The Seductions of Quantification and the Indicator Machine

The Book

The Seductions of Quantification was published by the University of Chicago Press in 2016. The publisher lists the cloth edition at 272 pages, with ISBN 9780226261287, and frames the book around the rise of global indicators for human rights, gender violence, and sex trafficking. The table of contents moves from the knowledge effects of indicators through human-rights measures, violence-against-women measures, sex-trafficking measures, local resistance, and the politics of measuring law and injustice.

Merry was a legal anthropologist at New York University whose work studied human rights, law, gender violence, colonial and postcolonial legal orders, and the translation of global norms into local settings. The official publisher page and university event materials emphasize that the book asks what happens when complex social problems are converted into numerical forms that can travel across international organizations, states, funders, advocacy networks, and reform programs.

This makes the book a necessary companion to Trust in Numbers, Seeing Like a State, The Audit Society, The Tyranny of Metrics, and The Ordinal Society. Those books explain why numbers become credible, why states and organizations simplify the world, why verification rituals spread, why metrics distort missions, and why ranking becomes a social order. Merry adds the missing translation layer: the process by which moral claims and lived harms are turned into indicators fit for comparison.

The Indicator as Translation Device

Merry's core insight is that an indicator is not just a measurement. It is a translation device. It takes a thick social condition, selects a few features, defines categories, assigns values, and makes the result portable. Once portable, the number can be compared, ranked, cited, funded, audited, displayed, and acted upon far from the communities and institutions that produced it.

More precisely, an indicator is a governed translation. It has a construct, a data source, a collection process, a denominator, an aggregation rule, a normalization method, a comparison frame, a display surface, and a decision setting. Each element carries judgment. The safety question is therefore not simply whether the number is accurate. It is what the number has been authorized to do.

The whole chain matters: source, category, rule, score, ranking, consequence. A contested definition of "trafficking," "violence," "risk," "capability," "toxicity," "quality," or "safety" can be hardened into a form field, then into a dataset, then into a dashboard, then into an institutional fact. Once that chain is operating, later users may see only the final number and not the decisions that made it.

That portability is powerful. International human-rights organizations need a way to compare states. Funders need a way to decide where money goes. Governments need evidence of compliance or progress. Advocates need public pressure. Journalists need a legible signal. A number can move through those settings more easily than testimony, case files, ethnography, legal detail, or local history.

The cost is compression. Definitions of violence, trafficking, rights compliance, reporting capacity, legal reform, and enforcement become part of the indicator. A state that records more cases may look worse because reporting improved. A state that criminalizes conduct may score differently from one that changes services, labor conditions, migration policy, or police practice. A jurisdiction's number can reflect harm, capacity, politics, paperwork, and classification choices all at once.

The indicator therefore creates a new object. It is related to the underlying social condition, but it is not identical to it. The danger begins when institutions forget the difference.

The Seduction

The title's word "seductions" is exact. Quantification attracts because it promises clarity without intimacy. It seems to let distant actors know what is happening, compare the incomparable, and make moral judgment operational. It offers a clean interface for messy reality.

That attraction is not foolish. Without numbers, severe harms can stay anecdotal, local, deniable, and easy for powerful actors to ignore. Indicators can make violence visible, force governments to report, create shared agendas, and help advocates show that a problem is structural rather than isolated. Merry's critique is not an argument for abandoning measurement.

Her warning is subtler. Indicators do not only describe policy fields. They reshape them. Organizations learn what must be counted. Governments learn what international bodies want to see. Reformers learn which categories funders recognize. Local actors may change language to fit global forms. A moral project becomes a reporting architecture, and the reporting architecture begins to feed back into the moral project.

This is the same recursive pattern that now runs through data-driven institutions: measure the world, act on the measure, change the world, then treat the changed measure as evidence. The loop can support accountability, but it can also create a model of reality that becomes more authoritative than the people and places it claims to represent.

Current Context

As of June 23, 2026, the indicator problem is no longer limited to human-rights reporting. It appears in AI benchmarks, safety scorecards, model cards, system cards, content-moderation reports, public-sector dashboards, vendor evaluations, procurement rubrics, and risk registers. These artifacts can improve accountability, but they also create the same hazard Merry describes: compressed translations travel farther than the evidence and context that produced them.

Current primary governance sources make that tension explicit. NIST's AI Risk Management Framework places measurement inside a broader cycle of govern, map, measure, and manage, and its Core treats risk work as continuous across the AI lifecycle rather than as a single score. NIST's Generative AI Profile extends the same discipline to generative systems by tying evaluation to lifecycle context, documentation, and risk management rather than treating a benchmark as a complete safety case.

NIST's AI test, evaluation, validation, and verification work makes the same point from the measurement side. NIST says trustworthy AI products and services depend heavily on reliable measurements and evaluations of the technologies and their use, and it identifies accuracy, robustness, bias, interpretability, and transparency as part of the measurement agenda. That framing matters because it refuses the easiest seduction: the idea that a public score is the evaluation. The evaluation is the designed inquiry that produced the score, the limits around it, and the decision it is meant to inform.

The EU AI Act gives the same problem legal form for high-risk AI systems. Article 12 addresses logging, Article 13 requires transparency and instructions that help deployers interpret outputs, Article 14 covers human oversight, Article 26 sets deployer obligations, and Article 27 requires certain deployers to perform fundamental-rights impact assessments before first use. In U.S. federal practice, OMB M-25-21 requires minimum risk-management practices for high-impact AI, including impact assessment, monitoring, human oversight, and appeal or remedies where appropriate. ISO/IEC 42001:2023 frames AI governance as a management system, while ISO/IEC 42005:2025 gives guidance for AI system impact assessment across foreseeable effects on individuals, groups, and society.

The practical implication is modest and important: a modern AI number should not travel alone. It should travel with purpose, scope, provenance, version, uncertainty, intended use, exclusion limits, and recourse. Otherwise an institution may mistake a useful indicator for a settled account of reality.

The AI Reading

AI systems extend the indicator machine. They do not usually begin with raw reality. They begin with quantified and categorized traces: labels, reports, scores, risk factors, forms, tickets, complaints, incident records, benchmark tasks, performance metrics, survey responses, rankings, disciplinary codes, and administrative histories. Merry helps explain why those inputs already carry politics before a model touches them.

A foundation-model benchmark is an indicator. It compresses capability into a score that can travel through product pages, investor decks, procurement memos, journalism, regulatory debates, and public belief. A risk score in welfare, insurance, hiring, policing, education, or content moderation is an indicator. It compresses a person, case, or event into a number or category that can trigger action. A dashboard for model safety, workplace productivity, civic service delivery, or platform trust is an indicator system. It turns uncertainty into a set of visible handles.

The same is true of safety evaluations. A red-team pass rate, toxicity score, cyber capability result, refusal metric, model-card summary, incident count, or benchmark leaderboard can be useful evidence. None of it is a full account of the system. The score depends on test design, sampled scenarios, threat model, evaluator access, version, deployment channel, user interface, monitoring capacity, and the incentives attached to the result. A result from a controlled development benchmark is not the same as evidence from a live deployment with real users, real incentives, and real failure reports.

The important boundary is where a score becomes a control. A benchmark percentage used for research comparison is one kind of artifact. The same number used as a release gate, procurement shortcut, compliance token, investor claim, school ranking, worker score, benefit flag, or policing priority is a different artifact. The governance burden rises when the indicator stops informing judgment and starts authorizing action.

The AI-era mistake is to treat the model as the first moment of abstraction. Merry shows that abstraction often happened earlier: when harm became category, category became form field, field became dataset, dataset became indicator, and indicator became institutional truth. The model may automate or intensify the process, but it inherits a world already shaped for machine reading.

This matters for governance. A model can be technically impressive and still be built on bad translations. It can optimize a number whose meaning is unstable. It can detect patterns in reporting behavior rather than underlying harm. It can rank systems that differ most in documentation capacity. It can convert historical institutional bias into predictive confidence. The problem is not only opacity inside the model. It is false clarity before the model.

When Measurement Governs

Merry's book is especially useful because it studies measurement as governance. Indicators do not need police powers to govern. They govern by defining problems, setting agendas, allocating attention, conditioning funding, shaping reputations, and creating incentives for reform performances. A state, agency, company, school, platform, or model lab can become oriented toward the measure because the measure is how it is seen.

That is exactly the terrain of modern AI policy. Model evaluations, safety cases, impact assessments, incident databases, risk registers, audit reports, system cards, transparency filings, and public leaderboards can all improve accountability. They can also become the objects that institutions optimize for while affected people remain outside the loop.

The governance question is therefore not "Should we measure?" It is "Who defines the measure, who can contest it, what does it hide, and what power follows from it?" If an AI benchmark is treated as proof of general intelligence, what domains are missing? If a fairness metric is treated as proof of civil-rights compliance, what harms escape the categories? If a safety score becomes a release gate, what incidents count as evidence? If a public-sector dashboard is treated as transparency, can the person governed by it do anything with the number?

The strongest governance use of indicators is therefore adversarial in a civic sense. A good metric should make a claim visible enough to dispute. It should invite questions about construct validity, data provenance, missing cases, subgroup effects, incentives, and decision consequences. A bad metric makes the institution harder to question because disagreement is treated as ignorance of the number.

Measurement becomes legitimate only when it remains attached to explanation, uncertainty, local knowledge, appeal, and repair. Detached from those conditions, it becomes a clean surface over unresolved power.

Governance and Safety

As of June 23, 2026, official AI governance materials have made Merry's problem operational. NIST's AI Risk Management Framework organizes risk work around governance, mapping, measurement, and management, while treating risk management as continuous across the AI lifecycle rather than a one-time score. The EU AI Act's high-risk framework likewise turns measurement into a documentation-and-accountability problem: logging, instructions for use, transparency to deployers, human oversight, deployer obligations, and fundamental-rights impact assessment all ask what evidence exists, who can interpret it, and what safeguards surround use.

The lesson for AI safety is concrete. A consequential indicator needs a stated purpose, validity domain, dataset provenance, collection incentives, version history, evaluation design, subgroup limits where relevant, uncertainty, known gaps, gaming risk, audit trail, human override, affected-person notice when people are governed by it, and a repair or appeal route. Without those conditions, an indicator can become an actuator: a number that does not merely describe a system but releases it, blocks a benefit, changes a price, flags a worker, deprioritizes a user, ranks a school, or allocates enforcement.

A useful safety review should therefore create an indicator dossier before it accepts the indicator's authority. The dossier should name the construct, source records, collection setting, denominator, aggregation and weighting rules, update cadence, uncertainty, intended use, prohibited uses, exclusion limits, affected-person path, audit trail, gaming risk, owner, and retirement trigger. If the number is attached to a release gate, benefit decision, enforcement priority, employment action, or public claim, the dossier is not optional bookkeeping. It is part of the evidence for whether the system should act.

The dossier should also classify the indicator's role. Learning indicators support inquiry. Monitoring indicators trigger attention. Command indicators change access, money, status, discipline, visibility, release, or enforcement. A command indicator needs stronger evidence, stricter change control, notice where people are affected, human override, post-deployment monitoring, and a route to suspend the metric when gaming, drift, or unmeasured harm appears.

That is why impact assessments and audits cannot stop at the metric layer. They have to ask how the target was chosen, how missing data behaves, how reporting incentives change under scrutiny, which communities become easier to count than to hear, and what happens when a dashboard contradicts lived evidence. The failure mode is not only bias or inaccuracy. It is administrative closure: the institution stops asking because the indicator has answered.

Where the Book Needs Friction

The Seductions of Quantification is strongest on human-rights indicators and the anthropology of global governance. It is not a technical manual for machine-learning evaluation, algorithmic auditing, statistical modeling, or benchmark design. Readers looking for model-validation procedures will need other sources.

The book also risks being misread as anti-number. That would weaken it. The real argument is not that numbers corrupt moral life by existing. It is that numbers have social lives. They are made by institutions, backed by definitions, used by actors, circulated through incentives, and mistaken for neutral facts when their construction disappears.

A serious reading therefore keeps both sides visible. Some harms require counting because otherwise they remain private, episodic, and deniable. But the count must never be allowed to become the whole moral field. Testimony, legal detail, local interpretation, historical context, and affected people's own accounts remain necessary because indicators are maps, not territory.

What This Changes

The practical lesson is to inspect every AI-era number as a translation, not a revelation. Before asking whether a model used a metric well, ask how the metric made the world legible. What did it compress? What definitions did it enforce? What reporting systems produced it? Who benefits when it travels? Who loses the ability to answer back?

This applies to AI benchmarks, safety ratings, trust scores, productivity metrics, policy dashboards, content-moderation statistics, risk scores, companion-safety claims, and public-sector AI inventories. The score is never just a score. It is a social object with a production chain and a destination.

Merry changes the AI governance conversation by moving attention upstream. The model is not the only machine. The indicator machine comes first: categories, forms, reports, rankings, dashboards, funding rules, legal definitions, procurement rubrics, and public comparisons. AI makes that machine faster and more persuasive. It can also make the machine harder to challenge because the old compression is hidden inside a new interface.

The deeper warning is about belief formation inside institutions. People believe what their systems are built to recognize. If an organization recognizes only the indicator, the indicator becomes reality for practical purposes. The work of governance is to keep the translation visible enough that people can still dispute it, correct it, and refuse to let the number become a substitute for the world.

Source Discipline

This review separates five source layers. Book facts come from the publisher, bibliographic records, interviews, and scholarly reviews. The conceptual claim comes from Merry's analysis of indicators as knowledge technologies and governance instruments. The current AI-governance claims come from primary regulatory, policy, and standards sources. The internal vocabulary comes from related pages on legibility, metrics, audits, impact assessments, recourse, and transparency. The AI-era interpretation is an analogy: benchmarks, risk scores, dashboards, and safety metrics are not identical to human-rights indicators, but they share the same problem of portable compression under institutional pressure.

For any indicator discussed on this site, source discipline means preserving the construct definition, data source, collection process, missing-data pattern, denominator, weighting rule, aggregation method, update date, decision consequence, and contestability path. For AI benchmarks and safety scores, it also means distinguishing a vendor claim from an independent evaluation, a development benchmark from deployment evidence, a leaderboard from an audit record, a compliance artifact from a repair mechanism, and a public dashboard from proof that affected people can contest the decision.

Official sources also need bounded reading. A statute or regulator page establishes duties and definitions; it does not prove compliance. An ISO page establishes that a standard exists and what it covers; it does not prove an organization is well governed. A NIST framework is voluntary guidance unless adopted into a contract, policy, or law. A public benchmark result is evidence about a test setting, not proof that a deployed product is safe for every workflow.

The bounded claim is that Merry's account of indicators explains why AI-era numbers become institutionally persuasive. This page does not claim that all measurement is corrupt, that every benchmark is useless, or that any AI system is conscious, divine, or AGI.

Related Pages

• Trust in Numbers on why quantified objectivity becomes attractive when social trust is weak.
• The Tyranny of Metrics on metrics, dashboards, and proxy optimization.
• The Audit Society on verification as institutional ritual.
• Seeing Like a State on legibility, simplification, and administrative power.
• The Ordinal Society on ranking as social infrastructure.
• Weapons of Math Destruction on opaque scores, feedback loops, and recourse.
• The Benchmark Becomes the Curriculum, How Data Happened, and Sorting Things Out on the production chains behind machine-readable categories.
• AI Governance, Algorithmic Impact Assessments, AI Audits and Assurance, AI Evaluations, and Reward Hacking for the governance vocabulary around indicators that become targets.
• Algorithmic Transparency, Notice and Appeal, Model Cards and System Cards, AI Data Provenance, Transparency and Public Registers, and Claim Hygiene Protocol for keeping indicator claims contestable.
• AI System Inventory, AI Audit Trails, AI Post-Market Monitoring, and Algorithmic Recourse carry the indicator dossier into operational records, live monitoring, and repair.

Sources

• University of Chicago Press, The Seductions of Quantification: Measuring Human Rights, Gender Violence, and Sex Trafficking, official publisher record, publication date, page count, ISBN, description, author note, and table of contents, reviewed June 23, 2026.
• New Books Network, interview with Sally Engle Merry about The Seductions of Quantification, November 7, 2016, reviewed June 23, 2026.
• Stanford Social Innovation Review, Fay Twersky, Can't Count on It, review of The Seductions of Quantification, Summer 2016, reviewed June 23, 2026.
• UCL Discovery, Catalina Turcu, The Seductions of Quantification review record, Critical Policy Studies, 2017, DOI and bibliographic metadata, reviewed June 23, 2026.
• Cambridge Core, Nicola Henry, The Seductions of Quantification review record, Law & Society Review, Volume 51, Issue 2, June 2017, DOI and bibliographic metadata, reviewed June 23, 2026.
• University of Leeds School of Law, event page for Sally Engle Merry's lecture on the politics of measuring human rights and gender violence, author bio and lecture summary, reviewed June 23, 2026.
• NIST AI Resource Center, AI Risk Management Framework Core, govern, map, measure, and manage functions, continuous lifecycle framing, and multidisciplinary perspective, reviewed June 23, 2026.
• NIST, AI test, evaluation, validation and verification (TEVV), official measurement and evaluation program page, reliability of measurements and evaluations, and accuracy, robustness, bias, interpretability, and transparency context, reviewed June 23, 2026.
• NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 publication record, lifecycle and risk-management framing, reviewed June 23, 2026.
• European Union, Regulation (EU) 2024/1689, Artificial Intelligence Act, official EUR-Lex text, especially Articles 12, 13, 14, 26, and 27 on record-keeping, transparency, human oversight, deployer obligations, and fundamental-rights impact assessment, reviewed June 23, 2026.
• Office of Management and Budget, M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, April 3, 2025, high-impact AI risk-management practices, reviewed June 23, 2026.
• ISO, ISO/IEC 42001:2023 Artificial intelligence management system, official standard page, reviewed June 23, 2026.
• ISO, ISO/IEC 42005:2025 AI system impact assessment, official standard page, reviewed June 23, 2026.

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

• Amazon, The Seductions of Quantification by Sally Engle Merry.

Return to Blog · Return to Books