Blog · Review Essay · Last reviewed June 25, 2026

New Laws of Robotics and the Human Expertise Rule

Frank Pasquale's New Laws of Robotics is a legal and political answer to automation fatalism: AI should not be treated as destiny, but as institutional design subject to democratic choice.

For this review, the human expertise rule means that AI should strengthen accountable human skill where rights, care, safety, education, work, or public authority are at stake. A system fails that rule when it converts professionals into liability buffers, hides managerial choices behind automation, or makes affected people contest a machine-mediated decision without access to the institution that used it.

Human expertise is not a sentimental "human in the loop" slogan. It is a capacity bundle: domain skill, context, time, independence, records, workload protection, authority to refuse, and a route for the affected person to be heard.

The Book

New Laws of Robotics: Defending Human Expertise in the Age of AI was published by Harvard University Press in 2020. Google Books lists Frank Pasquale as author, Harvard University Press as publisher, October 27, 2020 as publication date, and ISBN-10 0674975227 and ISBN-13 9780674975224 for the edition. Amazon uses the same ISBN-10 as its product identifier.

The book follows Pasquale's earlier The Black Box Society, but the emphasis changes. The earlier book made opacity the problem. This one asks what positive rules should govern robotics and AI when they enter medicine, education, policing, finance, social media, military systems, and professional work. Brooklyn Law School summarizes Pasquale's argument as a case that AI and robotics should complement, rather than replace, human labor, and that public institutions can help secure that outcome.

Pasquale's laws are not robot commandments. They are duties for people, firms, professions, and states: complement professionals rather than replace them, avoid counterfeit humanity, resist zero-sum arms races, and keep the identity of creators, controllers, and owners legible. That is why the book still matters after the generative-AI turn. It asks who is accountable before the system is treated as autonomous.

Current Context

As of June 25, 2026, Pasquale's argument is less about humanoid robots than about delegated authority in ordinary institutions. AI systems now appear as clinical documentation tools, tutoring assistants, hiring and workforce systems, public-service chatbots, legal research aids, coding agents, logistics optimizers, customer-service systems, and robots or embodied systems in warehouses, labs, hospitals, factories, and homes. The question is not whether every such system replaces a professional outright. It is whether the surrounding institution preserves expertise, evidence, recourse, and responsibility when automation changes the work.

Current governance language has begun to name the same problem, though dates and legal force vary by jurisdiction and system type. The EU AI Act's official text requires high-risk systems to be designed for effective human oversight under Article 14; Article 26 requires deployers to assign oversight to natural persons with the necessary competence, training, authority, and support, and includes a worker-notice duty before certain high-risk systems are used at work. The AI Act Service Desk timeline says the Act applies progressively and notes Digital Omnibus caveats around high-risk support tools. The practical lesson is stable even where transition dates move: oversight must be staffed, trained, resourced, and empowered before deployment.

NIST's AI Risk Management Framework gives organizations a lifecycle vocabulary of Govern, Map, Measure, and Manage, while its Generative AI Profile flags overreliance, automation bias, and human-AI configuration risks. ISO/IEC 42001 turns AI into a management-system problem: policies, responsibilities, monitoring, review, and continual improvement. The dated 2024 U.S. Department of Labor AI best-practices release is no longer a statement of current administration policy, but it remains a useful official record of worker-centered controls: worker input, meaningful human oversight for significant employment decisions, rights protection, training, transparency, and worker-data safeguards.

Agentic AI makes Pasquale's rule harder to fake. NIST's 2026 AI Agent Standards Initiative treats agents as systems needing standards work around identity, authentication, protocols, interoperability, security, and evaluation. A professional is not defended by adding a chatbot to the workflow. Expertise is defended when the system's authority, evidence, logs, tools, escalation paths, and failure ownership are clear enough for humans to govern, and when the institution can prove that the human role changed outcomes rather than merely blessed them.

The Expertise Rule

Pasquale's most useful move is to defend expertise without romanticizing professions. Nurses, teachers, doctors, designers, and other skilled workers are not sacred because they are human in the abstract. They matter because they hold situated judgment, ethical duties, interpersonal knowledge, and lines of accountability that cannot be reduced to throughput. A society that treats them as expensive obstacles to automation loses more than jobs. It loses institutions that know how to care, explain, contest, and revise.

Expertise has three layers. The first is craft: the trained ability to notice what a metric, prompt, robot, or dashboard misses. The second is duty: professional, legal, ethical, or public obligations that bind the person to patients, students, clients, workers, citizens, or the public. The third is institution: time, staffing, records, review, appeal, liability, and independence. AI can support the first layer while quietly weakening the other two. That is the failure mode this book helps name.

This is a direct challenge to the managerial story of AI adoption. The dominant pitch says: if a task can be automated, the worker is inefficient. Pasquale reverses the question. If an AI system can assist a professional, who designs the assistance, who owns the data, who sets the purpose, and who answers when the system fails? The problem is not automation as such. The problem is replacement framed as inevitability.

The rule is operational. A worker or professional must have the time, training, records, and institutional authority to challenge the tool. A patient, student, applicant, tenant, defendant, customer, or worker must have a path to reach a responsible human who can correct the record. Otherwise "human in the loop" becomes the ritual by which a machine-shaped decision borrows human legitimacy.

A useful test is the disagreement test. If the professional says the tool is wrong, can they see the source evidence, explain the disagreement, change the outcome, protect the affected person, and leave a record that improves the system? If the answer is no, the institution has not preserved expertise. It has preserved a signature line.

Law Before Autonomy

The title deliberately answers Asimov, but Pasquale's laws are aimed at institutions, not machines. That matters. Telling a machine to be ethical is a fantasy if the surrounding business model rewards surveillance, deskilling, deception, or arms-race behavior. The relevant law is not a command whispered into the robot. It is procurement, liability, labor law, professional regulation, transparency duties, antitrust, safety standards, and public investment.

That makes the book useful because it turns "AI alignment" back toward institutional alignment. An automated system can be technically impressive while misaligned with the public. A tutoring system can optimize engagement while weakening teachers. A clinical tool can increase throughput while narrowing care. A policing system can sharpen suspicion while degrading due process. Pasquale gives readers a vocabulary for asking whether the technology strengthens human institutions or drains them of judgment.

The site-wide governance lesson is concrete: a system that changes work also changes the record of reality inside an institution. It may decide what counts as a case, risk, diagnosis, exception, lesson, ticket, incident, or completed task. If those records are machine-shaped but legally human-owned, then human expertise must include the power to inspect and correct the record, not merely to accept an interface summary.

The Agent Reading

Read in 2026, the book is a guide to AI agents. Agents do not only predict; they take delegated actions. They draft, route, schedule, search, purchase, recommend, escalate, and sometimes execute tool calls across institutional systems. That makes Pasquale's emphasis on expertise sharper. If agents are inserted into professional workflows, they should expand accountable capacity, not become invisible managers that deskill workers while shifting blame back onto them.

The agent version of the expertise rule asks what the system can do, not only what it can say. Can it read records, write to systems, call tools, approve work, change schedules, route cases, make purchases, or update files? If so, governance has to separate read, write, send, publish, spend, delete, and permission-change authority. Logs, tool provenance, scoped credentials, rollback, and escalation are not technical extras; they are the conditions under which human expertise remains more than supervision after the fact.

A useful agent deployment needs an expertise handoff file. It should name the professional role being supported, the task boundary, permitted tools, forbidden actions, data touched, credential scope, approval gates, logging rule, rollback plan, escalation path, affected-person notice or appeal route, and the person or office that can pause the system. Without that file, the agent may become a new manager while the professional remains the visible signature.

The handoff file should also separate assistance from assessment. A system that helps a clinician draft notes, a teacher prepare feedback, or an engineer search documentation should not quietly become a productivity score, quality rating, or discipline record without a new review, worker notice where required, and an appeal path. The same trace that supports expertise can become surveillance if the institution changes its purpose after deployment.

NIST's AI RMF and Generative AI Profile support that lifecycle view, and the EU AI Act's official text gives human oversight, logging, documentation, risk management, and deployer obligations legal form in high-risk contexts. Those sources do not prove any given product is safe. They show the evidentiary burden Pasquale's argument implies: responsibility must attach to people and institutions across the system, not be displaced onto the machine's output.

Governance and Safety

The practical governance question is whether automation changes the location of responsibility. If an AI system recommends treatment, drafts a lesson, scores a worker, routes a benefits case, answers a resident, controls a robot arm, or operates through an agent connector, someone must be able to identify the system owner, use case, data sources, version, evaluation evidence, human role, appeal path, and stop condition.

Pasquale's frame turns human oversight into a capacity bundle. Oversight requires domain competence, time, uncertainty visibility, authority to refuse, protection from retaliation, access to logs and source material, and a way to repair the downstream record. A person who merely signs off on a machine-shaped output is not exercising expertise; they are absorbing liability.

The minimum artifact is an expertise-preservation plan. It should define the task; name the professional discretion that must remain; state which outputs are advice and which become action; document data provenance; require evaluation against the real workflow; set boundaries on secondary data use; preserve source and action logs; protect appeal and recourse; consult affected workers; monitor after deployment; and preserve a non-automated path where rights, safety, or care require one.

That plan should be tested as a workflow, not filed as a policy. Reviewers should be observed under real workload; override rates, appeal outcomes, near misses, and ignored warnings should be tracked; and the organization should ask whether the tool has shifted judgment from the professional to the vendor, dashboard, or productivity metric. Automation bias is not only a cognitive error. It is often an organizational design choice.

For embodied robotics, the safety case has an added physical layer: hazard analysis, operational design domain, sensor limits, fail-safe behavior, maintenance, update controls, human-robot interaction boundaries, and incident reporting. ISO's robotics materials list the 2025 ISO 10218 industrial-robot safety standards and the ISO/TS 15066 collaborative-robot specification as relevant safety standards. Those standards do not settle labor, care, or public-authority questions, but they make one point clear: a robotic system can be useful and still require stronger governance because its errors leave the screen and enter shared space.

The procurement checklist should therefore produce records, not only promises: model or robot description, vendor and owner, task boundary, risk tier, worker consultation, training plan, data categories, evaluation evidence, limitations, human authority, override and stop controls, logs, incident triggers, change notices, exit path, and post-deployment review date. If a buyer cannot get those terms, it is not purchasing reliable expertise support. It is purchasing dependency.

Where the Book Needs Care

The book's optimism about professional partnership is useful, but it can understate how weak many professions have become under cost pressure, platformization, staffing shortages, and managerial control. Complementarity can become a slogan if workers have no power to refuse bad tools, audit performance claims, protect clients, or slow deployment. A nurse "augmented" by software that intensifies surveillance and staffing ratios has not been defended.

The book also depends on law doing real work. That is the right arena, but not an easy one. Powerful firms can lobby, litigate, structure contracts, and define standards in their own interest. Pasquale is strongest when he treats governance as a political struggle, not a checklist. The new laws of robotics will not enforce themselves; they require institutions capable of saying no to profitable replacement.

A further limit is that "expertise" can be used defensively by professions that already exclude, overbill, or resist accountability. The right answer is not expert immunity. It is accountable expertise: public standards, patient and student rights, worker voice, auditability, and recourse for those harmed by both machines and professionals.

The AI-era correction is to watch for ceremonial expertise. A profession can be invoked to legitimize a system while the actual authority moves into vendor dashboards, staffing formulas, productivity metrics, or agent permissions. The test is not whether a professional appears somewhere in the workflow. The test is whether that professional can see the evidence, challenge the tool, protect the affected person, and change the outcome without retaliation.

What This Changes

New Laws of Robotics gives this site a practical test for AI deployment. Does the system strengthen human expertise, or make expertise cheaper to ignore? Does it identify its creators, controllers, and owners, or hide behind the interface? Does it reduce zero-sum arms races, or intensify them? Does it make care, judgment, and accountability more available, or merely make supervision more automated?

The strongest test is after deployment. Are workers more able to exercise judgment? Are clients and patients more able to understand and contest decisions? Are errors easier to discover? Are logs available? Are staffing and training protected? Or has the system made professional care look present while moving real authority into software and management metrics?

The book's lasting value is its refusal of inevitability. AI systems are not independent historical forces. They are designed, bought, regulated, marketed, resisted, and maintained. The question is not whether machines will replace humans. The question is which institutions will decide what replacement means, whether expertise is preserved as authority rather than branding, and whether the people affected will have any power in that decision.

The recurring theme is mediated reality with public memory. When a model, robot, or agent changes what the institution sees, the institution must preserve enough evidence for humans to answer back: inventories, logs, explanations, appeals, incident reports, procurement terms, and exit routes. Expertise survives automation only when the record remains governable.

Source Discipline

This review separates book facts, author and institutional context, legal duties, standards guidance, and dated policy statements. Harvard University Press, Google Books, Amazon, and Brooklyn Law School establish the book and author context. EUR-Lex is the primary source for EU AI Act text; AI Act Service Desk and Commission pages are implementation guidance and timeline context. NIST supplies voluntary risk-management and agent-standards vocabulary. ISO pages identify management-system and robotics safety standards. The Department of Labor release is cited as a dated 2024 official statement; its page warns that some older releases may not reflect current policies after January 20, 2025.

For deployment claims, the primary evidence is local: the actual system boundary, job or care workflow, reviewer training, staffing assumptions, logs, appeals, incident records, vendor terms, and proof that humans can reject or repair machine-shaped outputs. A policy that says "human oversight" is not evidence that oversight works.

Do not convert "human oversight," "human-centered," "worker input," or "professional augmentation" into magic phrases. A legal or standards document can establish a requirement, but the actual deployment still needs evidence about workflow, staffing, logs, data, appeals, incidents, training, and whether the human reviewer can change the result. Current book, legal, standards, regulator, and policy claims were rechecked on June 25, 2026.

This page makes no claim that any AI system is conscious, divine, or AGI. It treats AI systems as engineered tools and institutional processes that must remain answerable to human expertise.

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