The AI Governance Manifesto

AI should expand people's ability to understand, choose, and act—not diminish it.

We must improve our capacity to govern as deliberately as we improve the technologies being governed. These priorities apply to governance of AI, AI for governance, and governance of AI-enabled governance. The chapters develop the argument collected here.

Both sides of each comparison have value. These are priorities for judgment, not absolute rankings for every circumstance. They do not authorize violations of rights or legal obligations. A worthwhile purpose must be pursued through legitimate means.

We value collective capability over procedural compliance.

Governance should equip people and institutions to understand consequential situations, coordinate legitimate responses, and correct failures. Consistent procedures help make duties clear, preserve memory, and protect people against arbitrary treatment. Their purpose should guide their evaluation: completing a required process does not establish that it achieved its intended protection. Where an obligation needs improvement, change it through legitimate channels; a preferred outcome does not excuse disregarding it.

We value learning through action over advance planning.

Authorized experience can reveal what preparation alone cannot establish. Planning anticipates consequences, coordinates effort, and can prevent irreversible harm; it should help identify responsible opportunities to learn. Action may mean investigation, simulation, a bounded test, or an enforced restriction. Some proposals need extensive preparation, and some should be refused. The prospect of learning does not justify exposing people to unacceptable risks or treating them as involuntary experimental material.

We value useful evidence over comprehensive reporting.

Evidence should help establish what matters, expose uncertainty, and inform consequential decisions. Broad reporting can reveal omissions, preserve records, and prevent selective disclosure. Its usefulness depends on the reliability and intelligibility of its grounds, including unwelcome findings. Required disclosures and independent inquiry remain essential where justified. AI may help produce or examine information, but a fluent report cannot supply missing evidence, and cheaper preparation does not remove the work of verification.

We value public accountability over administrative efficiency.

Consequential power should be answerable to the people it affects through reasons, legitimate scrutiny, meaningful challenge, and effective remedy. Efficient administration saves resources and can make help available sooner. It should support these obligations, with attention to delays that make a remedy unusable. Accountability requires appropriate access and practical means to investigate; it does not require public disclosure of private records or sensitive security information. The audience and form of scrutiny must fit the responsibility.

We value meaningful choice over operational convenience.

People and societies should retain practical alternatives and the ability to question, change, replace, or decline consequential AI arrangements where appropriate. Convenient, integrated services can make useful capabilities accessible. Their benefits should be assessed alongside dependencies that could make replacement or recovery impractical. Choice respects others' rights and proportionate common rules; it does not promise costless customization or complete self-sufficiency. Cooperation can strengthen the freedom that a nominal alternative leaves unusable.

We value lasting benefit over immediate gains.

Today's decisions should preserve durable welfare, resilience, fair opportunities, and future people's ability to choose. Immediate gains also matter: they can relieve hardship, improve services, and create resources for longer benefit. Judge them alongside the dependencies and burdens they create, including who bears those burdens. Stewardship includes present people. Urgent, observable needs should not be sacrificed to speculative future benefits without compelling reasons, and short-term success should not silently close others' future choices.

Commitments that apply throughout

Those who govern must also be accountable. Using AI must not erase responsibility. A claim of improvement must be open to challenge.

Meeting today's standard does not remove the duty to reconsider its adequacy and improve it where justified. Learning must include what governance asks for, not only what it decides.

AI governance must maintain awareness as capability grows. That awareness must be supported by sufficiently current evidence, acknowledge knowledge limits, and remain open to independent inquiry and correction. No AI system should operate at a scope, scale, or level of autonomy beyond the responsible institutions' demonstrated capacity to oversee it and constrain its consequential effects. A material loss of visibility on which authorization depended warrants reassessment. Understanding must connect to legitimate authority and timely, effective response.

Note on the comparative form

The comparative form draws on the Manifesto for Agile Software Development (2001), whose value statements recognize value on both sides. These six priorities are this book's proposal; the influence does not establish the adequacy of software-development practices for public governance. Source: agilemanifesto.org.