practice locks in existing classification and wage grids: the year‑long strike that forced Tesla’s Swedish subsidiary to sign a sectoral agreement illustrates how unions can block AI‑enabled work reorganisation until those grids are honoured.
These rules make labour reallocation between firms unusually expensive.
Asymmetric regulatory risk is equally clear in product markets. Waymo vehicles have already driven more than 20 million fully driverless miles and sell rides in Phoenix, San Francisco and Los Angeles. No EU city has authorised comparable commercial service. The reason is two‑fold: the draft EU AI Act lists autonomous vehicles as “high‑risk”, triggering conformity‑assessment and oversight obligations that still lack implementing standards; and powerful taxi lobbies keep pressure on national regulators—Germany’s courts have already banned core Uber ride‑hailing products for licensing reasons.
Chinese factories can retool in months; US engineers move freely from shrinking to growing firms; but for European firms, the optimal strategy is often to keep doing yesterday’s work with yesterday’s technology.
Europe's regulatory framework empowers sectoral lobbies as effective veto players. Taxi unions locked down Brussels in 2022, stalling ride‑hailing deregulation. Swedish unions kept a year‑long blockade on Tesla until the firm signed a collective agreement, a signal that even global manufacturers must bow to Nordic wage norms. Courts reinforce that power: a 2019 Frankfurt ruling banned Uber’s core service nationwide for licence breaches. When organised groups can snarl city centres or swing coalition votes—and the law requires officials to bargain with them—they can block or delay any AI application that threatens their members’ jobs.
O-rings
In complex production processes, tasks are often complements rather than substitutes. You cannot replace one excellent surgeon with two mediocre ones during a delicate operation. You cannot substitute nine brilliant engineers for one incompetent one. The quality of the worst-performing essential task determines the value of the entire enterprise. Modern service delivery has this character everywhere. A restaurant can have the world's best chef, but if the waiter is surly and incompetent, the dining experience fails.
This insight, which was developed by Michael Kremer (1993) is known as the O-ring theory: the space shuttle Challenger exploded because a single rubber O-ring meant to seal the propellant tanks failed in cold weather; the aircraft worth billions became worthless because of one component failure.2
Europe institutionalizes these “weakest link” dynamics in its regulation and culture through the requirements of “human in the loop”. By deliberately inserting human oversight into critical processes, our systems ensure that automation never runs entirely ahead of human judgment–this also means the overall process can only move as fast as the slowest human involved. Laws and norms demand that humans remain responsible for key decisions across many sectors, effectively creating intentional bottlenecks in the name of safety, accountability, and ethics. Some illustrative examples include:
Aviation: Autopilot technology can technically handle most of a flight, but European aviation regulations still require human pilots on the flight deck at all times. Even as industry explores single-pilot or fully autonomous flights, EU authorities have pushed back. The European Aviation Safety Agency recently halted progress toward single-pilot operations, arguing that today’s most advanced flight control systems are not yet “smart” enough to replace the safety provided by two human pilots.
Finance: In algorithmic trading and financial services, EU rules mandate human oversight of AI-driven systems. For instance, Europe’s Markets in Financial Instruments Directive (MiFID II) imposes strict risk controls and requires that algorithmic trading programs be monitored by people who can intervene or shut them down if something goes awry. This ensures a human is in the loop for high-speed trading decisions. Similarly, EU AI Act regulations classify many financial AI applications as “high-risk,” meaning firms must build in transparency, accountability, and human override mechanisms for those tools. The intent is to prevent unchecked automation in markets–though it also means trading can only be as swift and efficient as human supervision allows.
Food Safety: Europe's food safety framework, as determined by the General Food Law (Regulation (EC) No 178/2002) and the Official Controls Regulation (EU) 2017/625, mandates a “human in the loop” for all food production, even with advanced automation. While food businesses are responsible for implementing safety systems like HACCP (Hazard Analysis and Critical Control Points), official human inspectors are legally empowered to verify compliance at every stage, from farm to fork. This ensures accountability and is reinforced by the precautionary principle, which allows human authorities to halt production under uncertainty.
This human-in-the-loop principle indeed upholds “European values” of oversight and caution, but it guarantees that human capabilities (attention spans, working hours, and cognitive limits) will bottleneck system performance in an era when technology itself could move much faster. The European model thus explicitly chooses a controlled, human-paced system over a potentially faster but unbounded automated system.
The future
Aggregate European growth reflects a weighted average across all sectors. If large portions of the economy—public services, small businesses, traditional crafts, relationship-intensive industries—remain constrained by human bottlenecks, they will drag down overall growth rates despite spectacular AI-driven gains in narrow areas.
It is not likely that AI will automatically solve Europe’s problems while preserving every existing protection and process. Labour market institutions, especially codetermination requirements for automation decisions and collective dismissal procedures that block workforce reallocation, are the biggest obstacle. Additional “human in the loop” mandates increase friction despite often serving no safety purpose. The alternative to fixing this is that other regions capture AI's benefits while European workers perform increasingly expensive human tasks for declining wages.
References
Aghion, Philippe, Benjamin F. Jones, and Charles I. Jones. Artificial intelligence and economic growth. Vol. 23928. Cambridge, MA: National Bureau of Economic Research, 2017.
Kremer, Michael. “The O-ring theory of economic development.” The Quarterly Journal of Economics 108, no. 3 (1993): 551-575.
This mechanism was first analysed for the AI case by Aghion, Jones and Jones (2017).
Incidentally, as Kremer (1993) shows, that production function implies talent segregates into industries: the best investment bankers work with other brilliant bankers and with the best programmers, administrators, support staff. Any weak link will doom the enterprise.
First of: very good analysis.
Let me offer a complementary perspective on Baumol’s disease informed by culture, before adding a few (pessimistic) thoughts on Europe.
I had a related conversation with a colleague last week about Baumol’s disease. Society has not become more productive at raising children, he said. How then can we not expect a substitution away from house-work---raising children---towards the more productive and better remunerated market-work activities? We quickly reached a consensus that looking at childrearing as a productive activity would result in fatalistic conclusions, much like Baumol’s string quartet I dare say: to be replaced with more productive uses of our time. I believe this view is mistaken. Childrearing much like playing in a string quartet is part of leisure. Many string quartets were never meant to be performed professionally, but rather were written to be played by amateurs. And to enhance the enjoyment of the string quartet and childrearing alike, technological progress can indeed come to the rescue. Better books make bedtime reading all the more enjoyable. And better violin education (e.g., Suzuki on a basic, Flesch on an advanced level) makes playing the great string quartets all the more achievable.
There is a deeper insight here: Baumol’s disease is not so much of a disease, but rather a profession of a natural law. When one person cares for another (hairdressing or amateur concerts included), society thrives most when it is done by amateurs, people who love doing this for others. Said less poetically and more in terms of an economic proposition, the professionalization of caring for one another and high economic growth are incompatible. Equivalently, to sustain high economic growth, individuals must enjoy doing as part of their leisure the services whose productivity by their very nature cannot increase.
I should add: the proposition that a society which entrusts activities where there is no productivity growth to amateurs instead of professionals will grow at a faster rate is not a normative judgment. It is an implication of Baumol’s disease: if market hours are spent on tasks for which there is no further improvement in productivity, the productivity of labour declines. If instead leisure is spent on these tasks, technology may well improve the experience of that leisure without raising its output. It means that a society which wishes to grow its economy should thrive to make caring a worthwhile rather than disliked use of individual time for leisure. (This is a really important point: once absorbed, think through the implications for housing, infrastructure, income tax, pension, volunteering in health care, etc..)
I personally think this is all well and just. The girls are happier if they braid each others’ hairs. Couples are happier if they learn to listen to each other and look after each others’ ailments like a sore muscle with tender care. Friends can enjoy teaching and honing their football skills. At parties, amateur musicians can play the music for food and amateur cooks supply the dishes for music. In a way, all of these are ways to creatively express our humanity: each and every age can contribute according to their talents. As a byproduct, self-rated happiness rises. On the flipside, European demographics are an economic problem when old people live on their own in institutionalized care, rather than being productively integrated within the larger family where they would not only require help but also could help with childrearing and housework. Then, suddenly, time use that would statistically fall into leisure becomes part of market work. That’s when economists start talking about Baumol’s disease. I feel that the disease started earlier on, when we replaced leisure that included caring for loved ones (something the state could encourage via the tax system) with bored binge-watching on Netflix. Recent practice of ending the life of the dying prematurely (a euphemism to be clear) reflects the economic incentives at work when ever more care becomes professionalized: palliative care is expensive, death by external force is cheap.
On Europe: European founders had given the European project simple rules and a simple objective. Free exchange of goods and services. No state aid. Constrained fiscal policy and a hard currency. I still believe this can work. Unfortunately, regulation has been a bit excessive as of late (e.g., supply chain reporting). Constrained fiscal policy went out of the window in the early 2000s. And since, the central bank has been doing its best to stretch its mandate to hide unsustainable levels of government debt that in come the next real-world crisis plenty of inflation will solve once more. Ironically, now a growing number of European leaders thinks that Europe is not growing because our state aid and M&A rules are too strict. (They are wrong.)
I emphasise: We are experiencing a crisis caused by bad governance by democratically elected governments, not unelected bureaucrats: Regulation often reflects preferences of individual member states or exists at the national level as the labour laws you cite. Meanwhile, fiscal policy rules were first breached by the heavyweight countries Germany and France. So mostly the nation states, not Brussels, will have to sort this mess. If European citizens would like to see more growth, they can vote for it.
The pessimistic note is that I see little indication that Europe will return to high growth rates anytime soon. Having cheap energy is paramount (which could be solved by a large-scale nuclear revival). Yet we are busy increasing its cost. Demographics are terrible. But no effort is underway to make would-be parents’ life financially any better (housing, tax breaks, infrastructure). Europe has taken in millions of immigrants. Why are no new cities built to house them and too many not integrated into the labour market 10 years after their arrival?
All said, renewal starts with the people, with culture, with how we spend our time, how much we care for one another (in our leisure). But it would be nice if politics got less into the way.
Great work here.
Unless there is a notable cultural/political shift in Europe, I fear that AI will exacerbate, rather than aid, the continent’s drift behind the US and China.
What the regulatory industrial complex often does is hold new technology to a higher safety standard to the status quo. This means new technologies, like autonomous vehicles, won’t be able to improve themselves.
It’s a catch-22 situation that trades reduced short-term risk for increased long-term risk.
In essence, we try building a safe world made of pillows, but end up suffocating beneath them.