Blog · Review Essay · Last reviewed June 23, 2026

Machine Learners and the Practice Behind Prediction

Adrian Mackenzie's Machine Learners is a useful antidote to AI mysticism because it treats machine learning as a practice: coded, documented, tuned, shared, benchmarked, and argued over by people working inside data cultures.

For this review, data practice means the repeatable work that turns messy records into model-ready reality: collection, cleaning, labeling, feature choice, training, evaluation, benchmark comparison, deployment, monitoring, and the institutional decision that an output is good enough to act on.

The article's practical standard is a pipeline receipt: a deployed model should leave enough trace for a reviewer to reconstruct the task, source data, labels, benchmark, threshold, override path, update history, and incidents that made prediction actionable.

The Book

Machine Learners: Archaeology of a Data Practice was published by the MIT Press in 2017. The MIT Press listing names Adrian Mackenzie as author, gives the paperback ISBN as 9780262537865, gives the hardcover ISBN as 9780262036825, and lists the book at 272 pages. Lancaster University's research record lists the book under Adrian Bruce MacKenzie, gives MIT Press as publisher, records print ISBN 9780262036825 and electronic ISBN 9780262342544, and links to the book's companion code repository.

The page-count discrepancy is a useful reminder about the book's own point. MIT Press currently lists 272 pages for the paperback and hardcover. Lancaster's institutional record lists 280 pages. This review treats MIT Press as the controlling source for retail metadata and Lancaster as the controlling institutional record for the author affiliation, publication record, electronic ISBN, and companion repository.

The book asks what changes when machine learning becomes an ordinary way of making knowledge. Mackenzie is not primarily interested in science-fiction intelligence or corporate demos. He studies machine learning as an operational culture: code examples, data settings, diagrams, mathematical formalisms, social organization, and local habits of judgment that teach practitioners what to count as learning.

That focus makes the book more durable than many AI books from the same period. It does not need a prediction about a future supermind. It studies the quieter transformation: institutions learning to treat prediction as a way of knowing, managing, ranking, and acting.

Current Context

Read on June 23, 2026, the book sits in a different public context than it did in 2017. Machine learning is now encountered through foundation-model products, retrieval systems, copilots, recommender systems, fraud tools, workplace analytics, public-sector decision systems, and tool-using agents. The interface may be conversational, but the governing object is still a data practice.

The regulatory context has also moved toward Mackenzie's terrain. NIST's AI Risk Management Framework frames AI risk management across design, development, deployment, evaluation, and use. Its Core functions - Govern, Map, Measure, and Manage - make risk management a lifecycle discipline rather than a one-time model score. NIST AI 600-1 extends that vocabulary to generative AI. NIST's 2026 AI Agent Standards Initiative adds a current standards agenda for agent security, identity, interoperability, and evaluation.

The EU AI Act makes the same shift legally concrete for covered systems. Article 10 requires data governance practices for high-risk systems that use training, validation, or testing datasets, including design choices, data origin, preparation, assumptions about what data measure, bias examination, mitigation, and relevant data gaps. Article 53 requires general-purpose AI model providers to draw up and make public a sufficiently detailed summary of training content according to an AI Office template; Article 113 makes Chapter V obligations apply from August 2, 2025, while the Regulation's general application date is August 2, 2026. The European Commission published the public-summary template on July 24, 2025.

The current pressure point is update governance. Retrieval corpora change, synthetic data enters tuning runs, benchmarks saturate, prompts and policies shift, adapters are swapped, and agent permissions expand. A system that looked bounded at launch can become a different data practice after a corpus refresh, threshold change, vendor update, workflow integration, or new tool permission.

Those sources do not prove that every AI deployment is dangerous or that every model has the same risk. They do show why Mackenzie's language of practice is now governance language. The practical question is no longer only "what is the model's benchmark score?" It is "what data practice, documentation practice, and accountability practice made this system deployable?"

Practice Before Oracle

The strongest move in Machine Learners is its refusal to treat the model as an oracle. Machine learning appears magical when the interface hides the work that precedes a prediction. Mackenzie puts that work back into view. Datasets are selected. Variables are shaped. Labels are assigned. Libraries carry assumptions. Benchmarks reward particular kinds of performance. Diagrams, tables, error measures, and code examples train practitioners in what to notice.

That emphasis relocates agency. The machine does not think from nowhere. It is assembled inside communities of practice that make some problems tractable and others invisible. This is not a claim that machine learning is fake. It is a claim that its reality is practical and institutional before it is metaphysical.

The useful definition is therefore narrow and demanding. A machine-learning system is not just a trained artifact. It is a sequence of choices about task framing, data access, data cleaning, labels, features, model class, optimization target, evaluation metric, deployment threshold, monitoring regime, and authority to act. A system can be technically impressive and still fail if those choices are undocumented, misaligned with the setting, or impossible for affected people to contest.

The review standard follows from that definition: do not ask only to see the model. Ask to see the practice. A serious review should recover the problem statement, source records, rights basis, label rules, evaluator instructions, metric choice, release threshold, deployment boundary, human override, monitoring plan, and incident route.

This is why the book belongs beside reviews of Data Feminism, Sorting Things Out, and How Data Happened. Each shows that records are not passive mirrors. They are built artifacts that can become institutional reality when a workflow gives them force.

The Pipeline Is the Object

Mackenzie's deepest contribution is to shift attention from the model to the pipeline. A model output is the last visible event in a longer chain. The earlier choices decide what the model can know, what it cannot know, which errors are easy to see, which errors disappear, and who is left with the burden of correction.

For safety, that means the audit target is the whole path from source data to acted-upon decision. The pipeline includes original collection purpose, consent or legal basis where relevant, data provenance, preprocessing, filters, label instructions, annotator conditions, feature engineering, train-validation-test splits, evaluation sets, benchmark selection, release criteria, logging, drift monitoring, incident reporting, and withdrawal authority. A model card without a data record is incomplete; a data sheet without deployment context is also incomplete.

Modern failure modes make this practical rather than academic. Data leakage can make a benchmark look stronger than deployment reality. Shortcut learning can reward a proxy that works in the dataset but fails in the world. Distribution shift can break performance after a population, policy, device, or language changes. Feedback loops can make a prediction alter the behavior it was built to measure. A recidivism score, hiring ranker, credit model, fraud detector, recommender, or agentic workflow can all become worse than its original test suggests once people adapt around it.

Good governance therefore has to preserve pipeline memory. It should be possible to reconstruct what data entered, what was removed, who labeled it, which assumptions were made, which metrics won, what thresholds triggered action, and who had authority to override or stop the system. Without that memory, "AI governance" becomes a front-end promise attached to a back-end mystery.

A pipeline receipt is not perfect reconstruction. It is the minimum evidence needed to dispute a consequential prediction without guessing at invisible preprocessing, hidden exclusions, stale evaluations, or private policy changes.

Prediction as Culture

Mackenzie helps explain why prediction has become a cultural form. A predictive model does more than output a score. It teaches an institution to see the world as a field of features, labels, losses, correlations, and future-facing interventions. Once that style spreads, people begin to ask model-shaped questions: what can be classified, optimized, recommended, ranked, or routed?

This is the quiet hinge between machine learning and belief. A model's output becomes persuasive not only because it is statistically impressive, but because the surrounding organization has learned to want that kind of answer. Prediction becomes believable when dashboards, procurement plans, research papers, product metrics, management incentives, and compliance documents all point in the same direction.

The model does not need consciousness to acquire authority. It needs a workflow ready to receive it. Once the workflow is ready, the score can become a ticket, a denial, a recommendation, a fraud flag, a ranking, a classroom intervention, a maintenance schedule, a workplace warning, or a next action in an agent chain.

That is also why prediction can become self-confirming. If a system routes more scrutiny toward a group, the resulting records may look like evidence that the group deserved scrutiny. If a recommender pushes attention toward a topic, the resulting engagement can look like evidence of natural interest. If a workplace dashboard measures only visible transactions, hidden coordination and care work may look like inefficiency. The data practice does not merely report culture. It can train culture to behave in ways the model knows how to count.

The safety issue is not only error; it is institutional training. Once employees, students, applicants, claimants, patients, or users learn what the system rewards, they may reorganize around the measurable proxy. The model then measures a world partly shaped by its own incentives.

Benchmarks and Public Memory

The book is especially helpful on the social life of benchmarks. Benchmarks are not neutral scoreboards floating above practice. They select tasks, datasets, metrics, and comparison rituals. They teach a field what progress looks like.

That can be productive. Shared evaluation helps researchers compare systems, find errors, and avoid pure marketing claims. But a benchmark also becomes a target. Once funding, publication, procurement, and product launches reward a score, the score begins to shape the work. Teams optimize for the public test, benchmarks leak into training sets, and capabilities that are hard to measure can disappear from the official story of progress.

The governance lesson is not to reject benchmarks. It is to demote them to evidence within a larger case. A high benchmark score should travel with dataset provenance, evaluation methodology, known limitations, subgroup performance where relevant, deployment assumptions, monitoring plans, and incident evidence after release. The review of AI Snake Oil makes a similar point from the prediction-hype side: predictive success in one setting does not license authority everywhere.

Public memory depends on this discipline. If the only surviving record is a leaderboard, the institution remembers victory and forgets the data practice that made victory possible. Model cards, data sheets, system cards, evaluation reports, and incident logs are memory devices. They preserve enough context for later users to ask whether the system still deserves trust.

The Agent Reading

Read in 2026, the book is especially useful for understanding AI agents. Agentic systems are usually discussed at the level of autonomy: what can they do without a person clicking each step? Mackenzie suggests a better question: what data practices make that autonomy legible, testable, and deployable?

An agent is not just a chat window with tools. It is a chain of prompts, retrieval rules, memory stores, permissions, evaluations, logs, handoffs, secrets, file access, network access, identity, and stopping conditions. Each part encodes a practice of judgment. The user may experience a smooth assistant, but the actual system is a practical culture of thresholds and conventions.

NIST's AI Agent Standards Initiative is useful here because it treats agents as a standards problem around secure operation, identity, authorization, interoperability, and evaluation. That is the right level of analysis. An agent mistake is not only a wrong sentence. It may be a wrong action in a filesystem, account, procurement flow, customer record, source repository, calendar, claims process, or ticket queue.

If the conventions are weak, the agent can convert provisional outputs into administrative action faster than accountability can follow. Agent governance should therefore ask practice-level questions: what sources can the agent read, what tools can it call, what identity does it act under, what logs survive, which actions require approval, what evidence is shown to the human reviewer, and how can an affected person challenge the result?

The site's Agent Tool Permission Protocol, Agent Audit and Incident Review, and agent sandbox essay extend Mackenzie's lesson from prediction to action. The data practice becomes a permission practice. The audit object becomes the whole action loop.

Governance of Pipelines

NIST's AI Risk Management Framework describes AI risk management through Govern, Map, Measure, and Manage, and presents AI risk as work across design, development, deployment, evaluation, and use. Read beside Mackenzie, that vocabulary becomes more concrete. The object to govern is not only a model. It is the pipeline that makes the model meaningful.

That means governance has to ask practice-level questions. Who defined the task? What data were excluded? Which benchmark became the target? What errors are tolerated because they are convenient to measure? Who can inspect the model's effect after deployment? What happens when the tool changes the behavior it was built to predict? A governance process that answers only "is the model accurate?" has already accepted too much of the machine-learning worldview.

Documentation practices are the bridge between critique and control. Datasheets for Datasets proposes documenting a dataset's motivation, composition, collection process, recommended uses, and related information. Data Cards treat dataset documentation as a user-centered product for stakeholders across the dataset lifecycle. Model Cards propose concise reporting on intended use, performance characteristics, evaluation procedures, and limits, including performance across relevant groups where applicable. The Data Provenance Initiative's 2024 audit of more than 1,800 text datasets found major gaps in license and attribution metadata, including widespread unspecified licenses and frequent miscategorization on popular hosting sites.

Those tools are not paperwork for paperwork's sake. They are ways to keep a model from laundering its history. A deployer should be able to see training data provenance, evaluation data provenance, label rules, model limits, human-feedback sources, intended use, disallowed use, known gaps, update cadence, monitoring duties, and incident response. If those records do not exist, the buyer is inheriting unknown legal, ethical, and operational risk.

A practice-level audit should ask whether a reviewer can reproduce the data lineage from source record to evaluation set; whether an affected person can challenge the category as well as the score; whether monitoring tests drift, feedback loops, subgroup performance, proxy leakage, and benchmark contamination after deployment; whether synthetic data, generated labels, model-assisted judgments, and human-feedback sources are logged; and whether a human owner can pause or withdraw the system when the practice fails.

The EU AI Act points in the same direction for systems within its scope. Article 10's data governance requirements for high-risk systems are pipeline requirements: data origin, preparation, assumptions, suitability, bias examination, mitigation, and gaps. Article 53's public training-content summary for general-purpose AI models is narrower than full provenance, but it sets a public baseline. The template does not replace internal records, licensing review, dataset documentation, labor disclosure, or deployment-specific evaluation. It is a floor, not the whole house.

For high-impact deployments, the minimum safety case should include task justification, data provenance, dataset documentation, model documentation, evaluation evidence, human oversight, logging, cybersecurity, privacy analysis, affected-person notice where required, appeal or correction path, monitoring, incident reporting, and a withdrawal plan. That is the practical version of Mackenzie's argument: make the practice visible before the output governs anyone.

Change Control and Drift

The article's pipeline receipt should not end at launch. A model is often approved once and then altered many times through quieter changes: a new data vendor, retrained classifier, refreshed retrieval index, changed prompt, threshold adjustment, policy update, safety-filter replacement, adapter swap, synthetic-data run, or expanded agent tool permission. Each can change the data practice even when the product name and interface remain stable.

The 2025 joint AI Data Security guidance from NSA, CISA, FBI, and international partners makes this operational. It identifies data supply-chain risk, maliciously modified data, and data drift as AI data-security concerns, and recommends provenance tracking, digital signatures, secure storage, trusted infrastructure, and lifecycle data protection. Read beside Mackenzie, that guidance turns "practice" into an incident-response requirement: if the system changes, the evidence trail has to show what changed, why it changed, who approved it, and how the old behavior can be recovered or retired.

A change-control record should therefore attach to the pipeline receipt. It should name the artifact changed, reason for change, affected users, data sources added or removed, evaluation repeated, regressions found, approval owner, rollback path, monitoring trigger, and notices or appeals required. For agentic systems, it should also name tool scopes, credentials, memory rules, sandbox boundaries, and approval gates. Without that record, monitoring becomes a dashboard without memory.

Drift is not only a statistical problem. It is institutional drift when a model built for one decision quietly becomes evidence for another, when a benchmark becomes procurement proof, when an internal assistant becomes a customer-facing agent, or when a pilot system becomes infrastructure before the affected population can contest it. The recurring risk is that a machine-readable practice acquires authority faster than its documentation, governance, or recourse can keep up.

The control layer is AI change management, model drift, AI post-market monitoring, AI system inventory, and AI incident reporting. Those pages extend Mackenzie's point from making machine learning to keeping deployed machine learning answerable over time.

Where the Book Needs Care

Machine Learners is dense and sometimes more archaeological than argumentative. Readers looking for an accessible account of AI harm, surveillance capitalism, labor extraction, or civil-rights impact will need companion texts. The book is strongest on how knowledge practices form, not on how specific communities experience automated decisions.

Its other limitation is historical timing. Published in 2017, it predates the public explosion of large language models, retrieval-augmented generation, synthetic-data loops, multimodal foundation models, and tool-using agents. But that limitation is also why it remains useful. Before the current layer of generative fluency, Mackenzie described the deeper habit: turning uncertain reality into a data practice that can be trained, measured, benchmarked, and operationalized.

The book can also tempt readers into abstraction. "Data practice" is useful only if the analysis stays close to the actual workflow. Which dataset? Which labels? Which benchmark? Which model family? Which deployment threshold? Which affected group? Which institutional decision? Which remedy? Without those questions, critique can become as vague as the hype it opposes.

Finally, the book is not an argument against machine learning as such. Some models improve science, accessibility, translation, logistics, medicine, climate analysis, and public administration. Mackenzie's value is that he makes those successes accountable to the practices that produce them. A useful system still needs a record of how it was made, where it works, where it fails, and who can stop it when the setting changes.

What This Changes

Machine Learners changes the first question. Do not begin by asking whether an AI system is secretly alive, prophetic, or inevitable. Begin by asking what practice gave it authority.

The practical checklist is direct: name the task, name the data source, preserve the category definitions, document cleaning and labeling, separate training from evaluation, disclose benchmark limits, test in the deployment setting, monitor after release, log consequential actions, provide correction and appeal paths, and assign a human owner with power to pause the system.

The deeper lesson is institutional. Machine learning becomes powerful when organizations reorganize around machine-readable reality. Once a system is embedded in procurement, dashboards, policies, workflows, and management incentives, the model's output can feel like common sense. The way to resist false authority is not to romanticize human judgment. It is to keep the whole practice inspectable.

That is the concrete tie to the site's recurring themes: belief, power, and public memory are now produced through records, interfaces, metrics, and automated action loops. A society that forgets the practice behind prediction will mistake operational convenience for truth. If no one can name the practice, no one should treat the prediction as governance-grade.

Source Discipline

This review separates book facts, interpretive claims, current governance claims, security guidance, and implementation claims. MIT Press and Lancaster University support bibliographic details. NIST sources support the risk-management and agent-standards context. EUR-Lex, the AI Act Service Desk, and the European Commission support claims about the EU AI Act and the training-content summary template. NSA, CISA, FBI, and partner guidance supports the data-security, provenance, and drift-control claims. Dataset documentation papers and the Data Provenance Initiative support the claims about datasheets, Data Cards, Model Cards, and provenance debt.

The AI-era reading is an application of Mackenzie's framework, not a claim that the book predicted every feature of 2026 AI systems. This page does not claim that AI systems are conscious, divine, or AGI. It treats them as institutional machinery that can classify, predict, recommend, summarize, and act when people build workflows around them.

Source discipline also means refusing to let product claims settle the case. A benchmark, launch blog, model card, or regulatory filing may be useful evidence, but none replaces independent evaluation, deployment logs, affected-person testimony, incident reports, procurement records, and the data documentation needed to understand how the system was made. Current legal, standards, and publication claims on this page were checked against primary or publisher sources on June 23, 2026.

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