Blog · Review Essay · Last reviewed June 25, 2026

AI Needs You and the Democratic Problem of AI

Verity Harding's AI Needs You argues that artificial intelligence should not be governed as a private destiny written by labs, firms, and security agencies. Its strongest contribution is the insistence that the future of AI is a political question, not a technical weather report.

For this review, democratic AI governance means more than public consultation, ethics principles, or asking users to trust a launch. It means allocating decision rights over purpose, procurement, evidence, limits, monitoring, appeal, and shutdown before an AI system becomes ordinary infrastructure.

The democratic problem is not solved by asking the public to endorse tools after they are built. It is solved, if at all, by giving affected people and public institutions power over goals, procurement, evidence, limits, oversight, recourse, and the decision not to deploy.

The Book

AI Needs You: How We Can Change AI's Future and Save Our Own was published by Princeton University Press on March 12, 2024. Amazon lists the hardcover at 288 pages, with ISBN-10 0691244871 and ISBN-13 978-0691244877. The Bennett School of Public Policy identifies Harding as Director of the AI and Geopolitics Project at the University of Cambridge and notes her earlier role as Google DeepMind's first Global Head of Public Policy.

The book's method is historical analogy. Harding draws on the space race, in vitro fertilization, and the early internet to argue that societies have governed difficult technologies before. That premise is a useful correction to AI fatalism. It says that law, standards, institutions, norms, and public pressure are not decorative afterthoughts. They are part of what a technology becomes.

History Against Destiny

AI Needs You is strongest when it refuses the mood that treats AI as either apocalypse or salvation. Harding's historical cases show a more ordinary pattern: capabilities arrive inside institutions, and institutions can be redesigned. The 1967 Outer Space Treaty made space activity a matter of international law and public interest. The UK's Human Fertilisation and Embryology Act 1990 created a statutory framework for embryo and fertility governance. W3C's history describes web standards as the work of a public-facing, international, multi-stakeholder community. None of these examples produced perfect governance. They show that governance is possible before every problem is solved.

That matters here because machine authority is an institutional habit before it is a model property. AI systems do not become powerful only because model weights improve. They become powerful when schools, hospitals, courts, employers, states, and platforms rearrange practice around their outputs. Harding's contribution is to push the reader away from spectacle and toward the machinery of social permission.

The space case is especially useful because it connects ambition to constraint. The Outer Space Treaty did not stop rivalry or technical acceleration, but it made states responsible for national space activities, including non-governmental entities, and required authorization and continuing supervision. That is the kind of analogy AI governance can actually use: not a promise that cooperation wins, but a reminder that private capability can be routed through public responsibility.

The analogies work best when read as constraints, not comfort. Space law did not abolish geopolitical competition. Fertility governance did not settle every question about reproduction, parenthood, access, or embryo research. Web standards did not prevent platform concentration. Their value is narrower and more useful: each case shows that technical possibility can be routed through law, institutions, standards, public values, and dispute. AI governance needs the same routing before deployment becomes dependency.

The Public as Infrastructure

The title is easy to misread as civic uplift. The better reading is infrastructural: public participation is not a comment box added after deployment. It is a design constraint. If AI changes labor, surveillance, welfare, education, policing, medicine, and knowledge work, then affected people need routes into agenda-setting, procurement, oversight, standards, and appeal. Otherwise "AI governance" becomes a conversation among builders, funders, consultants, and officials who already benefit from adoption.

Recent governance documents make that point concrete. NIST's AI Risk Management Framework treats AI risks as socio-technical and organizes practice around functions such as govern, map, measure, and manage. The OECD AI Principles frame trustworthy AI around human rights and democratic values. The EU AI Act, adopted as Regulation 2024/1689, lays down harmonized rules for AI systems in the European Union. Harding's book is not a manual for any one of these regimes, but it fits their shared implication: AI cannot be governed only at the level of model capability.

Democratic governance therefore means more than consultation. It means decision rights distributed across the lifecycle: the right to know that a system is being considered; the right to challenge the purpose before procurement; the right to inspect evidence before deployment; the right to appeal after use; the right to independent audit; and, in high-risk settings, the power to stop a system when safeguards fail.

That turns "the public" from an audience into part of the control system. Workers know where automation will become discipline. Disabled users know when a service path is inaccessible. Patients know when a model changes care into triage. Students and parents know when a learning system creates surveillance rather than support. Public participation is weak when it is limited to opinion; it becomes governance when it can alter specifications, budgets, deployment conditions, and remedies.

Governance and Safety

As of June 25, 2026, the policy environment has moved closer to Harding's premise. The European Commission says the AI Act entered into force on August 1, 2024, with prohibited practices and AI literacy duties applying from February 2, 2025, general-purpose AI governance applying from August 2, 2025, and most rules, Article 50 transparency duties, sandboxes, and enforcement moving into the August 2, 2026 phase. The Commission's current implementation page also reflects the 2025-2026 Digital Omnibus process, including later dates for some high-risk systems after a political agreement. That timing detail matters: democratic governance lives in the difference between final law, implementing guidance, political agreement, and operational capacity.

The Council of Europe Framework Convention on Artificial Intelligence, opened for signature on September 5, 2024, frames AI lifecycle activities around human rights, democracy, and the rule of law. Its public materials emphasize documentation sufficient for affected people to challenge AI-related decisions, complaint routes, procedural safeguards, risk and impact assessments, and possible bans or moratoria for some uses. OMB Memorandum M-25-21, issued April 3, 2025, directs U.S. federal agencies to accelerate AI use while applying minimum risk management practices for high-impact AI and maintaining inventories, compliance plans, and accountability roles; OMB M-25-22 adds acquisition guidance for federal AI procurement.

The safety implication is that democracy is not a soft value added after engineering. It is a risk-control mechanism. Public participation can surface missing harms, local constraints, inaccessible appeal paths, labor impacts, disability issues, procurement lock-in, and cases where a system's measurable target conflicts with the public purpose it claims to serve. Without those channels, an AI system can pass a benchmark and still fail the institution it enters.

Harding's argument becomes operational when translated into controls: public AI registers, plain-language notices, algorithmic impact assessments, procurement clauses that require audit access, independent testing, worker and community consultation, incident reporting, sunset dates, human review with real authority, and appeal processes outside the vendor workflow. These are not anti-innovation rituals. They are how a polity keeps technical capacity from becoming unaccountable administrative power.

A democratic safety file should be concrete enough to test. It should name the system owner, vendor, model or ruleset version, affected population, lawful purpose, non-AI alternative, evidence reviewed, known limits, data categories, human authority, appeal path, procurement rights, incident triggers, rollback conditions, and review date. If those facts cannot be found, the deployment is not only poorly documented; it is politically evasive.

The hard case is not the AI demo. It is the public service, employer, hospital, bank, school, border agency, platform, or court that quietly makes AI output part of a person's available world. Democratic governance has to follow the system into that setting, where errors become denials, rankings become opportunities, and convenience becomes a reason not to ask permission again.

The Democratic Mandate File

The practical artifact this review takes from Harding is a democratic mandate file: a public or inspectable record that states why an AI system is being considered before procurement, piloting, or deployment. It should name the public purpose, non-AI alternative, affected groups, decision rights, evidence required, procurement constraints, data boundaries, oversight authority, appeal route, incident trigger, shutdown condition, and review date.

The file is different from an ethics checklist because it assigns power. It should say who can change the specification, who can demand independent testing, who can pause deployment, who can force a public explanation, and who can obtain correction after use. If the only parties with practical control are the vendor, buyer, and model operator, the public has been consulted but not governed with.

For public-sector and workplace AI, the mandate file should connect the AI system inventory, procurement file, algorithmic impact assessment, audit plan, notice and appeal, and incident process. For agents, add agent identity, tool permissions, revocation owner, and run logs. Democratic control becomes real when these records identify who can stop the loop, not merely who may comment on it.

The refusal condition should be explicit: do not deploy when the purpose is unclear, the affected population cannot challenge errors, independent evidence is unavailable, the vendor controls the only meaningful logs, the system creates irreversible consequences without human authority, or the non-AI alternative has been removed before the public has a say.

The Agent-Governance Reading

The book becomes sharper when read against AI agents. Agentic systems do not need consciousness, divinity, or general intelligence to matter. They need delegated tasks, tool access, memory, workflow permissions, procurement contracts, dashboards, and managers willing to treat output as action. A customer-service agent can change labor. A coding agent can reshape review practices. A benefits-screening system can turn uncertainty into administrative burden. The democratic question is who gets to contest those arrangements before dependence hardens.

Agent governance should begin with permission, not personality. What tools can the agent call? What records can it write? What systems can it change? What human approval is required before an action becomes real? What logs are preserved? Who is liable when the agent acts outside the public purpose? The democratic question is not whether the agent sounds helpful. It is whether people and institutions retain authority over the loop that turns output into action.

NIST's 2026 AI Agent Standards Initiative adds a standards pressure point to that reading: agent governance has to specify identity, authentication, authorization, interoperability, security evaluation, and traceability before delegated action can be treated as governable. Those controls do not make an agent democratic by themselves, but they make it possible to know which system acted, under which authority, through which tool, and with which record left behind.

For public and workplace agents, the governance unit is the mandate. A mandate should state the task, data boundary, tools, accounts, forbidden actions, approval thresholds, logging requirements, escalation path, and person or office responsible for revocation. Without that mandate, agentic AI turns public administration into hidden delegation: a system acts, a human signs, and the affected person cannot tell where the decision was made.

In that sense, AI Needs You belongs beside AI Snake Oil, Human Compatible, and Tools for Conviviality. Those books ask, in different registers, how technical systems are bounded by evidence, human values, and usable autonomy. Harding adds the political demand: a society that will live with AI has standing to shape AI.

Where the Book Needs Care

The book's weakness is that "you" can sound too frictionless. Public participation is not evenly distributed. Time, expertise, language, disability access, immigration status, union power, technical literacy, money, and fear of retaliation all shape who can speak and who is heard. A democratic AI politics therefore needs more than invitation. It needs funded participation, worker consultation, procurement transparency, independent testing, enforceable rights, incident reporting, and ways to say no.

There is also a risk in historical analogy. Space law, fertility regulation, and web standards illuminate governance, but AI systems are unusually easy to copy, integrate, update, and hide inside ordinary software. Their consequences can be diffuse and hard to attribute. The lesson from history should not be reassurance. It should be discipline: identify the institution, define the boundary, log the decision, test the claim, preserve appeal, and make power nameable.

A final risk is governance theater. Councils, consultations, ethics boards, and public workshops can legitimate a decision already made elsewhere. Harding's democratic instinct is valuable only if it is tied to veto points, budgets, evidence duties, labor rights, regulator capacity, and remedies. Participation that cannot alter deployment is not governance. It is reputation management.

What This Changes

AI Needs You gives the archive a useful civic counterweight. Many reviews here track extraction, cult dynamics, automation bias, surveillance, and platform power. Harding's book says criticism should not end in spectatorship. If AI is becoming part of public memory and institutional action, then governance has to move from expert theater into ordinary channels of accountability.

The practical reading is modest and demanding: treat AI adoption as a public act whenever it reorganizes rights, work, education, care, access, or speech. Ask who was included before the system became normal. Ask what limits were chosen. Ask who can inspect, refuse, or appeal. Ask whether the record distinguishes model output from human judgment, vendor claim from independent evidence, and consultation from consent.

That is the link to the site's recurring concern with mediated reality: AI changes the world less by announcing a doctrine than by entering forms, queues, searches, dashboards, classrooms, benefits portals, hiring screens, and help desks. Democratic governance is the practice of keeping those mediation layers visible, contestable, and reversible where necessary.

A future shaped by AI will not be saved by optimism. It will be made less careless by institutions that let affected people exercise real power over the machines being installed around them.

Source Discipline

This review separates book facts, author biography, historical analogies, and current policy context. Amazon and publisher records support bibliographic details. Cambridge's Bennett School supports Harding's current role and professional background. UN, UK legislation, HFEA, and W3C sources support the three historical governance examples. EU, Council of Europe, OECD, NIST, and OMB sources support current AI governance context.

The source boundary matters. A historical analogy is not proof that AI governance will succeed. A regulation is not proof of compliance. A policy memo is not proof that a deployed system is safe. A public consultation is not proof of consent. Governance-grade claims need the system, version, setting, purpose, affected population, evaluation scope, decision authority, procurement terms, incident history, appeal path, and enforcement mechanism.

Timeline claims require special caution. The EU AI Act's baseline dates, Commission implementation pages, AI Act Service Desk timeline, and Digital Omnibus simplification materials are related but not identical sources. A review should not collapse "entered into force," "applies," "proposed," "politically agreed," "formally adopted," and "enforced in practice" into one status.

This page makes no claim that any AI system is conscious, divine, or AGI. It treats AI systems as sociotechnical arrangements whose power depends on institutional adoption, delegated authority, and public accountability.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books