Blog · Review Essay · Last reviewed June 23, 2026

Cloud Ethics and the Attribution Machine

Louise Amoore's Cloud Ethics: Algorithms and the Attributes of Ourselves and Others is a theory of machine-learning responsibility written before generative AI became ordinary office equipment. That timing makes it sharper, not weaker. The book studies the older ethical problem beneath today's chatbots, classifiers, copilots, agents, and risk systems: algorithmic decisions are made from partial attributes, uncertain relations, and institutional arrangements that no source-code inspection can fully settle.

The working definition: an attribution machine is a sociotechnical arrangement that turns traces into qualities assigned to people, cases, texts, images, or events, then lets an institution act as if the partial assignment were enough.

The ethical unit is not the model alone or the human reviewer alone. It is the attribution chain: source data, labels, embeddings, thresholds, interface cues, policy rules, human authority, institutional consequence, and the path for correction when the partial account is wrong.

The Book

Cloud Ethics was published by Duke University Press in May 2020. The publisher lists the book at 232 pages, with 27 illustrations, paper ISBN 978-1-4780-0831-6, hardcover ISBN 978-1-4780-0778-4, eISBN 978-1-4780-0927-6, and DOI 10.1215/9781478009276. Duke describes the book as an account of how machine-learning algorithms transform ethics and politics by operating through data attributes, incomplete self-accounts, human-machine relations, partiality, and opacity.

The table of contents makes the structure explicit. Amoore moves from condensed data and correlative reason to neural networks and recognition, then to authorship, aberrant acts, ground truth, partial accounts, and strategies for cloud ethics. JSTOR's book record surfaces the same architecture: the problem is not merely that algorithms hide decisions, but that they produce decisions through relations among traces, thresholds, training practices, institutional uses, and people who are acted upon through attributes.

Amoore is a professor of political geography at Durham University whose work focuses on geopolitics, security, biometrics, data, and algorithmic technologies. The European Research Council describes her as principal investigator of the Advanced Grant project Algorithmic Societies: Ethical Life in the Machine Learning Age. That background matters because this is not a general tech-policy book. It is a political theory of the scenes where algorithmic outputs become security, medicine, border control, policing, credit, risk, recognition, and administrative judgment.

The book belongs beside The Black Box Society, Algorithms of Oppression, A Prehistory of the Cloud, Seeing Like a State, The Audit Society, and Escape from Model Land. Those books explain opacity, classification, infrastructure, legibility, verification, and model error. Amoore adds a harder question: what if algorithmic responsibility cannot be found by isolating one hidden step, one biased variable, one accountable author, or one final output?

Attributes Before Answers

The title's word "attributes" does a great deal of work. A person does not enter a machine-learning system as a full person. A person enters through attributes: location, device, language, image features, purchase history, body pattern, risk marker, label, embedding, faceprint, behavioral trace, medical signal, credit relation, border-crossing history, or some proxy the institution treats as relevant.

An attribute is not a fact in miniature. It is a claim that has been made usable: a sensor reading cleaned into a field, a phrase embedded into a vector, a complaint translated into a category, a photograph passed through a recognition system, or a record linked to a population-level pattern. The same trace can become a different attribute when the purpose, dataset, threshold, or institution changes.

Attributes are not neutral fragments. They are selected, produced, cleaned, weighted, combined, and made actionable inside systems with purposes. A policing platform, an oncology robot workflow, an intelligence-analysis system, a credit model, a border-risk tool, a hiring screen, and a content classifier do not merely observe attributes. They decide which attributes can stand for suspicion, competence, disease, trust, relevance, fraud, promise, danger, or normality.

That is the first AI-era lesson. Before a model gives an answer, a world has been prepared for attribution. The person has become a bundle of machine-available handles. The institution has decided which handles count. The model then makes a relation among those handles operational.

This is why AI governance cannot stop at "what did the model say?" The deeper question is "what was the person allowed to become before the model spoke?" A model-mediated world can make attributes more durable than testimony. It can make correlation look like character. It can make a pattern extracted from others return as a judgment about this person, this case, this worker, this patient, this student, this traveler, or this applicant.

Opacity Is Not Only a Defect

Many algorithmic-accountability debates begin with transparency. Open the black box. Inspect the code. Disclose the variables. Produce the explanation. Trace the data. Those demands matter. Without them, institutions can hide consequential systems behind trade secrecy, vendor contracts, mathematical authority, or the claim that the machine is too complex to question.

Amoore's intervention is to show why transparency is necessary but insufficient. Machine-learning systems are not only opaque because someone refuses to reveal them. They are also opaque because their operation emerges from relations among training data, model architecture, parameters, labels, feedback, human practice, institutional uptake, thresholds, and future use. The system gives partial accounts because both human and algorithmic decisions are partial.

That does not excuse opacity. It removes a comforting fantasy. If accountability is imagined as a moment when the hidden truth of the algorithm is finally revealed, then governance will keep searching for the missing key. But some algorithmic decisions do not have a single key. They have conditions of production, chains of delegation, uncertain thresholds, and institutions that decide how much uncertainty they are willing to act on.

The useful response is not to abandon explanation but to widen it. A source-code explanation may be relevant, but so are the data setting, source provenance, label history, model version, retrieval corpus, user interface, reviewer incentives, escalation rule, and institutional threshold for action. Opacity is often produced by the arrangement, not only by the algorithmic object inside it.

For modern AI systems, this is a live problem. A large language model cannot fully account for why one completion appeared instead of another in ordinary human terms. A recommender system cannot reduce its social effects to a clean list of variables. A risk model may explain feature importance while leaving the social production of those features untouched. A safety dashboard may show a score while hiding which incidents, users, domains, and harms never became part of the evaluation frame.

The Attribution Machine

The strongest way to read Cloud Ethics now is as a study of attribution machines. These are systems that assign something to someone: risk, identity, authorship, relevance, intent, abnormality, trustworthiness, care need, fraud probability, border threat, health prognosis, creditworthiness, employability, or moral status. The machine does not only classify. It attributes a quality that can follow a person into institutional action.

Attribution is powerful because it feels smaller than judgment. A system may not say "this person is dangerous." It may say the record shares attributes with prior cases, the image has features associated with a class, the transaction resembles a pattern, the text has a score, or the case exceeds a threshold. The institution then converts the attribution into consequence.

That conversion is the political act. It is where a score becomes a denial, a flag becomes a search, a label becomes a queue position, a model output becomes a medical pathway, a prediction becomes a management decision, or a generated answer becomes the working memory of an office. Responsibility cannot be placed only inside the model because the harm often appears when the model's partial account is treated as enough.

The book is especially useful against the habit of treating "human in the loop" as a cure. A human reviewer can become the person who ratifies the machine's attribution under time pressure, incomplete evidence, interface nudges, organizational incentives, and fear of being blamed for ignoring a warning. The loop is not automatically democratic because a human appears in it. The question is whether the human has power to inspect, slow, contest, repair, and change the conditions under which the attribution is made.

The same point applies to explanation. A feature-importance chart, chain-of-thought-style rationale, confidence score, or audit summary may help, but none of them alone tells the affected person why an institution accepted one partial account and not another. A useful explanation names the attributes used, the relation inferred, the uncertainty retained, the human authority exercised, and the path for correction.

For high-consequence systems, the practical artifact is an attribution file or decision dossier. It should connect the system inventory entry, data provenance, model or vendor version, prompt or policy version, retrieved records, confidence or uncertainty signal, threshold, human approval or override, notice, appeal path, and retention rule. That file is not a philosophical solution. It is the minimum evidence record needed to make a later challenge more than a complaint against a blank surface.

The AI Reading

In 2026, Cloud Ethics reads like a prehistory of generative AI governance. The book is not primarily about chatbots, agents, retrieval systems, synthetic media, or model-routing layers. Yet its questions have become more urgent because generative AI wraps attribution in fluency.

A chatbot attributes relevance when it chooses which retrieved passages to synthesize. It attributes competence when it scores, summarizes, ranks, or recommends. It attributes normality when it rewrites a person's complaint into institutional language. It attributes authority when it presents a source list. It attributes intent when it moderates content. It attributes risk when it becomes part of fraud detection, child-safety screening, cyber defense, medical triage, classroom discipline, or legal review.

Retrieval-augmented generation makes this concrete. A search index, embedding model, access-control rule, reranker, and context window decide which records become the temporary ground of an answer. Agent systems extend the same pattern to action: they attribute reliability to tools, permission to credentials, priority to tasks, and adequacy to stopping conditions. Synthetic-media detectors and content provenance systems attribute authorship or manipulation. AI search engines attribute salience. None of these attributions is merely descriptive when an institution acts on it.

The interface can make these attributions feel like conversation instead of classification. That is the new danger. A generated answer seems less bureaucratic than a form, less punitive than a score, and less cold than a dashboard. But the same attribute machinery may sit underneath it: embeddings, labels, permissions, retrieval indexes, safety classifiers, memory stores, user profiles, and hidden policies.

This changes the belief problem. People trust conversational systems partly because they answer in the rhythm of social life. Institutions trust them because they convert unstructured material into usable handles. The result is a recursive reality loop: records become attributes, attributes become outputs, outputs shape decisions, decisions create new records, and the new records train or condition future systems.

Amoore helps name the governance failure in that loop. It is not only hallucination, bias, privacy leakage, or lack of explainability. It is the institutional decision to treat partial accounts as sufficient accounts because the machine has made them actionable.

What Accountability Requires

The practical value of Cloud Ethics is its refusal of simple cures. It does not let code inspection, ethical design principles, fairness metrics, transparency portals, or human review become final answers. Those tools can help, but they work only if they remain connected to the social and technical conditions that produce algorithmic judgment.

Accountability therefore has to travel across the whole arrangement. It needs data provenance, model documentation, evaluation design, uncertainty disclosure, appeal rights, independent audit, incident memory, interface review, procurement scrutiny, affected-person participation, and the authority to change or stop a system. It also needs a culture that can tolerate doubt instead of converting every ambiguous model output into operational certainty.

By June 23, 2026, current governance frameworks give that argument operational hooks. The EU AI Act implementation timeline says general-purpose AI rules applied from August 2, 2025, while most rules, including Article 50 transparency obligations and Annex III high-risk rules, begin on August 2, 2026, subject to the Commission's Digital Omnibus note for some high-risk timing. Article 10 makes data governance part of high-risk AI compliance by requiring attention to design choices, data collection and origin, preparation operations such as annotation and cleaning, assumptions about what data measures, bias examination, mitigation, and gaps. Article 27 requires certain deployers of high-risk AI to assess fundamental-rights impacts before first use and to identify affected groups, risks of harm, human oversight, governance, and complaint mechanisms. Article 50 requires clear notice when people interact with AI systems and marking of generated or manipulated content in specified cases.

U.S. federal policy points in the same direction from a different legal structure. OMB Memorandum M-25-21, issued April 3, 2025, directs agencies to apply minimum risk-management practices for high-impact AI and to discontinue use where the system is not performing appropriately and compliance cannot be achieved. It requires documented AI impact assessments before high-impact deployment, data and model fitness summaries, ongoing monitoring, human oversight, traceability where possible, and remedies or appeals where appropriate. NIST's AI Risk Management Framework remains voluntary, but it frames AI risk management across design, development, use, and evaluation; its Generative AI Profile adds a cross-sectoral map of generative-AI risks. ISO/IEC 42005:2025 supplies a standards-body vocabulary for AI system impact assessments across the lifecycle.

Provenance is the bridge between Amoore's political theory and operational governance. W3C PROV offers a common vocabulary for entities, activities, agents, derivation, versioning, and processing steps. The 2025 joint AI Data Security guidance from NSA, CISA, FBI, and international partners recommends reliable data sourcing, provenance tracking, and tamper-resistant records for data moving through AI systems. In attribution-machine terms, that means the question "why was this quality assigned here?" has to preserve a path back through source, transformation, authority, and use.

That last point is central. AI systems are attractive to institutions because they can make doubt manageable. They provide a score, a ranking, a summary, a risk flag, a next action, or a recommended response. But sometimes doubt is the ethical signal. A system that cannot know enough should not be permitted to turn insufficient attributes into authoritative action.

Good governance preserves friction at exactly that point. It asks where the model is uncertain, where the data is partial, where the category was contested, where the affected person can answer, where the institution can be forced to explain itself, and where no automated attribution should be allowed to decide the matter. In practice, that means logs tied to data and model versions, usable notices, human oversight with real authority, pre-deployment impact assessment, post-deployment monitoring, incident reporting, appeal paths, and procurement rights that let the deployer inspect enough evidence to own the decision.

Where the Book Needs Friction

Cloud Ethics is demanding. Its method moves through political theory, geography, philosophy, science and technology studies, and close attention to technical cases. Readers looking for a plain-language AI procurement checklist, a model-risk-management template, or a step-by-step audit procedure will need companion sources.

The book also says less than current readers may want about foundation-model supply chains, data-center infrastructure, annotation labor, platform concentration, copyright conflict, retrieval pipelines, agentic tool use, and the economics of model deployment. Those issues have become unavoidable since 2020. Amoore's framework should be joined to material accounts of extraction, labor, ownership, energy, and platform dependency rather than treated as a substitute for them.

There is also a political risk in emphasizing partiality and opacity. Bad institutions can weaponize complexity. They can say that because no account is complete, no one can be held responsible. A strong reading has to move in the opposite direction: precisely because the account is partial, institutions must carry more responsibility for when they choose to act on it.

The best use of the book is therefore not resignation. It is discipline. Do not pretend the machine is fully knowable. Do not let unknowability become immunity. Build accountability around the uncertain conditions where people, data, models, interfaces, vendors, and institutions compose a decision.

What This Changes

The book changes the unit of analysis. Instead of asking only whether an algorithm is fair, transparent, aligned, accurate, or explainable, ask what arrangement of attributes is being made authoritative. What traces were accepted as evidence? What relation was inferred? What threshold turned relation into decision? What institution acted? What options did the affected person have after the attribution was made?

This applies to AI companions, search answers, medical triage, student models, workplace dashboards, border tools, synthetic-media detectors, content moderation, hiring systems, credit decisions, policing platforms, insurance models, and agentic workflows. The visible output is only the last surface. The ethical problem begins earlier, when the system learns what a person, event, risk, claim, source, or action can be reduced to.

Cloud Ethics is valuable because it resists the lure of machine certainty without retreating into anti-technical moralism. It treats algorithms as sociotechnical participants in judgment, not as neutral tools or alien minds. That is exactly the register AI governance needs. The question is not whether machines have values hidden inside them. The question is how human and machine arrangements generate value judgments, attach them to people, and make institutions comfortable acting on the result.

The lesson is severe: whenever an AI system turns attributes into action, the partial account must remain visible enough to be disputed. Otherwise the attribute becomes the person, the score becomes the fact, the output becomes the institution's memory, and uncertainty disappears at the moment it should have forced responsibility.

That visibility is not just disclosure to the public. It is structured contestability: affected people need notice, reviewers need evidence, auditors need logs, deployers need vendor documentation, and decision-makers need authority to stop or narrow use. A partial account can be used ethically only when it remains revisable.

Source Discipline

This review uses Duke University Press and JSTOR for bibliographic facts, table-of-contents structure, and publisher framing; scholarly reviews for reception; ERC author context for Amoore's research position; and regulator, government, and standards sources for current governance claims.

Those sources support a specific claim, not a mystical one. The review does not treat algorithms as conscious, divine, autonomous moral agents, or AGI. The accountability target is the arrangement that lets people and institutions convert partial machine-generated accounts into action.

Legal and standards claims should be read by scope. The EU AI Act provisions cited here apply by role, system category, date, and jurisdiction. OMB M-25-21 governs U.S. federal agency use of AI, not private-sector systems in general. NIST AI RMF, W3C PROV, AI Data Security guidance, and ISO/IEC 42005 are risk-management, provenance, security, and impact-assessment references; none proves that a particular deployment is safe or accountable.

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