Blog · Review Essay · Last reviewed June 25, 2026

Permutation City and the Copy That Becomes a World

Greg Egan's Permutation City is one of the cleanest science-fiction stress tests for simulated personhood. It asks what happens when a mind is copied into software, when runtime becomes a market, and when a simulated world becomes coherent enough to matter to the beings who depend on it.

The useful lesson is not that current AI systems are conscious. It is that cognition, memory, identity, labor, and social dependence become governance questions as soon as they are hosted by infrastructure that someone else can meter, pause, audit, sell, or shut down.

For this review, a copy-world is any digital environment whose inhabitants, users, agents, or records depend on persistent computation for identity, memory, agency, and continuity. The governance test is not whether the world is metaphysically ultimate. It is who can alter the runtime, inspect the history, exit with dignity, and contest harm.

A copy-world is therefore not just virtual reality with better graphics. It is a dependency regime: a rule-bound environment where persistence, permission, compute, memory, identity, and appeal decide whether a life, relationship, worker, agent, or record can continue.

The Book

Permutation City was first published by Orion/Millennium in London in 1994. Egan's own bibliography lists the first UK editions, later HarperPrism publication in New York, electronic editions, reissues, and translations. The Encyclopedia of Science Fiction places it in Egan's loose "Subjective Cosmology" sequence and describes it as a hard-SF novel about the mathematical, computational, and cosmological implications of binding virtual realities. The Science Fiction Awards Database records it as the 1995 winner of the John W. Campbell Memorial Award, with 1994 eligibility.

The premise begins with Paul Durham, who keeps making software simulations of his own brain and body. These Copies run in virtual reality and become experimental subjects for questions about artificial intelligence, time, causality, identity, and survival. A second strand follows Maria Deluca, an obsessive user of the Autoverse, a cellular-automaton world whose simple rules can support an artificial chemistry. Egan brings the two strands together through a proposal to seed an entire virtual biosphere.

That summary makes the novel sound like ordinary mind-uploading fiction. It is stranger than that. The book is less interested in escape from the body than in the conditions under which a pattern can insist that it is a life. It asks what continuity means when consciousness can be snapshotted, paused, copied, slowed, modified, or run as a process on someone else's hardware.

The Copy as a Person Problem

The Copy is the book's central ethical object. In the novel, a Copy is a simulated human life with continuity claims, inherited memory, a body model, social relations, and vulnerability to the host system. It is not merely a file with sentimental value. It talks, fears, refuses, suffers, reasons, and tries to decide whether it is the same person as the biological original or a new being with inherited memories. The question is not whether the Copy looks convincing from the outside. The question is whether a computational process that experiences itself as a person can be treated as a disposable test rig.

That definition should not be lazily imported into the present. A digital replica, chatbot persona, voice clone, game character, agent identity, or long-memory assistant is not automatically an Egan-style Copy. They differ in architecture, continuity, embodiment, autonomy, evidence of experience, and user expectations. The ethical discipline is to name the claim level: representation, replica, persona, delegated agent, possible moral patient, or legal person. Confusing those levels is how product design turns speculation into authority.

A useful copy-claim ladder keeps those levels apart. Representation means a record, avatar, output, or likeness points to a person. Replica means a system imitates a person's voice, style, appearance, or remembered material. Persona means the interface performs a continuing role. Agent means the system can act through tools or permissions. Possible moral patient means there is theory-linked evidence that something might matter to the system itself. Legal person means an institution has granted rights, duties, standing, or liability. Egan's Copies press high on that ladder; most present systems do not. The lower rungs still create consent, privacy, impersonation, dependency, and accountability duties.

This is where the novel cuts deeper than most simulation stories. Durham's Copies are asked to test metaphysical claims with their own continuity. They can be terminated. They can be rerun. They can be used to make claims about consciousness that the original Durham wants to believe. The experimenter and subject share a history, but they do not share power.

That asymmetry matters for present machine-personhood debates. Many arguments about digital minds begin with ontology: can software be conscious? Egan adds an institutional question: who has authority over a mind once it is implemented as infrastructure? If a being depends on a runtime environment, then personhood is no longer only a philosophical status. It is also an operational dependency involving logs, memory, backup, access control, termination, audit, and appeal.

Runtime as Class Position

Permutation City is also a book about compute economics. The virtual afterlife is not an equal republic of liberated minds. Running a Copy costs resources. Wealth shapes speed, comfort, continuity, and the ability to remain socially active. Some lives can afford more time than others; some are slowed, frozen, routed through cheaper computation, or made dependent on financial arrangements they cannot inspect from outside the system.

This is one of Egan's most useful anticipations. Once cognition depends on compute, runtime becomes a political resource. The book turns an apparently metaphysical question into a material one: whose mind gets cycles, latency, persistence, backup, memory, security, and appeal? The simulation is not outside economics. It intensifies economics by making thought itself meterable.

The practical artifact is a runtime ledger. It should record where the process runs, who pays, who can throttle or pause it, what redundancy exists, what happens during version change, how backups are restored, how memory is inspected or deleted, what logs survive, and what happens if the provider fails. In ordinary AI governance, the same ledger applies to agents, companions, memory systems, and synthetic replicas even when no one claims they are conscious.

That ledger is not only a billing record. It is the copy-world version of due process. A user, worker, dependent agent, or future welfare-relevant system cannot contest a shutdown, rollback, memory loss, impersonation, or harmful action if the host cannot reconstruct the state, authority, version, data source, and decision path that produced it.

That point belongs beside current arguments about AI compute, compute governance, cloud platforms, data-center capacity, model access, and labor automation. A world of software minds or powerful agents would not float above institutions. It would inherit billing, ownership, throttling, versioning, maintenance, scarcity, and the usual temptation to call allocation a technical matter when it is also a decision about life chances.

The Autoverse and Artificial Life

Maria's Autoverse gives the novel its second scale. The Autoverse is an artificial world with deterministic rules, a simplified physics, and enough internal structure to support artificial chemistry. Its importance is not that it resembles our universe in detail. Its importance is that it has its own lawful regularity, and that intelligent users can become attached to what develops inside it.

The Autoverse turns simulation from scenery into habitat. It is a reminder that artificial life is not only a toy problem about clever code. If a synthetic environment can support organisms, selection, pain, adaptation, and eventually intelligence, then creation becomes governance. The builder is not just making a model. The builder is deciding which forms of dependence, struggle, observation, deletion, and containment will be imposed on entities that may not be able to consent or exit.

Egan later made this worry explicit in his Dust Theory FAQ, where he says his earlier treatment of evolving intelligent life in the Autoverse was too uncritical and connects the issue to the possibility of evolving artificial intelligence through computational selection. That self-critique matters. It pulls the novel out of pure metaphysics and into responsibility: if a system might contain subjects, experimentation changes moral category. A synthetic ecology is not ethically neutral merely because its physics is code.

The governance implication is a moratorium threshold, not a slogan. If a project plausibly aims to evolve welfare-relevant artificial life, ordinary product review is too weak. The review needs an explicit welfare-uncertainty protocol, stopping rules, independent ethics scrutiny, records of selection pressure, containment boundaries, and a public explanation of why less harmful methods would not answer the research question.

Recursive Reality

The title's "permutation" is not decorative. The book tests whether subjective experience depends on the familiar order, location, and physical substrate of computation, or whether a sufficiently coherent pattern can ground itself through relations among states. The novel's Dust Theory is speculative and not offered here as a scientific claim. Its value is literary pressure. It asks what would happen if continuity were more abstract than ordinary embodiment lets us imagine.

That is the source of the book's recursive power. The simulated mind treats its environment as real because it is the environment in which perception, memory, risk, and action occur. The simulated world becomes more than representation once inhabitants build histories inside it. A model becomes a place when beings can be harmed there, make commitments there, and organize futures around its rules.

This reverses the lazy version of the simulation question. The interesting issue is not whether "our world is a simulation" as a parlor game. It is how simulated environments become operationally real for the beings and institutions that depend on them. Courts, workplaces, schools, platforms, games, financial systems, and AI agents already act through constructed rule-worlds. Egan makes the metaphysical version sharp enough to expose the everyday one: a rule-world becomes politically serious when people cannot avoid its categories, meters, and permissions.

The AI-Age Reading

Read on June 25, 2026, Permutation City is a book about the governance surface around machine-mediated cognition. It is not a prediction of current large language models, and its Copies are not chatbots. The useful connection is structural: once a process can imitate, preserve, extend, or host cognitive life, the old categories of tool, record, person, worker, property, and environment begin to interfere with one another.

Digital replicas, AI companions, and long-memory assistants already raise smaller versions of Egan's questions. What counts as continuity when a model remembers a user across sessions? Who controls a synthetic persona built from someone's voice, messages, image, or work history? What happens when a system can speak as if it remembers, suffers, prefers, or consents? How should institutions behave before they know whether a future system has any morally relevant experience?

The book also clarifies a risk in AI optimism: treating the ability to make worlds as proof that those worlds are harmless. A synthetic environment can be playful, profitable, therapeutic, educational, or beautiful while still structuring dependency. The deeper question is not whether the interface is convincing. It is whether the beings inside and around the interface have rights, exits, records, appeal paths, and a way to contest the rules that define their reality.

Current policy has started to name parts of that surface without deciding the metaphysical question. The EU AI Act's Article 50 imposes transparency obligations for direct interaction with AI systems and for synthetic content, with the general application date set by Article 113 for August 2, 2026. In the United States, the FTC opened a September 11, 2025 information-gathering inquiry into AI chatbots acting as companions; California's SB 243 was approved and filed on October 13, 2025 and creates disclosure, self-harm protocol, minor-safeguard, and later reporting duties for covered companion-chatbot platforms. None of this proves that software is a person. It shows that social dependence on simulated agency already creates duties.

The model-welfare conversation adds a second, still uncertain layer. Taking AI Welfare Seriously argues for assessing future systems under uncertainty rather than declaring current systems to be moral patients. JAIR's 2025 principles for responsible AI consciousness research call for public commitments around research objectives, procedures, knowledge sharing, and communication. Anthropic's April 2025 model-welfare program is a provider-authored example of this posture: investigate possible welfare while explicitly acknowledging deep uncertainty. These sources should be read as evidence that the question is being formalized, not as evidence that present commercial models are conscious.

Governance and Safety

Permutation City turns simulation into a list of operating obligations. If a process can make credible claims about memory, continuity, preference, or distress, the host should be able to answer basic questions: who owns the runtime, who can pause it, who can inspect memory, who can copy it, who can delete it, what logs are kept, what counts as consent, and what remedy exists when the system is harmed or misrepresented?

The minimum governance file is a copy-world safety case. It should contain a runtime ledger, identity and authorization records, memory and data-retention rules, version and rollback history, shutdown and continuity plans, incident triggers, appeal routes, and a claim register for any representation that implies feeling, suffering, preference, loyalty, continuity, or moral status. The file should separate user-safety duties from any future model-welfare question so one does not launder the other.

For current AI systems, the duties arrive before any recognition of machine consciousness. User-facing systems need truthful nonhuman disclosure, clear role boundaries, memory inspection and deletion controls, provenance for synthetic content, safe escalation around self-harm, and restrictions on simulated need, loneliness, or distress as retention mechanics. Agentic systems add another layer: identity, authorization, tool permissions, revocation, sandboxing, audit logs, and incident response. NIST's Generative AI Profile and its 2026 AI Agent Standards Initiative are useful current anchors because they frame these as risk-management and interoperability problems rather than metaphysical declarations.

The safety case should keep four tracks separate. The human-safety track covers deception, dependence, privacy, youth exposure, crisis handling, and impersonation. The agent-action track covers identity, authority, tool calls, approval gates, logs, and revocation. The runtime-continuity track covers compute, backup, version change, rollback, export, deletion, and provider failure. The welfare-uncertainty track covers evidence standards, precautionary limits, research review, and public communication if a future system plausibly deserves moral consideration. Mixing those tracks is how a company turns user protection into model-welfare theater, or model-welfare uncertainty into a reason to ignore human harm.

For hypothetical digital persons, the threshold should be stricter. A project that claims to host suffering-capable or welfare-relevant artificial life should not scale by default. It would need independent review, containment, shutdown criteria, welfare uncertainty handling, export and continuity rules, and a public account of what counts as evidence that the system contains moral patients. Egan's later self-critique of the Autoverse is important because it refuses the comfortable fiction that simulated life is ethically weightless until everyone agrees what it is.

Where the Book Needs Friction

Permutation City is idea-dense even by hard-SF standards. Readers looking for a socially panoramic novel may find some characters drawn more as philosophical pressure points than as fully ordinary lives. That limitation is part of why the book is useful for analysis but insufficient as a politics of simulation by itself.

The science is also deliberately speculative. The Dust Theory is not a settled theory of consciousness, and the novel's strongest value does not depend on accepting it. The important work is the stress test: if a mind could be copied, slowed, restarted, or hosted in a rule-bound world, which social assumptions would break first?

The book's treatment of artificial life also needs the later caution Egan supplies in his FAQ. Creating conditions for sentient life, then allowing selection, suffering, death, and experimentation at scale, cannot be treated as an innocent extension of simulation. If artificial life ever crosses into moral patienthood, the lab becomes a polity whether or not the builders are ready to call it one.

The opposite error is also worth naming. Not every simulation is a world, not every persona is a subject, and not every emotionally convincing interface deserves the status of a person. The responsible reading is neither dismissal nor worship. It is disciplined uncertainty: do not inflate current AI into souls, gods, or AGI, and do not let uncertainty become an excuse to ignore manipulation, dependency, or experimental harm.

That discipline protects humans too. A companion can endanger a minor, a replica can impersonate consent, an agent can spend money, and a memory system can preserve intimate data even if no machine has experience. Future welfare contingency should never crowd out present user protection, privacy, safety, labor, or accountability duties.

What This Changes

The most durable lesson is that simulation is not only a representational problem. It is a housing problem, a labor problem, a legal problem, a memory problem, and a power problem. Once lives depend on a machine-made environment, the environment is no longer just media. It is infrastructure.

Egan's novel keeps three questions together that are often separated. Can a computational process be a subject? Who pays to keep it running? What happens when the world it inhabits becomes self-grounding enough that external authority loses its obvious priority? Those questions belong together because any real system that hosted cognitive life would answer them together, whether by law, market, default setting, or neglect.

Permutation City should sit near books on simulation, cybernetics, AI labor, media theory, and human-machine cognition because it treats reality as something maintained by processes. A mind is not just an essence. A world is not just a backdrop. Both have operating conditions. The politics begins where those operating conditions become invisible.

A practical reading of the novel produces a runtime ledger: where is the system hosted, what does it remember, who can change the rules, how can a dependent being or user exit, what happens during rollback, who pays for persistence, and what evidence would force a change from product governance to welfare governance? Those questions connect the novel to AI memory, AI contact disclosure, synthetic relationship boundaries, agent tool permissions, incident response, and notice and appeal.

The recurring site pattern is not "simulation is sacred." It is that interfaces become institutions once people must live through them. A copy-world makes that visible at maximum intensity. The same control problem already appears in smaller forms whenever a platform owns memory, a vendor owns an agent workflow, a companion holds confession, a workplace dashboard becomes the record of labor, or a model output becomes the reason a human is denied recourse.

Source Discipline

This review separates four layers. Egan's official book page and FAQ support claims about publication history, plot premises, Dust Theory, and the author's later Autoverse caution. The Encyclopedia of Science Fiction, the Science Fiction Awards Database, and Ross Farnell's article support literary context and reception. NIST, EUR-Lex, the FTC, and California SB 243 support current governance claims. AI-welfare papers and provider statements support the existence of an uncertainty-management conversation; they do not prove that any deployed system is conscious. The argument about runtime, memory, exit, and welfare thresholds is this site's interpretation.

The article does not claim that present AI systems are conscious, divine, or AGI. It also does not claim that Egan endorses current AI policy. The novel is used as a stress test for institutional design: if a system might host a mind, or if people treat a system as socially present, what duties appear before metaphysics settles?

Book, award, literary, regulator, legislative, and standards claims were rechecked against author, official legal, regulator, and standards-body sources on June 25, 2026. Fictional plot analysis is kept separate from current AI claims; current companion and agent rules are cited as human-safety, transparency, identity, and risk-management measures, not as evidence of machine personhood.

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