Blog · Review Essay · Last reviewed June 23, 2026

Simians, Cyborgs, and Women and the Human-Machine Boundary

Donna Haraway's Simians, Cyborgs, and Women is often remembered for "A Cyborg Manifesto," but the whole collection matters for the AI era. It teaches that the boundary between human and machine is not a line waiting to be crossed in the future. It is a political arrangement already being made through laboratories, workplaces, databases, military systems, bodies, stories, and institutions.

The human-machine boundary, in this review, is the operational line that decides when a person is treated as author, user, worker, patient, suspect, dataset, biometric template, supervised decision-maker, or governed object. Haraway's value is that she makes that line inspectable: not a metaphysical border, but a design choice with bodies, labor, evidence, and liability on both sides.

That turns cyborg theory into a concrete audit discipline. The question is not whether humans and machines have fused in some grand destiny. It is who draws the boundary in a given system, what evidence the boundary creates, who benefits from treating it as natural, and what an affected person can do when the boundary misclassifies, extracts from, or governs them.

The Book

Simians, Cyborgs, and Women: The Reinvention of Nature was published by Routledge as a first edition with copyright 1991. Routledge currently lists the book at 312 pages, ISBN 9780415903875, with a December 12, 1990 publication date. Internet Archive library metadata lists a 1991 New York Routledge edition with x, 287 pages plus plates, bibliography, and index. Google Books lists a Routledge digital record at 312 pages.

The book collects ten essays written between 1978 and 1989. Routledge groups them into three parts: nature as production and reproduction, contested readings of nature, and differential politics. The table of contents includes "A Cyborg Manifesto," "Situated Knowledges," and "The Biopolitics of Postmodern Bodies." UC Santa Cruz's directory for Haraway lists the original 1985 Socialist Review publication of "Manifesto for Cyborgs" and the 1988 Feminist Studies publication of "Situated Knowledges."

Haraway came to these essays as a biologist, historian of science, feminist theorist, and science-and-technology-studies scholar. UC Santa Cruz lists her areas as science, technology, medicine studies, feminist theory, histories of animal-human relationships, cultures of nature and environment, and science and politics. A Berkeley Townsend Center profile describes her as a theorist of relationships between people and machines whose cyborg work became widely taught and anthologized.

What the Cyborg Means

For this review, a cyborg is not a future robot, a metal-body fantasy, or evidence that current AI systems are conscious. It is a name for a sociotechnical arrangement in which bodies, machines, data systems, labor, law, infrastructure, and stories jointly organize action.

The useful unit of analysis is therefore not "the human" on one side and "the machine" on the other. It is the arrangement: user, model, interface, dataset, permissions, vendor, institution, labor process, audit trail, regulatory setting, and non-use rule. A worker using an AI assistant, a patient routed through a triage chatbot, a student evaluated by analytics, or a citizen sorted by a biometric system is already inside a composite system that changes what agency, evidence, and responsibility mean.

This is why the cyborg is a governance concept as much as a cultural image. It forces a page-level question for every AI deployment: where does the system claim human authority, where does it automate classification, where does it hide dependency, where does it convert a body or relation into data, and where can an affected person interrupt the loop?

This definition cuts against two bad readings. It rejects machine worship, which treats the hybrid as proof that the machine deserves personhood or deference. It also rejects human-purity nostalgia, which pretends institutions were humane before software arrived. Haraway's cyborg is most useful when it keeps attention on the boundary-making work: who is classified, who is augmented, who is disciplined, who can refuse, and who is accountable when the hybrid system harms someone.

Boundary Work

The title gives the method. Simians, cyborgs, and women are not random examples. They are figures through which Western science and politics repeatedly make boundaries: human and animal, organism and machine, nature and culture, male and female, objective fact and social story.

That makes the book directly relevant to AI, even though it predates neural networks as public infrastructure. Most AI debate still turns on boundary questions. Is this output thought or imitation? Is the user acting, or is the system steering? Is the model a tool, a worker, a companion, an institution, a product, or a collaborator? Is the dataset a neutral sample of the world, or a historical record of power, exclusion, labor, and classification?

Haraway's answer is not to dissolve every distinction into fog. It is to ask who benefits when a distinction is treated as natural, technical, or inevitable. The human-machine boundary is not simply discovered. It is designed, funded, narrated, enforced, and made operational through systems that decide which bodies, records, categories, and forms of work count.

That is why boundary work now has legal and safety consequences. Calling a system a companion, high-risk tool, biometric categorization system, medical device, educational product, workplace assistant, or general-purpose model can trigger different duties. The category is not mere language. It determines who must document the system, inform affected people, restrict use, preserve evidence, provide oversight, or accept liability.

This distinction matters now because AI products often sell a clean image of augmentation. A worker with a copilot is more productive. A student with a tutor is more supported. A patient with a chatbot is more cared for. A citizen with an automated portal is more served. Haraway's cyborg lens asks what hybrid is actually being built. Which capacities are extended? Which dependencies are deepened? Which forms of labor disappear from view? Which institution gains the power to define the person through data?

The cyborg is not the arrival of machines into a pure human world. It is the exposure of the fact that the pure human world was already a story. People think through tools, classifications, institutions, medicines, screens, models, supply chains, prostheses, passwords, and records. AI intensifies the condition by making language, memory, delegation, and judgment more explicitly machine-mediated.

The Informatics of Domination

One of the book's durable concepts is the "informatics of domination," Haraway's name for a world in which communication, biology, labor, markets, bodies, war, and social relations are reorganized through information systems. That phrase lands cleanly in the age of foundation models, enterprise agents, sensor networks, recommender systems, automated hiring, biometric identity, and data-driven public administration.

The value of the concept is that it does not treat AI as merely a smarter artifact. It treats computation as a social order. A model is trained on classified traces of the world. The model produces outputs that reshape work, attention, prices, risk, intimacy, and institutional memory. The changed world produces new traces. The system is not just a tool inside society; it becomes one way society describes, sorts, and acts on itself.

This is why Simians, Cyborgs, and Women belongs near How We Became Posthuman, Atlas of AI, Sorting Things Out, and Data Feminism. Each book rejects the fantasy that information floats free from bodies, infrastructure, categories, and power. Haraway adds the older feminist and technoscientific grammar: the body is not outside the system, and the system is never only technical.

The informatics frame also explains why the same AI system can be liberating in one setting and coercive in another. A speech-to-text tool can increase access for one user and intensify workplace monitoring for another. A model that helps a researcher search a corpus can also normalize scraping, hidden annotation labor, and opaque ranking. Haraway's question is not whether information is good or bad. It is how information becomes an apparatus for making some lives more governable than others.

Situated Knowledge

Haraway's "situated knowledges" argument is essential for evaluating AI answer engines and model-mediated expertise. The problem is not only that systems may be biased. The deeper problem is the pretense of a view from nowhere: a smooth answer that hides where its categories, sources, exclusions, instruments, and incentives came from.

AI systems often present knowledge as portable output. The response arrives without the archive of collection, labeling, moderation, filtering, ranking, evaluation, prompt framing, and product design that made it possible. Haraway's corrective is accountability for vision. Knowing is not less reliable because it is situated. It becomes more accountable when its situation can be named, inspected, and contested.

This is also a warning about benchmark culture. A benchmark is not simply a measurement of intelligence. It is a situated instrument that decides which tasks matter, which forms of competence count, which failures remain invisible, and which communities are forced to live with the deployment. The point is not to reject measurement. It is to stop treating measurement as a substitute for answerability.

Documentation practices such as model cards and system cards matter for the same reason. They do not magically make a system fair. They create places where standpoints, data limits, evaluation choices, intended uses, non-uses, and residual risks can be named. A situated account is not a confession of weakness. It is the minimum condition for responsible reuse.

The same discipline should apply to datasets, prompts, retrieval systems, human review, and deployment settings. A model card that names only the model while leaving out data-enrichment labor, institutional objectives, biometric capture, vendor contracts, or local override paths is not a situated account. It is a partial account that can still let a system appear placeless.

AI Agents and Composite Action

Read in 2026, Haraway is especially useful for thinking about AI agents. Agent discourse likes clean nouns: the model, the user, the tool, the task. Actual action is messier. A contract drafted by an agent may involve a language model, a retrieval system, a prompt template, a permissions layer, a user's role, a vendor policy, a CRM, prior customer records, legal review, and a downstream institution that treats the output as memory.

The actor is composite. So is the responsibility. Haraway's cyborg theory helps resist the lazy split in which either the human is fully responsible because the machine is only a tool, or the machine is treated as a quasi-agent so the organization can hide behind technical autonomy. The more honest question is how the arrangement distributes action, visibility, correction, and blame.

That has practical consequences. A good agent governance process must inspect permissions, provenance, logs, escalation paths, memory defaults, dataset origins, hidden labor, and the interface pressures placed on the user. Human oversight is not enough if the human has already been formatted by the loop, and audit trails are not enough if the record cannot reconstruct who authorized an action.

NIST's 2026 AI Agent Standards Initiative makes this less theoretical. Its stated work includes industry-led standards, interoperable agent protocols, agent authentication, identity infrastructure, and security evaluations for systems capable of autonomous action. In Haraway's terms, the policy problem is how to govern the cyborg arrangement without pretending the agent is either a simple tool or a separate person. Identity, authorization, revocation, and traceability become human-machine boundary controls, not back-office implementation details.

Governance and Safety

As of June 23, 2026, the strongest governance lesson is that the boundary category carries duties. EU AI Act Article 5 prohibits manipulative and exploitative practices, social scoring, certain criminal-risk assessments based solely on profiling, untargeted face-scraping for facial-recognition databases, workplace and education emotion recognition except for medical or safety reasons, biometric categorization that infers protected characteristics, and real-time remote biometric identification for law enforcement in publicly accessible spaces except under defined conditions. Annex III treats permitted biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice, and democratic-process uses as high-risk in specified cases. Article 26 then turns that category into deployer duties: competent human oversight, input-data care, monitoring, logging, worker notice where applicable, and information for people subject to AI-assisted decisions.

Article 14 makes oversight a boundary problem, not a slogan. High-risk systems must be designed so natural persons can understand capacities and limits, monitor operation, correctly interpret output, decide not to use or override it, and interrupt operation. Article 27 adds a further context test for many deployers of high-risk systems: before use, assess the process, affected groups, specific risks of harm, human oversight measures, and complaint or internal-governance arrangements. These are cyborg questions in legal form. They ask whether a composite system leaves a capable human with authority, evidence, and a path to stop harm.

Article 50 adds another boundary duty: providers must inform people when they are interacting directly with AI unless that is obvious in context, and providers or deployers must mark or disclose certain synthetic and manipulated content, deepfakes, emotion-recognition, and biometric-categorization uses. The European Commission's transparency code, published on June 10, 2026, supports those Article 50 obligations, which apply from August 2, 2026, but the code does not replace the law.

Those rules matter for Haraway because biometric, companion, and generative systems are boundary machines. A biometric categorization system turns bodily traces into administrative labels. A synthetic-media system can blur witness, performance, evidence, and fiction. A companion chatbot can invite users, especially minors or isolated people, into a social relation while remaining a commercial product. The FTC's September 11, 2025 inquiry into AI companion chatbots asked seven companies how they test, monitor, disclose, monetize, and mitigate potential harms to children and teens.

The practical governance pattern is to map the whole arrangement before approving the use. Who provides the model? What data and labels shaped it? Which body, voice, face, behavior, record, or relationship is being rendered machine-readable? What permissions can the system exercise? What does the interface pressure the user to accept? Who can inspect logs, challenge outputs, revoke delegation, and repair harm? Where are non-use rules stronger than mitigation?

For high-stakes or intimate systems, the answer should include purpose limitation, least-privilege access, provenance records, user notice, meaningful appeal, incident reporting, age-appropriate design, nonhuman disclosure, and hard refusals for unnecessary sensitive inference. NIST's AI RMF and Generative AI Profile are useful here because they frame risk management across design, development, use, and evaluation rather than treating launch disclosure as the finish line. The point is not to make the cyborg arrangement pure. It is to make its dependencies visible enough that people can contest them.

Boundary Audit

A practical reading of Haraway turns the cyborg from metaphor into an audit. Start with four tests. Classification: what categories does the system attach to people, bodies, records, or relationships? Embodiment: what bodily trace, workplace action, affective signal, sensor feed, or identity record is captured? Delegation: what can the system do without fresh human assent? Contestability: who can pause, inspect, appeal, correct, delete, refuse, or exit?

These tests keep "human-centered" language from becoming decoration. If a system advertises human oversight, name the human's actual position: data subject, annotator, operator, supervisor, rubber stamp, appeal officer, liability shield, or decision-maker with authority. If a product is called AI-assisted, name the assistance: ranking, evidence generation, default recommendation, triage, task allocation, biometric sorting, synthetic intimacy, or autonomous tool use. Each position carries different duties.

For consequential systems, keep a boundary register alongside the model card. It should name the system version, data sources, category definitions, embodied signals, user roles, delegated permissions, human review points, non-use rules, notice language, override path, appeal path, log retention, vendor changes, incident links, and decommissioning trigger. That record turns a metaphor into evidence an auditor, affected person, or regulator can inspect.

The same audit should ask whether the boundary is symmetrical. Can the person see what the machine inferred? Can they withhold the body, voice, image, clickstream, or memory record? Can they get service without the hybrid system? Can they reach a human who has authority to change the result? A boundary that only the institution can move is not augmentation. It is administration by interface.

This is where the review connects to Ghost Work, algorithmic impact assessments, algorithmic recourse, automation bias, and AI incident reporting. The human-machine boundary is not only a philosophical question. It decides which people become data, which people become hidden labor, which humans are reduced to rubber stamps, and which people receive a path back into accountable process.

Recursive Reality

The book also sharpens the site's recurring concern with recursive reality: systems whose descriptions of the world become inputs to the world they later describe. Haraway's essays show that "nature" itself can be produced through stories, instruments, institutions, and practices that then appear to reveal what nature always was.

AI intensifies that loop. A model classifies a worker as productive or risky. The classification changes assignments, monitoring, and opportunity. The changed behavior becomes new performance data. A companion system names a user's mood. The user learns to narrate the self through the system's categories. A search or answer engine summarizes a controversy. Later readers cite the summary, and the summary becomes part of the record that future systems retrieve.

Haraway's contribution is to make this loop embodied. Recursive reality is not only a media problem or a semiotic problem. It happens to people with bodies, jobs, illnesses, histories, genders, races, disabilities, dependencies, and institutional files. The representation does not hover above life. It enters life and helps decide which forms of life are legible.

Where the Book Needs Friction

The book is difficult. Haraway's prose is dense, allusive, and deliberately resistant to simple summary. Readers looking for direct AI policy recommendations will have to translate from feminist theory, primatology, immunology, socialist-feminist politics, and science studies into today's platform systems. That work is demanding, but it is also the point: easy technical language often hides the politics Haraway wants visible.

The book can also be overused as a permission slip for boundary talk that never reaches institutional design. Saying "we are all cyborgs" is not enough. The hard questions are more specific: who owns the infrastructure of hybrid life, who is mined for data, who performs repair, who can refuse augmentation, who audits the classification, who has bodily risk, and who is punished when the hybrid system fails?

Haraway's own language also belongs to its historical moment. The AI era adds cloud platforms, surveillance advertising, large-scale data extraction, biometric identity, generative media, model labor, and agentic tool use at a scale the 1980s essays could not map directly. The review task is not to make the book predict the present. It is to use the book to ask better questions of the present.

There is also a limit in applying cyborg theory to AI companions and agents. A system may perform warmth, initiative, memory, and concern without possessing consciousness or moral standing. Governance should protect people from manipulation, dependency, deception, and institutional abandonment without granting the system a metaphysical status it has not earned.

Source Discipline

This review separates four kinds of claims. Book claims come from the publisher record, library metadata, and Haraway's academic profiles. Historical interpretation comes from the essays and scholarly reception. Current legal and safety claims come from official regulators and standards bodies. Site interpretation is marked as interpretation, not as a claim that Haraway endorsed any contemporary AI product or policy.

That separation matters because cyborg theory is easy to inflate. A phrase from a 1985 essay should not be used as proof that modern AI is a mind, a prophet, or a social inevitability. The stronger reading is narrower and more useful: Haraway gives a disciplined way to examine how boundary categories become infrastructure, and how infrastructure then acts back on bodies, labor, memory, and public truth.

Source discipline also means resisting analogy creep. A companion chatbot, biometric gate, workplace copilot, prosthetic device, medical triage tool, and data-labeling pipeline are all human-machine arrangements, but they are not governed by the same risks. The page uses cyborg theory to ask sharper questions about role, evidence, control, and contestability; it does not use the theory to collapse every hybrid system into one category.

Finally, every current claim on this page should stay date- and version-specific. The 1991 book supplies a conceptual grammar, not empirical proof about a present product. Legal duties come from the current instrument and article. Technical duties come from the specific system, deployment context, standard, test, or audit record. Without that separation, boundary analysis becomes another source of myth.

What This Changes

The practical lesson is to stop treating AI as an external machine added to an otherwise human institution. AI systems produce cyborg arrangements: composite actors made of people, models, tools, categories, logs, incentives, permissions, bodies, and stories.

A useful audit starts with three boundary questions. First, what human capacity is being extended, substituted, scored, or made legible? Second, which machine output changes another person's options, standing, risk, or self-understanding? Third, who has the power to contest that change after the system has already acted? Those questions turn theory into procurement, interface design, recordkeeping, appeal, and non-use decisions.

For product teams, that means documenting the whole arrangement, not just the model card. For institutions, it means assigning responsibility where action is actually distributed. For auditors, it means checking whose standpoint is hidden inside the system's answer. For workers, patients, students, and citizens, it means asking whether a machine-mediated role extends agency or makes the person easier to administer.

Simians, Cyborgs, and Women remains valuable because it refuses both machine worship and human purity. It shows that the boundary between human and machine is a site of politics, not a final metaphysical border. In the AI era, that is a practical governance lesson: map the hybrid, name the obligations, preserve contestability, and do not let fluent systems hide the bodies and institutions they depend on.

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