Acting too late
Technological and institutional lock-in from acting too late
Harmful practices, market structures, or technical dependencies become conditions that are costly to reverse.
AI governance from first principles
Govern while choices are still choices.
The definition · every clause opens the framework
AIGovSpec is a source-linked guide for learners, policy researchers, practitioners and people affected by AI decisions. Its editorial framework begins one level above the AI system, with the incentives, institutions, infrastructures and safeguards that turn capabilities into outcomes, tests each arrangement for power, legitimacy, distribution and adaptability, and asks whether it is still a choice.
How to use AIGovSpec
Choose a question to start with.
New to AI governance?
Start with the accountability lesson. Follow a fictional decision from a broad principle to the people, mechanisms and evidence needed to make it work.
Examining a proposal?
Read a hypothetical worked analysis, then use eleven questions to identify assumptions and unresolved choices in your own proposal.
Tracing a concept?
Explore definitions, relationships and disagreements, then inspect the claims and sources behind them.
This is a deliberately limited prototype, organized around an editorial framework rather than a universal standard. Use it to ask better questions and inspect evidence; it does not determine legal duties or certify compliance. Read the method and limitations.
Where governance acts
Four objects, one system around the AI system.
Analytical lenses, not an exhaustive taxonomy. Together they show where a proposal can act, and why it may fail even when its principles are sound.
- 01IncentivesWhat behavior does the system reward?A proposal tells actors to behave responsibly without changing what the environment makes rational, so the stated principle loses to the incentive every time.Open the object
- 02InstitutionsWho has authority, capacity, and accountability?The institution exists formally but fails operationally: a regulator without staff, a safety team without escalation power, an oversight body captured by the organization it supervises.Open the object
- 03InfrastructuresWhat is AI built on, and who controls it?A mechanism assumes access the actor does not have, or an intervention on infrastructure reduces one risk while consolidating power among those already able to comply.Open the object
- 04SafeguardsWhat actually prevents, detects, or limits failure?A principle is mistaken for a safeguard. 'AI should be safe' is a principle; testing a model against specified dangerous capabilities before deployment is a safeguard. 'AI should be accountable' is a principle; a named actor required to keep records, investigate incidents, and bear defined consequences is a mechanism.Open the object
Four cross-cutting tests
The objects tell us where governance acts. The tests help us examine whether the arrangement is defensible.
Power
Who has the ability to shape the system, set terms, control infrastructure, or define the problem?
Legitimacy
By what authority are decisions made, who is represented, and who can contest them?
Distribution
Who captures the benefits, who bears the costs and risks, and who is excluded?
Adaptability
Can the governance mechanism learn, revise itself, and recover when its assumptions prove wrong?
The temporal principle
Govern while choices are still choices.
Early in a technological transition, change is easier but information is scarce; later, the consequences are clearer but change is harder. This is the Collingridge dilemma, and AI governance faces it in two directions at once.
Acting too early
Governance lock-in from acting too early or too rigidly
Governments can institutionalize the wrong metric, threshold, category, or compliance architecture before the technology is sufficiently understood.
The appropriate response is not simply earlier governance. It is anticipatory and adaptive governance: intervene when consequential choices remain reversible, specify how mechanisms will be reviewed, gather evidence about whether they are working, and preserve the ability to revise them when assumptions prove wrong. Govern while choices are still choices, but build governance that can learn.
Inquiry and practice
Eleven questions for any proposal.
A diagnostic, not a checklist. Apply it to a law, corporate policy, treaty, national strategy, or technical standard, or read the worked analyses first.
Open the diagnosticWorked analysis · hypothetical · corporate levelReleasing a significantly more capable frontier modelA frontier AI developer must decide whether to release a model that is significantly more capable than anything currently deployed.Worked analysis · hypothetical · national levelLicensing an imported AI triage systemA health ministry in a middle-income country must decide whether to license an AI triage system that was trained and evaluated elsewhere.Worked analysis · hypothetical · sectoral levelAdopting an AI shortlisting tool under a risk-based lawAn employer operating in the European Union must decide whether to adopt a vendor's AI system to shortlist job applicants.Worked analysis · hypothetical · community levelA school decides whether to pilot an AI tutoring toolA secondary school's leadership must decide whether to run a one-term pilot of a commercial AI tutoring tool for pupils aged 11 to 16.Worked analysis · hypothetical · community levelA community organisation deploys a multilingual AI assistantA community organisation must decide whether to launch an AI chat assistant that answers residents' questions about public services in several languages.Worked analysis · hypothetical · national levelA child-safety regulator responds to a companion chatbot teenagers useA national child-safety regulator must decide how to respond to a companion chatbot that many teenagers use. It could issue guidance, open an investigation, require design changes within its powers, or recommend new rules.
Evidence, not false certainty
Follow a claim back to its sources.
88
source-linked claims
91
sources in the register
11
contested entities
6
worked analyses