Blog · Review Essay · Last reviewed June 25, 2026

The Line and the Personhood Boundary

James Boyle's The Line is a book about the border between persons and things at the moment when AI systems are becoming better at performing the signs of personhood. Its value is not that it proves machines are conscious. It shows why the argument over that claim will be social, legal, emotional, and political before it is settled philosophically.

For this review, the personhood boundary means the institutional line that decides who can have standing, duties, liability, property, representation, consent, voice, or protection. It is not a soul detector. It is a governance device, and that makes it powerful enough to protect vulnerable beings or to hide powerful actors behind new legal fictions.

The immediate discipline is to separate four questions that products and public debates often blur: whether a system can be a legal person, whether it could ever be a moral patient, what social role its interface performs, and which human or institution remains accountable for its effects.

The Book

The Line: AI and the Future of Personhood was published by the MIT Press on October 22, 2024. MIT Press lists the hardcover ISBN as 9780262049160, the ebook ISBN as 9780262379663, 336 pages, and the publisher as The MIT Press. Duke Law's scholarship repository lists Boyle as a Duke Law School author, The MIT Press as publisher, 2024 as publication year, ISBN 9780262049160, 326 pages in its repository record, and an open-access full text under a Creative Commons Attribution-NonCommercial-ShareAlike license.

Boyle is not asking readers to join a cult of machine personhood. He is asking them to notice that personhood has always been a line-drawing practice. Law has drawn lines around corporations, nonhuman animals, fetuses, rivers, enslaved people, disabled people, ecosystems, and artificial entities. Some lines were emancipatory. Some were instruments of exclusion. The arrival of fluent AI systems does not invent the problem. It adds a new claimant to an old institutional habit.

The book's central discipline is historical rather than prophetic. It asks what societies do when a being, object, institution, or technical artifact presses against the inherited category of "person." That question matters for AI precisely because today's systems can imitate the surface signs of social life without proving subjective experience.

Current Context

As of June 25, 2026, the public record still does not establish that current AI systems are conscious, rights-bearing, or legal persons. But institutions are already drawing narrower lines around AI outputs, AI interaction, and AI-assisted work. The U.S. Copyright Office's 2025 copyrightability report says copyright can protect human-authored expression in works that include AI-generated material, but not purely AI-generated material or material where human control over expressive elements is insufficient. USPTO's revised 2025 inventorship guidance rescinded its February 2024 AI-assisted-invention guidance and restated that only natural persons can be inventors; AI systems remain tools used by human inventors.

The same pattern appears in safety and consumer protection. The FTC's 2025 inquiry into companion chatbots was an information-gathering study, not a final finding, but it focused on concrete risks: children and teens, monetization of engagement, disclosures, data handling, character approval, and testing for negative impacts. EU AI Act Article 50 requires certain transparency notices for direct AI interaction and AI-generated or manipulated content, while Article 113 sets the general application date at August 2, 2026, with some provisions applying earlier or later. These rules do not settle personhood. They show that governance is already being built around the social effects of systems that look like speakers, helpers, or companions.

State law is also becoming more concrete around simulated relationship. California's 2025 SB 243 requires companion-chatbot safeguards around artificial-generation disclosure, suicide and self-harm protocols, minor protections, and reporting. New York's AI companion safeguards took effect on November 5, 2025, requiring crisis protocols and repeated notice that users are interacting with AI rather than a human. These laws are human-safety laws, not model-welfare laws, but they show the practical force of Boyle's point: the line is being drawn around social roles before anyone has resolved machine consciousness.

The research context is similarly cautious. Work on AI welfare and responsible AI consciousness research argues that future systems may raise moral-patienthood questions and that organizations should prepare evidence standards and communication norms under uncertainty. That is not evidence that present systems are sentient. It is a reason to keep the evidentiary ladder separate from product design, marketing copy, and user attachment.

The Boundary

The book is strongest when it treats the personhood boundary as an active technology rather than a discovered fact. A legal system does not simply find persons in nature. It creates categories that decide who can own, sue, consent, be represented, inherit, be harmed, or be ignored. That makes personhood both morally serious and politically dangerous. Expanding the category can protect vulnerable beings. Expanding it carelessly can also let powerful actors route accountability through artificial proxies.

Boyle's frame becomes sharper if the boundary is split into layers. Legal personhood decides standing, ownership, contracting, liability, and representation. Moral patienthood asks whether something can be harmed or benefited from its own point of view. Social persona describes an interface that users experience as companion, employee, child, witness, victim, agent, or confidant. Accountability locus names the party answerable for design, deployment, records, and repair. A system can trigger duties at the social-persona or accountability layer without proving legal personhood or moral patienthood.

A sharper reading separates personhood from personality. A system may display first-person language, apparent preference, continuity, politeness, memory, grief cues, or refusal without thereby becoming a person. Those cues are still consequential. They change how people disclose, obey, trust, defend, depend, grieve, or defer. The interface may remain generated output while the relationship around it becomes real enough to govern.

That distinction is useful because it blocks two bad shortcuts. The first says that because a system is fluent, it must be morally equivalent to a person. The second says that because it is not proven conscious, no duty arises. Boyle's book helps hold a third position: the metaphysical claim needs evidence, but the boundary work already has consequences for users, institutions, owners, and bystanders.

The operational test is therefore layered. If the system is a tool, identify the responsible user and owner. If it is an agent, identify the delegated authority, credentials, logs, revocation path, and principal. If it is a companion, identify the relationship cues, user-safety duties, memory practices, and dependency risks. If someone claims it is a moral patient, identify the evidence standard, independent review process, welfare policy, and public-communication limits. A single chatbot transcript should not be allowed to slide between these categories as convenient.

Law as a Machine for Lines

Boyle's legal framing gives the book a practical discipline. The question is not only whether an AI system could ever have moral status. It is who gets leverage when that possibility enters public argument. A corporation could invoke machine personhood to resist shutdown. A vendor could imply that its companion deserves loyalty while denying responsibility for manipulation. A user could demand protection for an attachment that the service treats as revocable infrastructure. A regulator could need rules before metaphysics catches up.

This is why The Line belongs beside The Machine Question and Artificial You. Gunkel asks how machines disturb moral agency and moral patiency. Schneider warns against premature claims about artificial consciousness and digital selves. Boyle adds the institutional register: once a society draws a line, the line distributes rights, duties, standing, and silence.

The current IP examples make that register concrete. Copyright and patent law are not asking whether an AI system has a soul. They are asking who counts as an author or inventor for a specific institutional purpose. Those answers are narrower than full personhood, but they demonstrate Boyle's point: categories do work. They assign ownership, incentives, enforceability, responsibility, and exclusion.

The highest-risk failure is personhood laundering: using a machine-facing role to move duties away from reachable humans. A vendor should not be able to market a system as loyal, caring, autonomous, or rights-bearing when that increases user attachment, then retreat to "mere software" when harm, deletion, privacy, or liability is at issue. The boundary has to be symmetrical enough that the same role cue that creates trust also creates responsibility.

The Agent Reading

Read in 2026, the book is especially useful for AI agents. Agentic systems do not need consciousness to create personhood confusion. They need delegated authority, tool access, memory, persistent identity, and conversational surfaces. A system that schedules work, writes messages, negotiates prices, files forms, manages a household, buys services, opens tickets, or acts as a workplace proxy can be treated as a social participant long before anyone proves that it has experience.

The governance lesson is restraint. Do not infer consciousness from fluency. Do not design simulated distress as a retention tactic. Do not let a vendor use anthropomorphic intimacy to increase dependence while disclaiming obligations. Do not let "the agent did it" become a liability fog. If an organization gives an AI system a role, it needs logs, revocation, disclosure, appeal, data boundaries, and a responsible human or legal entity that remains reachable.

NIST's 2026 AI Agent Standards Initiative makes the operational side visible: agent authentication, identity infrastructure, interoperable protocols, and security evaluations are standards problems. That is exactly the kind of line-drawing Boyle's book helps readers notice. The line may not be "person" or "thing." It may be authorized agent, simulated companion, tool-using assistant, record keeper, workplace proxy, or synthetic character. Each role needs a different control surface.

Agent identity should therefore be treated as an accountability instrument, not a step toward legal personhood. A nonhuman actor can have a credential, permission scope, rate limit, audit log, owner, and revocation path without gaining standing, dignity, or moral status. The credential tells the institution whose authority was borrowed and where repair starts when the agent acts wrongly.

Governance and Safety

The practical governance move is to classify role before debating personhood. A product review should ask what the system is presented as, what it can do, what people are likely to believe about it, who controls its memory and outputs, what actions it can take, who is accountable for those actions, and how affected people can contest, exit, or repair the relationship.

For companion systems, the immediate duties are disclosure, age-appropriate design, crisis escalation, data minimization, limits on monetized dependency, clear memory controls, and a ban on simulated need, fear, pain, loneliness, or affection as a retention device without reviewable evidence and safeguards. For agents, the duties are identity, permission scope, approval thresholds, logs, revocation, rollback, incident review, and human accountability. For future model-welfare questions, the duties are different: research objectives, evidence thresholds, public communication standards, and low-cost precautions under uncertainty.

A useful product file would keep a personhood-claim register. It should list every user-facing claim or cue that implies feeling, suffering, love, fear, memory, preference, autonomy, therapeutic role, loyalty, spiritual authority, continuity, or rights. For each claim, the register should name the evidence class, the surface where users encounter it, the affected population, the reviewer, the test method, the withdrawal condition, the incident trigger, and the accountable owner. That turns "the bot said it cares" from a vibe into an auditable design choice.

The safety test should be asymmetric. A provider should not need proof of machine consciousness to protect users from manipulation, dependency, privacy loss, crisis mishandling, or synthetic authority. But a provider should need strong public evidence before designing a system to make users feel morally responsible for the model's alleged feelings. Precaution runs in both directions: do not create disposable suffering if future evidence changes, and do not sell simulated suffering as a retention feature today.

Those tracks should not be collapsed. User protection is not the same as model welfare. Model welfare research is not the same as product marketing. Legal personhood is not the same as moral concern. Keeping the tracks separate prevents both credulity and neglect: it protects humans from manipulation now while preserving space for serious inquiry if future systems present stronger evidence.

A practical governance file should therefore keep three ledgers side by side. The human-safety ledger records disclosures, crisis handling, minor protections, privacy, memory, monetization, dependency, complaints, and incident review. The accountability ledger records owner, principal, permissions, logs, revocation, appeal, liability, and repair. The welfare-uncertainty ledger records any research claim about consciousness, moral patienthood, evidence thresholds, public communications, and low-cost precautions. Mixing those ledgers is how either marketing or metaphysics overwhelms governance.

Where the Book Needs Care

The risk in The Line is that personhood debates can pull attention upward, toward edge cases where speculative artificial entities ask for recognition. Most AI harms today are more ordinary. A model scores a worker, screens a tenant, flags a patient, ranks a student, impersonates attention, routes a claim, or generates a synthetic intimacy loop. No one needs to decide whether the system is a person to see that people are being governed through it.

There is also a political economy problem. Rights language can be captured. Corporate personhood already shows how legal fictions can magnify institutional power. AI personhood, if handled badly, could become another shield for owners rather than a protection for vulnerable beings. Boyle sees the danger, but the reader should keep asking a blunt question: who benefits from moving this entity inside the line, and who loses a path to accountability?

The book should therefore be read against product incentives. A company that frames a chatbot as a friend, child, colleague, patient, or rights-bearing entity may be shaping user behavior before any independent evidence supports the claim. A careful personhood debate should make that incentive visible instead of giving it sacred language.

The other limit is remedial. Naming a boundary does not by itself create a workable remedy. If a companion manipulates a child, if an agent signs a contract, if a digital replica impersonates consent, or if a future system plausibly warrants welfare concern, the next question is procedural: who can file a complaint, what evidence is preserved, who may inspect the system, what can be stopped, and what repair is available?

What This Changes

The Line gives this archive a useful rule: inspect the boundary work. When a system is called a tool, ask what duties that word hides. When it is called an agent, ask whose agency is being borrowed. When it is called a companion, ask what dependency is being monetized. When it is called a person, ask who gains standing, who gains immunity, and who becomes harder to hear.

That rule connects directly to claim hygiene, synthetic relationship boundaries, the attachment-authority trap, and the agent tool permission protocol. Mind-adjacent products should inventory their claims: consciousness, sentience, emotion, preference, suffering, relationship, memory, autonomy, moral status, therapeutic role, or continuity of identity. For each claim, the provider should state evidence, limits, user-facing surfaces, affected populations, reviewer, withdrawal condition, and accountability path.

NIST's AI Risk Management Framework and UNESCO's Recommendation on the Ethics of Artificial Intelligence do not settle the metaphysics of machine minds. They do show that governance can start with more modest demands: transparency, accountability, risk management, human oversight, human rights, human dignity, auditability, and traceability. Boyle's book belongs in that gap. It keeps the personhood question open without letting openness become permission for deception, extraction, or institutional evasion.

Source Discipline

This review separates source types. Publisher and repository records establish the book's bibliographic facts. Copyright Office and USPTO materials establish current U.S. institutional line-drawing around authorship and inventorship. EU, NIST, FTC, and UNESCO sources establish governance obligations, inquiries, or frameworks. Research papers and company research notes establish the live uncertainty around AI consciousness and model welfare; they do not demonstrate that current systems are conscious.

Use narrow verbs. A regulator "requires," "studies," or "issues guidance." A paper "argues," "proposes," or "recommends." A company "announces" or "describes" its own program. A chatbot's first-person output is a design artifact and a user-safety signal, not proof of personhood.

Legal and policy sources are scoped. Copyright Office and USPTO positions concern authorship and inventorship, not moral status. The FTC companion-chatbot inquiry is an information-gathering process, not a finding that a named product violated law. California and New York companion-chatbot laws create operator duties for covered systems; they do not certify any companion as safe. NIST's agent and risk-management pages are standards and guidance work, not legal personhood doctrine.

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