European AI Office

Stated Theory of Change

The EU AI Office's stated theory: binding law changes company behavior in ways voluntary commitments cannot. The AI Act creates enforceable obligations on GPAI model providers, backed by penalties of up to 7% of global annual turnover. The GPAI Code of Practice provides a "presumption of conformity" framework. Companies that want access to the 450-million-person EU market must comply. This creates a "Brussels effect" that raises safety standards globally, because companies find it cheaper to build to the highest regulatory bar rather than maintain separate processes.

Simultaneously, the AI Office claims to foster European AI innovation through compute infrastructure (AI factories, gigafactories), funding programs (GenAI4EU, EUR 700M), and regulatory sandboxes. Sioli: "We look at the innovation policy and the trust policy as two sides of the same coin."

Revealed Theory of Change

The AI Office's actions suggest a different set of priorities than its stated mission:

What the AI Office actually optimizes for:

1. Institutional survival within the European Commission's political dynamics (securing budget, mandate, and political backing from Virkkunen/von der Leyen)
2. Getting major AI companies to sign the Code of Practice (treated as the primary success metric, with Big Tech signatures described as "the currency of the Code's success")
3. Demonstrating innovation-friendliness to counter the "overregulation" narrative from the US and industry
4. Maintaining pace with the AI Act's phased implementation timeline (despite delays)
5. Building international legitimacy through AISI network participation

Where actions diverge from stated theory:

• The binding-law advantage is theoretical if no enforcement actions are taken. 13+ months of active obligations and zero fines, zero compliance orders, zero public sanctions.
• The Code of Practice was weakened by industry lobbying in a 4th drafting round that excluded civil society. Emergency preparedness, pre-deployment reporting, and whistleblower protections were removed -- precisely the mechanisms that would give the Code real teeth.
• The Digital Omnibus delays high-risk obligations by 16 months. The AI liability directive was cancelled. These retreats erode the binding-law credibility that is the AI Office's core differentiator.
• 4 of 6 units focus on innovation, not regulation or safety. The budget, staffing, and institutional attention are heavily skewed toward innovation promotion.
• The Brussels Effect is weakening. The US under Trump explicitly rejects convergence with EU AI regulation. Few countries are copying the AI Act framework. Brookings: "The Brussels Effect may produce not convergence but divergence."

The revealed theory of change is closest to: The AI Office is building regulatory infrastructure that can be activated when political will exists, while in the meantime prioritizing industry engagement and innovation promotion to maintain political support. It is betting that the legal framework will matter more in the future (when fines begin in Aug 2026 and enforcement capacity grows) than it does now. This is a bet on regulatory maturation, not current impact.

Key Assumptions

1. Binding law without enforcement is still more effective than voluntary commitments.

• Evidence for: Companies are actively engaging with the Code of Practice. 26 signed. Legal teams at AI companies are building compliance processes. Non-signatories know they face "enhanced scrutiny." The mere existence of potential penalties shapes behavior.
• Evidence against: Zero enforcement actions in 13+ months. If companies learn that non-compliance has no consequences, the deterrent effect erodes. Meta refused to sign and faces no visible consequences. The GDPR analogy cuts both ways: GDPR enforcement was slow but eventually produced large fines (EUR 1.2B against Meta in 2023), suggesting patience may be warranted.
• Testable: Watch for the first GPAI enforcement action after Aug 2026. If the AI Office demonstrates willingness to fine a major company, the theory is validated.
• If wrong: The EU has created the world's most comprehensive AI law that functions as a paper tiger.

2. The AI Office can build sufficient technical capacity to evaluate frontier AI models.

• Evidence for: The AI Safety unit has genuinely strong hires (Brauner from Oxford/RAND, Moller from Google, Grosse-Holz from UK AISI). The Scientific Panel will provide independent expert support.
• Evidence against: ~125 staff (mostly policy/legal) vs UK AISI's ~250 (90 technical). EUR 46.5M vs GBP 100M. Lead Scientific Advisor position vacant for over a year. No open-source tools, no published evaluations, no visible technical output. EU nationality requirements and Brussels location permanently constrain the talent pool.
• Testable: Will the AI Office publish model evaluation results, release evaluation tools, or produce a trends report comparable to UK AISI's?
• If wrong: The AI Office can levy fines but cannot assess whether companies are actually compliant. It becomes a regulatory body without regulatory intelligence.

3. The dual mandate (regulate + promote innovation) does not compromise regulatory independence.

• Evidence for: Other agencies manage dual mandates. The EU has experience with regulatory independence within the Commission structure (e.g., DG Competition). Sioli genuinely believes trust is essential for innovation.
• Evidence against: The cancellation of the AI liability directive, the Digital Omnibus delays, the Big Tech capture of the Code drafting process, and the innovation-heavy unit allocation all suggest the innovation mandate is winning. MEP Geese: "Centralizing AI supervision in the Commission's AI Office could be an advantage for enforcement... but risks concentrating enforcement in a body that is neither politically neutral nor sufficiently independent."
• Testable: Track enforcement decisions. Does the AI Office enforce against companies that are also partners in its innovation programs?
• If wrong: The AI Office becomes an innovation promotion body that also has nominal regulatory powers it rarely uses.

4. The Brussels Effect will extend EU AI standards globally.

• Evidence for: GDPR precedent. The Code of Practice's model-level requirements mean companies cannot easily develop different models for different jurisdictions. Many companies will standardize to the EU bar. Countries in the EU's sphere of influence may adopt similar frameworks.
• Evidence against: Brookings analysis finds the Brussels Effect for AI "more limited than presented by EU policymakers." The US under Trump explicitly rejects convergence. China has its own AI governance regime. Few countries have copied the AI Act. CEPA: AI Act "inspires few copycats."
• Testable: Track how many countries adopt AI Act-aligned frameworks by 2028.
• If wrong: The EU regulates in isolation, imposing costs on companies operating in Europe without raising global standards. This would validate the competitiveness critique.

Strengths

Binding law is a real structural advantage. The EU is the only major jurisdiction where AI regulation has the force of law with financial penalties. This means the AI Office has tools the UK and US lack entirely: the ability to fine, issue compliance orders, and legally compel cooperation. Whether these tools are used is a separate question from whether they exist.

The Code of Practice's Safety chapter is genuinely rigorous. Independent analysts consistently assess it as more demanding than any company's voluntary framework. Pre-defined risk thresholds, mandatory assessment of loss-of-control risk, external evaluations, and incident reporting are real advances. CSET: "significantly more rigorous and comprehensive than current best practices."

Regulatory adaptability on display. The AI Office's response to DeepSeek -- considering threshold adjustments based on new evidence that frontier capabilities can be reached with less compute -- demonstrates the AI Act's built-in mechanisms for keeping pace with technology. This is better than static regulation.

EA/x-risk awareness inside a regulatory body. The MLex report on effective altruism links in the AI Safety unit means at least some people inside the regulatory body take catastrophic risk seriously. Jan Brauner's background (Oxford AI safety, RAND, published in Science) suggests genuinely safety-oriented technical staff. This is rare in any government body.

Massive market leverage. The EU's 450-million-person consumer market gives it leverage that smaller countries lack. Companies that want to sell AI products in Europe must comply with the AI Act, period. This is not a voluntary ask.

Weaknesses and Risks

Zero enforcement credibility. The most critical weakness. 13+ months of active obligations, zero enforcement actions. The binding-law advantage is meaningless without demonstrated willingness to use it. Every month without enforcement erodes the deterrent effect.

Catastrophically under-resourced relative to mandate. The AI Office has fewer staff and less budget than UK AISI, but a vastly larger mandate (implement comprehensive legislation, manage innovation programs, participate internationally, enforce against all GPAI providers in Europe). This is not a capacity gap -- it is a capacity chasm.

Industry capture of the regulatory process. Corporate Europe Observatory documented Big Tech receiving dedicated workshops while civil society got emoji reactions. The Code was weakened in a 4th round accessible only to US tech companies. Emergency preparedness and pre-deployment reporting were removed. This process damage is already done.

Political retreat under pressure. AI liability directive cancelled. Digital Omnibus delays high-risk obligations by 16 months. Virkkunen "eyes amendments." Each retreat signals to companies that political pressure works, encouraging more lobbying for further weakening.

No pre-market approval power. Despite having binding law, the AI Office cannot stop a dangerous model before it reaches consumers. The Code only requires reporting after deployment. This is a fundamental structural gap that undermines the safety function.

Dual mandate conflict. The same body that regulates AI manages billions in AI investment promotion. Four of six units focus on innovation. This creates structural pressure to prioritize industry engagement over enforcement. The FAA analogy (promoting aviation + regulating safety) was historically considered a failure that was eventually separated.

Brussels Effect may be failing. The US under Trump explicitly rejects EU-style regulation. Few countries are copying the framework. If the EU regulates alone, it bears compliance costs without raising global standards.

Cross-References

UK AI Security Institute: Has 2.5x the budget, 2x the staff, open-source tools (Inspect), GBP 40M in alignment research grants, and a chief scientist (Irving) who speaks publicly about loss-of-control risk. But it has no binding law, no enforcement powers, and no ability to fine non-compliant companies. The EU AI Office has the legal tools the UK lacks; the UK has the technical capacity the EU lacks. Together they cover the full range; separately each has a critical gap.

US CAISI: Has been defunded and politically redirected under Trump. $10M/year budget (less than the EU AI Office). No binding law, no enforcement, no alignment research. The comparison is favorable to the EU on every dimension except talent pool (US has the frontier labs and researchers). Both the US and EU demonstrate that government AI safety bodies are politically fragile.

Ada Lovelace Institute: Argues "a testing regime is only meaningful with pre-market approval powers underpinned by statute." The EU has the statute but not the pre-market approval powers. Their framework for what effective AI regulation requires (mandatory pre-deployment testing, FOIA transparency, enforcement powers, statutory authority) provides a roadmap for what the AI Office would need to become effective.

Centre for Future Generations: Their AISI network analysis concluded the EU AI Office "is not an AI Safety Institute" and proposed a separate "CERN for AI" entity focused on science rather than regulation. This suggests the EU may need two bodies: the AI Office for regulation and a separate institute for safety research.

ControlAI / Alfour critique (from UK AISI analysis): The argument that "evals move the burden of proof away from AI corporations" applies here too. The Code of Practice's voluntary evaluation framework may provide false assurance while delaying the kind of mandatory pre-market approval that would actually constrain behavior.

What Would Change This Assessment

Upward:

• The AI Office levies a significant fine against a major GPAI provider. This would demonstrate that binding law translates to binding enforcement.
• The AI Safety unit publishes model evaluation results or releases open-source evaluation tools. This would demonstrate technical capability.
• The AI Office blocks or conditions the deployment of a GPAI model in Europe. This would demonstrate pre-market authority.
• Budget and staffing increase substantially (to UK AISI levels or beyond). This would address the resource gap.
• Meta faces consequences for refusing to sign the Code. This would demonstrate that non-compliance has costs.

Downward:

• Further Digital Omnibus-style delays or weakening of obligations. More retreat = less credibility.
• Evidence that enforcement actions are avoided for political reasons (pressure from Virkkunen, member states, or industry).
• Major AI safety incident in Europe while the AI Office has published no evaluations and taken no enforcement actions.
• Additional talent drain from EU AI Office to UK AISI or private sector.
• The Brussels Effect demonstrably fails (AI Act framework not adopted by any other major jurisdiction by 2028).

Self-Critique

Weakest claim: That zero enforcement in 13 months constitutes a failure. GDPR enforcement was also slow initially, with years passing before the first major fines. The AI Office may be deliberately building cases, waiting for the Aug 2026 enforcement date for GPAI fines, and conducting preparatory investigations that are not publicly visible. My assessment may be too impatient with a regulatory body that is less than two years old.

Potential bias: I may be unfairly comparing the AI Office to UK AISI. The two bodies have fundamentally different mandates (regulator vs evaluator), different legal frameworks, and different institutional contexts. Comparing staff counts and tool releases may be misleading because the AI Office's value lies in its legal authority, not its evaluation capability. A small regulator with binding law may be more effective than a large evaluator with none.

What I should have checked but could not: Internal AI Office documents. The MLex article on EA influence (behind paywall -- only 149 words visible). EA Forum posts about working for the EU (blocked domain). Whether the AI Office has conducted model evaluations privately. Budget breakdown by unit. Which companies are under active investigation.

What a thoughtful disagreer would say: "You are judging a two-year-old regulatory body against the standard of a mature enforcement agency. The GDPR took 4+ years before the first billion-euro fine. The AI Office needs time to build cases, develop standards, hire staff, and establish precedent. The legal infrastructure is in place; enforcement will follow. The Code of Practice, even weakened, establishes more rigorous safety requirements than anything any company does voluntarily. And the binding-law advantage becomes decisive when the first fine is levied -- at that point, the entire regulatory landscape shifts. Every company will know the EU is serious. Patience is warranted."

This is a reasonable argument. But it requires accepting that during the multi-year maturation period, frontier AI models are being deployed across Europe without meaningful government evaluation or constraint. The pace of AI development is fast enough that the gap between "law exists" and "law is enforced" could be consequential if capabilities advance significantly during that window.

Single weakest claim: That the dual mandate (regulate + promote) necessarily compromises regulatory independence. Some regulators manage dual mandates effectively. The Commission has institutional experience with independent enforcement (DG Competition). It is possible that the innovation mandate is complementary rather than conflicting. However, the pattern of evidence (cancelled liability directive, Digital Omnibus delays, industry capture of Code drafting) is consistent with the innovation mandate taking priority.

Information that would most change my view: Evidence that the AI Office has conducted private model evaluations, is building enforcement cases, or has privately influenced a deployment decision. If significant behind-the-scenes work is happening, the public record dramatically understates the AI Office's impact.