Blog · Review Essay · Last reviewed June 25, 2026

AI Ethics and the Machine Moral Infrastructure Problem

Mark Coeckelbergh's AI Ethics is a compact map of the ethical disputes around artificial intelligence. Its deeper value is showing why ethics cannot be a decorative policy layer added after machine systems already shape work, judgment, public memory, and institutional authority.

Machine moral infrastructure, in this review, means the practical machinery that turns computational output into authorized action: categories, data, labels, objectives, interfaces, permissions, contracts, logs, audits, escalation rules, appeal paths, and liability assignments.

The test is whether ethical language changes the machine's authority chain. A principle that cannot alter data collection, model design, procurement, human oversight, incident response, or recourse is not yet governance. It is a wish attached to an already-moving system.

The Book

AI Ethics was published by The MIT Press in 2020 as part of the MIT Press Essential Knowledge series. MIT Press lists Mark Coeckelbergh as the author, the paperback ISBN as 9780262538190, the publication date as April 7, 2020, and the length as 248 pages. Amazon lists the same book with ISBN-10 0262538199 and ISBN-13 978-0262538190. MIT Press identifies Coeckelbergh as Professor of Philosophy of Media and Technology at the University of Vienna.

The book moves through familiar AI narratives, debates over the difference between humans and machines, questions of moral status, machine learning and data science, privacy, responsibility, delegated decision making, transparency, bias, work, and policy. Coeckelbergh treats AI ethics as a field where metaphors, technical design, institutional authority, and political ideals all meet.

Current Context

As of June 25, 2026, the field around AI Ethics has shifted from principle-setting toward operational proof. UNESCO's Recommendation on the Ethics of Artificial Intelligence remains a global reference point for human rights, dignity, fairness, transparency, and human oversight. The OECD AI Principles, first adopted in 2019 and updated in 2024, still frame trustworthy AI around human rights, democratic values, robustness, safety, and accountability. Those sources establish a vocabulary, not a deployment record.

The implementation layer is now more concrete. NIST's AI Risk Management Framework is voluntary but widely used as a lifecycle vocabulary for governing, mapping, measuring, and managing risk; its generative AI profile adds risks such as confabulation, data provenance, privacy, cybersecurity, harmful bias, and misuse. ISO/IEC 42001:2023 treats AI as a management-system problem, while ISO/IEC 42005:2025 gives guidance for AI system impact assessments. These standards make ethics administrable, but they still require local evidence.

The legal context has also changed. The EU AI Act entered into force on August 1, 2024, with phased application through 2025, 2026, and 2027. The Commission's policy page lists broad applicability on August 2, 2026 with exceptions, while the AI Act Service Desk records obligations such as record-keeping for high-risk systems, human oversight, and fundamental-rights impact assessments for certain deployers. In the United States, OMB memoranda M-25-21 and M-25-22 connect federal AI use and acquisition to innovation, governance, public trust, performance, risk management, interoperability, and acquisition evidence. The practical result is that "AI ethics" increasingly means dated records, named owners, and reviewable controls.

Machine Moral Infrastructure

The phrase "machine moral infrastructure" is useful because it shifts attention from abstract machine morality to the parts of a system that make moral consequences durable. An AI system becomes ethically important when its outputs enter a workflow that decides who is trusted, delayed, promoted, denied, watched, routed, ranked, or believed.

That infrastructure has at least four parts. Representation decides what counts as a case, a risk, a need, a violation, or a successful outcome. Authority decides when an output becomes a recommendation, a default, a command, or a record. Recourse decides whether affected people can see, correct, appeal, or refuse the result. Maintenance decides whether errors, drift, incidents, and vendor changes are detected before they become ordinary procedure.

On that reading, ethical AI is not a moral personality test for machines. It is an institutional design problem. The question is whether a system preserves human agency, evidence, and repair at the points where automated classification meets real power.

A useful definition therefore has to name the transfer of authority. A model output becomes morally infrastructural when it changes a queue, file, score, route, price, grade, diagnosis, eligibility decision, moderation action, security alert, or worker instruction. The ethical unit is not only the model. It is the model plus the workflow that makes other people live with its categories.

Ethics Is Not a Layer

The strongest lesson of AI Ethics is that ethics is not something added to a system after the engineering is done. A model already carries choices: what data counts, which labels are available, what error is tolerable, who is measured, who is excluded, which objective is optimized, and what form of explanation a user receives. Those choices may be hidden inside procurement documents, dashboards, product defaults, moderation queues, training pipelines, and organizational incentives, but they are still moral and political choices.

This is why a review of AI ethics belongs beside questions of belief, labor, governance, and memory. A system does not need consciousness to become morally consequential. It only needs to mediate hiring, policing, credit, medical triage, education, insurance, welfare, work allocation, search, speech ranking, or intimate companionship. The ethical question is not whether the machine has a soul. It is how human institutions use machine outputs to authorize action.

The weak version of AI ethics asks whether a product has principles. The stronger version asks where the principles are enforced. Are they in the training data and evaluation set? In the procurement contract? In the user interface? In the manager's incentives? In the appeal process? In the incident log? In the budget for independent audit? If the answer is nowhere operational, the ethic is only a caption.

The same point applies to documentation. A model card, system card, impact assessment, or audit report is ethical only if it can constrain use: delay deployment, narrow scope, trigger retesting, require notice, force a human review path, change procurement terms, or preserve evidence for a person seeking repair. Otherwise documentation becomes a museum label for a machine that continues to operate unchanged.

Moral Panic and Administrative Power

Coeckelbergh is useful because he does not let AI ethics collapse into either nightmare or salvation. The MIT Press description notes that the book discusses narratives from Frankenstein to transhumanism and technological singularity, but the book's center of gravity is more practical: responsibility, privacy, transparency, bias, work, and policy. That is the right emphasis. AI panic can make the machine appear larger than the institutions deploying it. AI boosterism can do the same.

The more durable problem is administrative power. When a score, ranking, chatbot, classifier, recommender, or agent is inserted into an organization, it changes who must explain themselves. The person denied a benefit may be asked to contest an opaque process. The worker may be asked to satisfy a dashboard. The student may be routed by an unseen model. The user may mistake synthetic fluency for institutional care. AI ethics begins where that asymmetry becomes visible.

Moral panic also gives institutions a way to avoid ordinary accountability. If the machine is treated as mysterious, managers can blame "the algorithm." If it is treated as neutral, they can blame "the data." If it is treated as inevitable, they can blame "the future." Coeckelbergh's practical emphasis helps cut through that fog: ask who chose the purpose, who validated the evidence, who benefits from speed, who absorbs the error, and who has authority to stop the system.

Administrative power is why ethics has to include notice and appeal. A person cannot meaningfully contest a machine-mediated decision if the institution cannot name the system, explain the rule or evidence at the right level, preserve logs, identify the accountable office, and correct the downstream record. Without recourse, ethical principles may improve the institution's self-description while leaving the affected person in the same maze.

The Agent Reading

Read in June 2026, the book also helps with AI agents. Agentic systems make an old ethics problem sharper because they connect model outputs to tools, files, workflows, and external actions. The ethical boundary moves from content quality to delegated authority.

That makes Coeckelbergh's attention to responsibility and delegation especially relevant. The responsible party is not the model. It is the institution that gave the system a goal, access, budget, user interface, escalation path, monitoring regime, and authority to act. A serious agent policy therefore needs scope limits, logging, permissions, rollback, human authorization for consequential steps, incident review, and a real appeal path for affected people. Without those controls, "human in the loop" can become a phrase that hides responsibility instead of assigning it.

Current standards work makes this concrete. NIST's 2026 AI Agent Standards Initiative treats agent authentication, identity infrastructure, interoperable protocols, and security evaluations as standards problems. That is exactly the right level of analysis. A tool-using agent is not only a conversational surface. It is an account with permissions, a chain of tool calls, a record of actions, and a possible source of institutional commitments.

For safety, the practical questions are plain. What systems can the agent read or write? What credentials does it use? What actions require confirmation? What actions are forbidden? What logs are preserved? Can the action be reversed? Who reviews anomalies? Who notifies affected people? If those questions are unanswered, the agent is not governed; it is merely supervised after the fact.

The agent version of machine moral infrastructure therefore needs least privilege, scoped identity, action receipts, rollback where possible, and a rule that separates drafting from sending, recommending from deciding, and internal analysis from external effect. Ethics fails when an agent's permission boundary silently becomes the institution's moral boundary.

Governance After Principles

Since AI Ethics appeared, the public governance layer has become more concrete. UNESCO's 2021 Recommendation on the Ethics of Artificial Intelligence frames AI ethics around human rights, dignity, transparency, fairness, human oversight, and policy action areas. The OECD AI Principles, adopted in May 2019 and updated in May 2024, emphasize trustworthy AI, human rights, democratic values, transparency, robustness, security, safety, and accountability. NIST's AI Risk Management Framework 1.0, released on January 26, 2023 and under revision in 2026, gives organizations a voluntary framework for AI risk management across design, development, use, and evaluation.

The legal and management layers have also hardened. The European Commission's AI Act page says the Act entered into force on August 1, 2024, with phased application; the Commission and AI Act Service Desk should be checked for current timing because implementation and omnibus proposals can change details. ISO/IEC 42001:2023 gives organizations an AI management-system standard for establishing, implementing, maintaining, and continually improving governance around AI products and services. ISO/IEC 42005:2025 adds impact-assessment guidance focused on effects on individuals, groups, and society. For U.S. federal agencies, OMB M-25-21 and M-25-22 connect AI use and acquisition to risk management, fit-for-purpose performance, vendor sourcing, data portability, and interoperability.

Those frameworks confirm the book's main point but also raise the bar. Principles are not enough. Values have to become inventories, risk owners, documented purposes, data lineage, testing records, procurement requirements, human oversight roles, incident triggers, public notice, appeal paths, audit access, and exit plans. Otherwise, AI ethics becomes institutional theater: a vocabulary for sounding responsible while the decision system remains unchanged.

Human oversight is the clearest test. The EU AI Act's Article 14 requires high-risk systems to be designed so natural persons can effectively oversee them, understand relevant capacities and limitations, monitor operation, avoid over-reliance, interpret outputs, disregard or reverse outputs, and interrupt operation where appropriate. That is not a ritual approval click. It is a role with time, information, training, authority, and a real ability to stop or repair a decision.

A minimum ethics file should therefore identify the use case, system owner, affected population, data provenance, model or vendor version, risk assessment, chosen controls, human oversight capacity, audit or evaluation evidence, incident trigger, appeal route, retention rule, and exit plan. The file is not bureaucracy for its own sake. It is the memory that lets an institution prove what it authorized and lets a person challenge what happened to them.

Where the Book Needs Care

The book's compactness is both its strength and its weakness. It is an excellent entrance into the field, but a short survey can make conflict appear smoother than it is. "Ethical AI" often sounds like a shared aspiration. In practice, it is a fight over power, money, liability, labor, surveillance, public procurement, military use, platform governance, environmental cost, and the right to refuse automation.

Coeckelbergh gives readers the vocabulary needed to enter that fight, but the next step is harsher and more institutional. Ask who benefits from calling a system ethical. Ask who can inspect it. Ask who can stop it. Ask whether a person harmed by the system has time, money, legal standing, and evidence enough to contest the decision. Ask whether workers affected by automation have bargaining power or only training modules.

The second limit is that ethics can become too model-centered. Many harms come from the surrounding structure: legacy databases, bad categories, forced adoption, vendor lock-in, weak procurement, manager incentives, inaccessible notices, and appeal systems that exhaust the person seeking repair. A fairer model inside an unfair workflow may only make the workflow more durable.

A third limit is that moral status debates can absorb energy that should be spent on present accountability. Questions about future machine personhood are philosophically legitimate, and Coeckelbergh treats them as part of the field. But most current harms do not require a conscious machine. They require an institution willing to let a system classify, route, rank, persuade, or act without enough evidence, oversight, and recourse.

AI Ethics belongs in this archive because it refuses the fantasy that machine intelligence is only a technical achievement. It is also a moral infrastructure project. The book is most valuable when read not as a final code of conduct, but as a warning about where the code of conduct must go next: into the places where machines are attached to authority, labor, memory, and belief.

Source Discipline

This review separates three layers of evidence. Publisher and author pages support bibliographic facts and the book's stated scope. Standards bodies, regulators, and official government documents support current governance context. The interpretive claims about machine moral infrastructure are this review's argument, not claims made by the sources themselves.

The distinction matters. A regulation is not proof that a deployed system is compliant. A standard is not proof that an organization is well governed. A vendor assurance page is not an audit. A "human in the loop" workflow is not meaningful oversight unless the human has usable information, enough time, and authority to intervene.

For current claims, this page prefers primary records: publisher pages for the book; UNESCO and OECD for principles; NIST and ISO for voluntary frameworks and standards; European Commission and AI Act Service Desk pages for EU timing and obligations; and OMB memoranda for U.S. federal AI use and acquisition. Secondary commentary can explain disputes, but it should not carry the legal, standards, or deployment claim alone.

This page makes no claim that any AI system is conscious, divine, or AGI. It treats AI systems as sociotechnical systems whose power comes from delegated authority, institutional uptake, and the records they leave behind.

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