Blog · Review Essay · Last reviewed June 25, 2026

The Unaccountability Machine and Accountability Sinks

Dan Davies's The Unaccountability Machine is a systems book for the age of automated administration. Its key idea is not that machines make decisions alone, but that organizations build structures where authority, evidence, remedy, and operational control are separated until responsibility disappears into procedure, software, committees, markets, and metrics.

The Book

The Unaccountability Machine: Why Big Systems Make Terrible Decisions and How the World Lost Its Mind was published in the United Kingdom by Profile Books in 2024 and in the United States by the University of Chicago Press in 2025. The University of Chicago Press lists Dan Davies as author, gives the hardcover ISBN as 9780226843087, lists additional ebook and audio ISBNs, and describes the book as 304 pages. Profile Books lists the hardback ISBN 9781788169547 and the paperback ISBN 9781788169554.

The book is built around a useful diagnosis: complex systems often create outcomes that no one claims to have chosen. Davies calls the mechanism an accountability sink, a structure where decisions are delegated to procedures, rule books, markets, standards, software, or committees in ways that make the responsible actor hard to find when damage appears.

Current Context

As of June 25, 2026, Davies's argument lands in a policy environment that is trying to turn accountability into records, duties, and remedies rather than slogans. The EU AI Act applies in phases, with its general application date set for August 2, 2026 and additional timing rules for some high-risk obligations. Its high-risk system regime includes requirements for risk management, data governance, technical documentation, record keeping, transparency, human oversight, accuracy, robustness, cybersecurity, and deployer duties. That is governance infrastructure, not a guarantee that every harmed person will receive compensation.

The liability side remains uneven. The EU's revised Product Liability Directive treats software, including AI systems, as a product for no-fault product liability purposes and applies to products placed on the market or put into service after December 9, 2026, while the separate proposed AI Liability Directive was withdrawn by the European Commission in October 2025. In the United States, there is still no single federal AI liability statute. NTIA's 2024 AI Accountability Policy Report instead frames accountability as a chain of information flow, independent evaluation, and consequences, while the FTC has continued to use existing consumer protection authority against deceptive or unfair AI-related claims.

That current context supports the book's central warning. A legal duty, audit, model card, or risk framework can create evidence, but it does not by itself create answerability. The hard governance question is whether evidence can force a decision owner to repair the harm, alter the workflow, compensate the injured person, pause deployment, or change the procurement terms that created the failure.

The Accountability Sink

For this review, an accountability sink is a decision arrangement in which authority, evidence, remedy, and operational control are split across so many actors and artifacts that harm can be explained but not corrected by anyone the affected person can reach. A loan denial, benefit suspension, hiring screen, insurance triage, content moderation action, or medical scheduling decision can all have explanations in this weak sense. The stronger test is whether the explanation points to a person or office with power to change the result.

The idea clarifies a recurring problem in AI liability and accountability. A bad automated decision is rarely just the output of one model. It is usually the product of a procurement contract, a dataset, a policy goal, a workflow, a risk appetite, a user interface, a dashboard, a manager, and an institution that benefits from treating the whole arrangement as neutral process. The system does not need consciousness to become difficult to challenge. It only needs enough layers that everyone can point elsewhere.

That places Davies beside The Black Box Society, Automating Inequality, and The Glass Cage. Those books show opaque scoring, welfare automation, and automation bias from different angles. Davies adds a management theory of disappearance: the harm persists because feedback has nowhere authoritative to land. The practical counterparts are audit trails, notice and appeal, and algorithmic recourse that bind a complaint to an institutional obligation.

Cybernetics Without Mysticism

The book's recovery of Stafford Beer is especially relevant to this site. Profile Books says Davies casts new light on Beer's writing and Beer's claim that organizations should be treated as artificial intelligences capable of decisions distinct from the intentions of their members. That is not a claim that organizations are conscious, divine, or persons. It is a sober systems claim: organizations process information, dampen feedback, allocate attention, and act.

Cybernetics helps because it asks whether feedback reaches the part of the system that can change behavior. A complaint form that never alters policy is not feedback. A dashboard that displays failure while rewarding throughput is not control. A model audit that cannot stop deployment is theater. For AI systems, the cybernetic question is practical: where does evidence of harm go, who must respond, and what authority can alter the machine or the workflow?

A useful feedback loop has at least three properties. It reaches a decision owner, it can change constraints, and it leaves a record that supports review or remedy. Without those properties, the loop is decorative. This is why algorithmic impact assessments should not be treated as paperwork at launch only. They matter when they connect expected risks, actual incidents, post-market monitoring, procurement duties, appeal outcomes, and management authority.

The Agent Reading

Read in 2026, The Unaccountability Machine is a useful warning about AI agents. Agentic systems are attractive because they promise to turn intention into action: summarize this file, update that record, route this case, contact that customer, purchase this service, close that ticket. The danger is not that the agent becomes a mind. The danger is that delegation becomes a new excuse for missing responsibility.

NIST's AI Risk Management Framework treats trustworthy AI as something built into design, development, use, and evaluation, and its Generative AI Profile is explicit that the framework is voluntary. NIST's 2026 AI Agent Standards Initiative also points toward identity, authorization, interoperability, and security as agent infrastructure. Davies supplies the missing organizational caution: a standard or control can still become an accountability sink if no one has the power, time, or incentive to act on what the control reveals.

For AI agents, accountability has to attach to the delegated action, not just to the model call. The system needs a stable agent identity, scoped tool permissions, approval rules for risky actions, logs that preserve prompts, retrieval, tool calls, approvals, and outputs, and an incident owner with rollback authority. Otherwise the institution has built a faster route from instruction to consequence without building a usable route from consequence back to responsibility. The local design problem is covered by the agent tool permission protocol and agent audit and incident review pages.

Governance and Safety

The governance implication is concrete: do not deploy a decision system unless accountability has an address. A minimum architecture includes a named business owner, a system inventory entry, procurement evidence, data and model provenance, version records, tested failure modes, human oversight rules, audit logs, appeal paths, incident reporting, rollback authority, and vendor obligations that survive a public complaint or litigation hold.

Safety also requires restraint. An accountability system should not become indiscriminate surveillance of workers or users. Logs should be purpose-limited, access-controlled, integrity-protected, and retained only for defined governance, safety, legal, and appeal needs. The point is to preserve enough evidence to reconstruct a harmful action and remedy it, not to hoard every private prompt forever.

The same discipline applies to vendors and platforms. A procurement team that accepts a black-box promise, a vague indemnity clause, or a dashboard screenshot has not solved accountability. It has purchased a future dispute about who knew what, who controlled what, and who had power to intervene. The safer pattern is to specify audit access, incident cooperation, change notification, performance boundaries, data rights, subcontractor disclosure, and termination rights before the system is embedded in a public-facing workflow.

Where the Book Needs Care

The book can sometimes make cybernetics sound like the road not taken that might have spared institutions from decades of dysfunction. That is compelling, but the practical task is harder. Public agencies, firms, platforms, and hospitals do not lack only a better theory of feedback. They also face politics, budgets, union power, vendor lock-in, liability strategy, debt, regulatory capture, and career incentives. A feedback loop can be beautifully drawn and still be ignored by the people who profit from not listening.

It also needs a sharper labor reading when applied to AI. Accountability sinks are often built on unequal work. Call-center agents, caseworkers, content moderators, data labelers, and compliance staff absorb the anger generated by systems they did not design. If AI agents automate more of the visible interaction while leaving workers to manage exceptions, the sink deepens rather than disappears. Real accountability asks whether front-line workers can halt the process, correct the record, escalate a systemic failure, and refuse unsafe automation without being punished for slowing throughput.

There is a civil-liberties limit too. More documentation is not automatically better governance. Overbroad logging can expose sensitive user data, intensify workplace monitoring, or create records that are useful to institutional self-defense but useless to affected people. Davies's concept is strongest when it is paired with recourse: records, explanations, and audits must be designed around correction, not merely institutional memory.

What This Changes

The Unaccountability Machine gives the archive a name for a familiar institutional pattern. People experience a decision as machine-made, but the machine is only one component in a larger evasion structure. The user is denied. The worker is measured. The manager cites policy. The vendor cites configuration. The regulator asks for documentation. The harm circulates until it exhausts the person least able to force a response.

The practical test is blunt. When an AI system or automated workflow is proposed, ask where accountability will land when the system fails. Ask who can see the logs, who can change the rule, who can compensate the harmed person, who can stop deployment, and who is named when the dashboard says everything went according to process. A machine that cannot be governed is not intelligent infrastructure. It is an institution hiding from its own decisions.

Source Discipline

This review uses publisher pages for book metadata and framing, primary legal texts for enforceable duties, and official regulator or standards-body pages for policy context. It treats standards as evidence of governance practice, not proof of compliance. It treats enforcement announcements as evidence of regulator posture, not proof that every similar claim is unlawful. It treats retail listings as retail metadata, not authority on the book's argument.

The same discipline should govern any accountability claim about an AI system. Prefer statutes, regulations, contracts, procurement records, incident reports, model or system cards, audit scopes, versioned logs, appeal outcomes, and regulator orders over vendor summaries or promotional claims. A source trail is part of the feedback loop: if a factual claim cannot be traced to a responsible record, it may already be inside the sink.

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