Institutions: Can What We Have Metabolise This?
The Claim
Institutions are how humans coordinate in large groups. They're the rules, procedures, and shared expectations that let companies hire, courts adjudicate, governments legislate, and economies function. They work because they're stable. They're stable because they change slowly.
This is a problem when the underlying conditions change fast.
The AI transition, if it proceeds at current or accelerating pace, may require institutional adaptations that existing institutions are structurally ill-suited to make. Not because the institutions are broken (many aren't), but because their tempo is mismatched with the pace of technological change.
Whether institutions metabolise is one of the three big questions from chapter 01. This chapter goes deeper.
Government
Several challenges for governments specifically:
Speed mismatch
Legislative cycles are measured in years. AI capability cycles are measured in months. By the time a law is drafted, debated, passed, and implemented, the technology it was designed for has been superseded.
The EU AI Act is instructive. Drafted starting 2021, based on the AI landscape of that time. By the time it's in force (2024-2026), generative AI has dramatically changed that landscape. Key provisions arguably address the wrong generation of technology.
This isn't a criticism of specific legislators. It's a structural feature of how laws get made vs how fast AI is moving.
Knowledge asymmetry
Legislators and regulators often know less about AI than the firms they regulate. This is normal for technical domains; it's particularly acute for AI because the technology itself is novel and rapidly changing.
Responses vary. Some jurisdictions hire technical experts into government (NIST in the US, various EU bodies). Some rely on external advisors. Some set up dedicated regulators (UK's AISI, EU's AI Office). All are partial solutions.
Scope confusion
What's the right regulatory frame? Product safety? Consumer protection? National security? Data privacy? Labour? Competition? Intellectual property? All of the above apply to AI; none is adequate alone. The regulatory patchwork produces inconsistent outcomes and is expensive for actors to navigate.
Geographic fragmentation
AI is global; regulation is national. A rule that binds US firms doesn't bind Chinese ones. A permissive jurisdiction attracts firms that want fewer rules. This race-to-the-bottom dynamic is familiar from other industries; AI is specifically bad because the externalities cross borders.
International coordination is weak. The voluntary frameworks (G7, G20, OECD) have no enforcement. More serious efforts (UN AI discussions, proposed treaties) are early and fragile.
Markets and Firms
Markets are supposed to be quick. They can be. But they have their own failure modes.
Short time horizons
Public companies are scored on quarters. AI deployment often takes longer to pay off; its risks often manifest over longer horizons still. Markets may underinvest in safety and long-term thinking.
Externalities
Many of AI's most serious risks are externalities: harm to people not party to the transaction. Markets handle externalities poorly without regulatory intervention. Climate is the obvious parallel.
Winner-take-most dynamics
AI markets show strong network effects, data moats, and infrastructure advantages. This produces concentrated markets quickly. Market self-correction becomes unreliable once incumbents dominate.
Information asymmetry
Customers often can't evaluate AI systems well: they don't know what the system was trained on, how it fails, what its biases are. This undercuts market discipline.
Capital allocation
The firms best at raising capital don't always make the best technical or ethical choices. Market discipline rewards some things; careful safety work isn't obviously among them.
Courts
Legal systems face specific challenges:
Pace
Cases take years to resolve. AI-related legal questions pile up faster than courts can address them. Some important legal questions (training data copyright, liability for AI harm, personality rights) are being litigated now; answers won't fully arrive for years.
Doctrinal fit
Many AI questions don't fit existing legal categories cleanly. Is AI-generated content copyrightable? Is an AI system's creator liable for its outputs? Can an AI be "the employer" of decisions it makes? Courts have to stretch or adapt existing doctrines.
International divergence
Different legal systems are reaching different conclusions. US fair use doctrine is relatively permissive about training; EU data protection is stricter. Uniformity across jurisdictions is unlikely.
Resource asymmetry
Large AI firms have substantial legal resources. Plaintiffs often don't. This affects which cases get litigated and how.
Professions
Professional bodies (medicine, law, accounting, education) are significant governance actors. AI challenges them directly.
Licensing and gatekeeping
Many professions restrict who can practice. AI may perform tasks (medical diagnosis, legal advice, tax preparation) that licensing restricts. How do professions respond? Tightening licensing (limiting AI use) or adapting licensing (accepting AI as an assistant) are both happening, unevenly.
Standards
What does "acceptable AI use" look like in a profession? Professional bodies are writing standards, often in response to member concerns. These standards shape practice more than statute often does.
Training and credentialing
If AI can do significant parts of junior professional work, what happens to how professionals are trained? The junior work was how juniors learned. Without it, the pipeline to senior expertise may break. No profession has fully answered this.
Economic disruption
If AI substantially displaces professional labour, what happens to the professions as institutions? They derive power partly from labour scarcity; abundant AI-assisted labour changes that. Some professions will absorb AI as a tool; others may fragment.
International Bodies
The international governance landscape is thin.
UN
The UN has begun AI discussions. A General Assembly resolution in 2024 encouraged international cooperation. An Advisory Body on AI reports. These are early and non-binding.
A formal UN treaty on AI is, as of writing, not realistic in the short term. Sovereignty concerns, geopolitical competition, and the tempo mismatch all make it hard.
OECD
The OECD has AI principles (2019, updated). Broadly sensible; not binding. Useful as a baseline that many governments adopt into their own frameworks.
G7 and G20
Voluntary coordination. The 2023 Hiroshima Process on Generative AI was an early effort. Produces statements more than actions.
Bilateral and plurilateral
The US-UK AI Safety Institute exchange, the EU-US Trade and Technology Council, various bilateral agreements. Useful incremental coordination; not comprehensive.
Private-sector coordination
The Frontier Model Forum (Anthropic, Google, Microsoft, OpenAI) and similar efforts: voluntary safety commitments, information sharing, joint research. Not a substitute for public governance but meaningful.
Historical Comparisons
How have institutions handled fast technological change before?
Well, sometimes
Electricity grids: adapted over decades with public investment and regulation. Nuclear energy: international safeguards work (imperfectly but meaningfully). Aviation: globally coordinated regulation, mostly successful at safety.
Poorly, sometimes
Social media: institutions are still catching up ~15 years later. Financial derivatives in 2008: institutions failed to catch dangerous innovations until after a crisis. Pharmaceuticals occasionally: regulation has failed catastrophically in specific cases (opioids).
The pattern
Institutions can adapt, but typically:
- With a lag
- Often after a crisis
- Better with specific technical expertise and dedicated bodies
- Better when international coordination exists
- Worse when change is sudden and wide-ranging
AI may not fit the cases where adaptation worked. It may fit the cases where it didn't.
What Adaptation Looks Like
Specific things institutions are doing or could do:
Dedicated AI bodies
The UK AI Safety Institute, the US AI Safety Institute (now within NIST), Japan's AISI, Singapore's AI Verify, similar in several other countries. Dedicated technical bodies that can move faster than full legislation. Partial solution, useful.
Sector-specific regulation
Healthcare, finance, employment, education: each sector regulating AI applications in its domain. Less clean than cross-cutting rules; more aligned with actual practice.
Rapid-response legislation
Some jurisdictions passing more agile AI laws (less comprehensive than the EU AI Act but faster to update). The UK's pro-innovation approach, for instance.
International coordination expansion
The AI Seoul Summit (2024), the AI Action Summit (2025), various bilateral and multilateral efforts. The cadence is insufficient; it's not zero.
Capacity building
Training government staff, supporting academic research, funding civil-society AI work. Multi-year projects that pay off slowly.
Public investment
Publicly funded AI research, public computing resources, public datasets. Pushes against concentration.
What Adaptation Requires
For institutions to actually metabolise AI well, several things need to be true:
- Political will: leaders who treat AI as important and allocate attention accordingly
- Technical capacity: staff who understand the technology
- Legitimate process: rules produced through processes the public accepts as fair
- Coordination: between governments, within governments, with private actors
- Resources: funding for enforcement, research, capacity
- Humility: acknowledging that early answers will be partial and need updating
Most jurisdictions have some of these some of the time. None has all of them reliably.
The Honest Answer
Can institutions metabolise this?
- On some dimensions: yes, given time and will
- On other dimensions: uncertain; depends on pace of AI vs pace of institutional response
- In a worst case: no, particularly if capabilities advance faster than expected
The prudent position is to work as if institutions might not keep up automatically, and help them keep up specifically.
Common Pitfalls
"We just need the right regulation." Regulation is part of the answer. Institutional adaptation is broader: courts, markets, professions, international bodies all matter. A single perfect law wouldn't suffice
"Innovation requires no rules." Specific rules can protect innovation; others suppress it. The absence of rules advantages incumbents with legal resources, not innovators generally
"Rules stifle progress." Some do; some enable. Regulated industries (aviation, pharmaceuticals, finance) can be innovative. The right question is which specific rules, not whether to have rules
"Let experts decide." Experts disagree. Democratic accountability matters for consequential technology. Narrow expert governance has its own failure modes
"Global coordination is impossible." Hard, not impossible. Nuclear non-proliferation, ozone layer protection, various financial coordination show hard things can get done. AI is hard but comparable
Next Steps
Continue to 06-capability-timelines.md for the question of when transformative AI arrives, and why the answer is genuinely uncertain.