Themes structure the research program. Each has its own landing page, survey status, participation route, and eventual public notes.
Survey Now Open!
Responsible AI is not something you approve at the beginning and file away. In government, the real risks often appear only after a system is live: when staff start using it, when edge cases emerge, when performance shifts, or when public scrutiny arrives. The book treats governance as an ongoing discipline—something built into rollout, monitoring, escalation, and revision—not as a one-time checkpoint.
Most governments do not suffer from a shortage of AI ideas. They suffer from too many ideas arriving at once, with too little capacity to sort them well. One of the central themes of the book is disciplined selectivity: choosing use cases that fit the mission, match institutional readiness, and can survive legal, operational, and political scrutiny. In public-sector AI, restraint is often a sign of seriousness, not hesitation.
AI programs are easy to describe in broad, optimistic terms. They are much harder to justify once questions turn to outcomes, costs, and tradeoffs. This book returns again and again to a simple standard: if leaders cannot define what success looks like, measure it against a baseline, and show why the system is worth keeping, then they are not ready to scale it. Public value has to be demonstrated, not assumed.
Many AI efforts do not fail because the technology is weak. They fail because the institution is unprepared. Data is fragmented, ownership is unclear, teams are too thin, approvals are inconsistent, and no one is sure who is responsible once the pilot ends. A major theme of the book is that readiness—organizational, operational, and data-related—is what determines whether AI becomes useful infrastructure or just another short-lived experiment.
Government cannot adopt AI credibly without being clear about limits. What data can a system touch? What stays off-limits? Who gets access? What does the vendor get to retain, change, or learn from use? The book argues that trust is not built through vague assurances. It is built through boundaries: privacy safeguards, security controls, contractual protections, auditability, and clear decisions about where automation should stop and human judgment must remain.
Deployment is not the finish line. It is the moment the real work begins. Once systems are live, leaders have to manage adoption, monitor drift, control cost, respond to failures, and decide whether the system still deserves to remain in use. This book treats production as the proving ground. What matters is not whether a tool can be launched, but whether it can be run responsibly, improved over time, and retired when it no longer delivers.
In government, a program is not truly successful if it depends on one champion, one budget cycle, or one unusually motivated team. It has to survive turnover, scrutiny, leadership change, legislative pressure, and the slow erosion of institutional memory. The final theme of the book is durability: how to build AI programs that can outlast the people who started them and remain legitimate, useful, and governable over time.