The EU and the not-so-simple macroeconomics of AI - Luis Garicano

Discussion of whether AI will cause explosive GDP growth acceleration. Garicano argues that institutional bottlenecks (FDA approval, organizational adoption, regulatory constraints) will slow AI's impact, contrasting with Silicon Valley's extrapolation from coding productivity gains. He distinguishes between autonomous and non-autonomous AI as the key threshold for discrete productivity jumps.

• •Autonomous AI creates discrete productivity jumps; non-autonomous AI remains bottlenecked by human supervision speed
• •Organizations and institutions accommodate new technologies slowly - regulatory approval, adoption, and learning create friction
• •Customer service chatbots show the transition: initially augmenting juniors, then potentially replacing them entirely
• •Silicon Valley overestimates speed by observing only best-case applications like coding

Analysis of AI's short-term macroeconomic impacts through a two-sector model. If AI makes one sector's output nearly free (e.g., legal/medical services), this creates massive consumer surplus but also requires painful labor/capital reallocation, potential unemployment, and relative price changes rather than traditional inflation.

• •Zero-price sectors create huge consumer surplus but GDP measurement problems and unemployment during transition
• •Human capital depreciation is severe: neurosurgeons becoming plumbers represents complete skill devaluation
• •Price shocks in remaining sectors, not inflation - demand increases while supply slowly reallocates
• •Professional associations will lobby for regulatory barriers to slow AI adoption (bar associations, medical boards)

Reconciling contradictory evidence on AI's labor market effects. Brynjolfsson's field experiments show junior worker productivity gains, but aggregate data reveals hiring freezes for entry-level positions. The resolution: AI complements top workers (superstar effects) while substituting for juniors, with the threshold rising as AI improves.

• •Field experiments (customer service, software developers) show 20%+ productivity gains for juniors initially
• •Aggregate data shows significant drops in 22-25 year old hiring in AI-exposed occupations (Brynjolfsson, Lichtenberg papers)
• •Superstar effects at top: limited substitutability between quality and quantity creates massive leverage for AI-augmented experts
• •Hiring freezes rather than layoffs - firms stop recruiting juniors as AI handles routine tasks (McKinsey example: deep research, PowerPoint)

Original theoretical framework showing how AI disrupts the apprenticeship model. Junior workers traditionally 'paid' for training through menial tasks; AI devalues this currency. Whether training continues depends on the ratio of AI's complementarity with experts versus substitution for juniors. Risk of losing tacit knowledge transfer.

• •Apprentices pay for training through menial work (contract review, spreadsheets, basic research) - AI makes this worthless
• •Critical ratio: expert productivity gain with AI vs. junior worker replacement determines if training survives
• •Tacit knowledge cannot be codified in manuals - requires master-apprentice transfer that may disappear
• •Long-term growth problem: no pipeline of intermediate skilled workers to supervise AI in 10-15 years

Garicano's two-tier educational approach: first-year microeconomics uses traditional blue books to ensure foundational thinking skills, while second-year policy class mandates AI use for complex analysis. The challenge is preventing AI from undermining basic skill acquisition while leveraging it for advanced work.

• •Basic level: traditional exams and problem sets to build mental models - AI is 'the enemy' here
• •Advanced level: mandatory AI use for policy analysis, modeling, data work previously requiring PhD-level skills
• •Students can now tackle complex trade models, political economy analysis that was previously impossible
• •Homework in supervised classroom settings may be necessary to ensure basic skill development without AI shortcuts

Garicano challenges Acemoglu's view that 'we' can direct AI development to remain complementary to labor. Three problems: US-China strategic competition makes slowdown impossible, the 'we' is undefined (whose interests?), and attempting control risks unintended consequences as seen in EU regulation.

• •Game theory problem: US-China competition means unilateral slowdown isn't viable - strategic interaction dominates
• •The 'we' is ambiguous: conflicting interests between countries, firms, workers, professions, AI industry
• •As AI advances, inherent shift from complement to substitute appears unavoidable - not a policy choice
• •EU's regulatory approach (GDPR, AI Act) shows risks of intervention: growth stagnation without technology benefits

Detailed breakdown of EU AI Act's four risk tiers and their impact on startups. High-risk applications (including education and health) require error-free training data, 10-year documentation, conformity assessments, and registration with 55 different EU authorities. Combined with GDPR, this creates insurmountable barriers for small companies.

• •Forbidden uses: emotion detection, government social scoring (blocks potentially valuable applications)
• •High-risk category includes education and health - requires error-free training data (impossible standard)
• •Compliance: 10-year data logs, documentation, conformity assessment, registration with 55 EU authorities needing AI expertise
• •15+ papers unanimously find GDPR reduced VC investment, startups, increased compliance costs - AI Act compounds this

Europe has ~2 foundation models versus 50 in the US, despite having researchers, capital, and ideas. Garicano proposes 'smart second mover' strategy: free-ride on US/China infrastructure investment, focus scarce resources on implementation layer, ensure data sovereignty and interoperability to capture value downstream.

• •Europe has researchers, savings pool, ideas but businesses don't scale - regulatory environment is key barrier
• •Smart second mover: accept infrastructure/LLM layers are lost, focus on implementation where Europe can compete
• •Ensure interoperability and data portability to prevent lock-in (contrast with social media's network effects)
• •Value chain: hardware (US/China win), cloud (switching costs risk), LLMs (competitive if interoperable), implementation (Europe's opportunity)

Europe lacks effective error correction mechanisms. The same Commission, Parliament, and Council that passed the Green Deal and digital legislation (2019-2024) are now tasked with undoing it. Political coalitions (center-right, center, center-left, Greens) invested in original legislation resist reversal despite Trump wake-up calls.

• •Von der Leyen Commission that created legislation now tasked with undoing it - structural conflict of interest
• •Coalition (EPP, Renew, S&D, Greens) that passed laws must now reverse them - politically difficult
• •Multiple 'wake-up calls' (Trump election, Ukraine) haven't triggered action - Europe keeps 'going back to sleep'
• •First deregulation attempt (corporate reporting) failed in Parliament despite being 'guaranteed to pass'

Applying Caballero et al.'s asset demand-supply framework to Europe. AI increases interest rates (R) globally through higher productivity and investment needs, but Europe may not get growth (G) due to regulatory barriers. With 120% debt-to-GDP ratios and 3-4x GDP in pension liabilities, higher R without G creates fiscal crisis.

• •Global AI boom increases asset supply and reduces demand → higher interest rates (lower asset prices)
• •Europe faces higher R without higher G if regulation blocks AI adoption - 'R without G' scenario
• •France: 116-120% explicit debt + 300-400% GDP implicit pension debt requires growth to finance
• •Demographic collapse (fertility rates plummeting globally, not just developed world) makes AI productivity gains essential
• •Europe as net saver benefits from higher R, but distributional issues: only top 3-5% have financial asset exposure

In semi-endogenous growth models, declining populations threaten growth unless AI can substitute in idea production. Evidence from Terence Tao (larger math teams enabled by AI proof-checking), protein folding Nobel Prize, and combinatorics suggests AI is already accelerating research - crucial given demographic constraints.

• •Semi-endogenous growth models: need population growth for sustained growth - demographics make this impossible
• •AI's biggest impact may be in R&D rather than as capital - accelerating idea production (Aghion, Jones, Jones framework)
• •Terence Tao: AI enables larger math collaboration teams through decentralized proof verification
• •Protein folding Nobel Prize, combinatorics results show AI already contributing to scientific discovery
• •Care economy applications: robots may be preferred for personal care tasks over human caregivers