Blog · Review Essay · Modified August 12, 2026 · Last reviewed August 12, 2026

Reality+ and the Reality of Virtual Worlds

David J. Chalmers's Reality+ argues that virtual worlds can be genuine worlds, virtual objects genuine objects, and virtual lives meaningful lives. Its durable contribution is to break the lazy equation of virtual with fake.

This review adds a necessary distinction: metaphysical reality does not automatically produce legal rights or legitimate rule. A mediated environment crosses into operational reality when people rely on persistent identities, relationships, assets, or records inside it and an operator can alter those dependencies or use them to trigger consequences outside the interface. At that point the question is not whether the world is real. It is which rights, evidence, safety controls, and exits its power requires.

The Book

Reality+: Virtual Worlds and the Problems of Philosophy was published by W. W. Norton in the United States and Allen Lane in the United Kingdom on January 25, 2022. Norton's official record gives the U.S. hardcover as 544 pages with ISBN 9780393635805; its paperback record gives ISBN 9781324050346 and a February 7, 2023 publication date. The purchase links on this page point to that U.S. paperback edition.

Chalmers describes the project as technophilosophy: technology creates new philosophical problems and supplies new cases for old ones, while philosophy clarifies the concepts technology unsettles. His teaching guide organizes the book around knowledge, reality, mind, and value, with additional routes through AI, virtual realism, and the simulation hypothesis. The current Penguin author note identifies him as University Professor of Philosophy and Co-Director of the Center for Mind, Brain, and Consciousness at New York University.

This breadth keeps the book from being merely an artifact of the early-2020s metaverse boom. Virtual worlds become a test bench for Cartesian skepticism, the simulation hypothesis, mind and consciousness, extended cognition, ethics, political philosophy, language, and science. The central proposal is not that every digital experience is as good as every physical one. It is that digital composition alone does not make a world, object, or life unreal.

Four Senses of Real

"Real" is not one scalar property. At least four questions are being asked when a platform calls an object or world real:

  1. Ontological: does the environment contain stable entities, relations, and causal rules rather than a mere depiction?
  2. Experiential and social: do people form memories, relationships, identities, practices, or projects through it?
  3. Institutional: do organizations treat its records, scores, currencies, permissions, or events as grounds for action?
  4. Legal and political: who has enforceable rights, jurisdiction, remedy, and authority when the system changes or fails?

A yes at one level does not settle the next. An avatar can be a genuine social identity without being legally portable. A virtual item can be a real digital object while a user holds only a revocable contractual license. An emotionally important exchange can be real for a user without showing that the software has feelings. A generated scene can exist as a digital artifact without being accurate evidence of the external world.

For governance, the threshold is therefore not metaphysical prestige but dependency under power. Ask whether the state persists; whether a person or institution relies on it; whether an operator can unilaterally alter, rank, erase, monetize, or disclose it; and whether the state can trigger consequences for money, access, reputation, work, safety, or official records. When those conditions converge, "virtual" is no defense against duties.

The duty follows the function. Evidence needs provenance and corroboration. Identity needs authentication, impersonation controls, recovery, and a usable challenge process. Property-like interests need clear license, transfer, termination, and remedy terms. Relationship-like systems need role disclosure, privacy boundaries, and limits on dependency engineering. Training or evaluation environments need versioned claims about coverage, fidelity, and transfer to the real task.

This distinction turns a recurring feedback loop into an audit target. A model or world begins as representation; people coordinate around it; platforms attach permissions and incentives; institutions act on its records; later systems learn from those actions as if they were fresh observations. The representation has not become metaphysically ultimate. It has become operational infrastructure.

Virtual Realism

Chalmers's core claim is virtual realism: virtual worlds can be genuine worlds, virtual objects can be genuine objects, and life in virtual environments can be meaningful. In his earlier paper "The Virtual and the Real," he develops the point by treating virtual objects as digital objects, not as nonexistent physical objects. A virtual tree need not be a biological tree to be a genuine virtual tree.

This is a strong correction to media criticism that uses "just pixels" as a solvent. Bank balances, shared documents, game items, social graphs, avatars, and platform reputations can organize obligation, conflict, labor, identity, and status. Their dependence on computation and convention does not make their effects imaginary.

But three distinctions must remain intact. Real is not accurate: an existent generated scene may misrepresent a city or event. Meaningful is not conscious: a relationship may matter to a human without evidence of machine experience. Persistent is not owned: an item may endure in a database while its user lacks a transferable property right. Virtual realism removes one dismissive argument; it does not certify truth, moral status, title, safety, or justice.

The political question therefore begins where the ontological question ends. Who controls the servers, world state, identity namespace, terms, payment rails, moderation, recommender layer, backups, and exit? A digital object can be genuine while the user's access remains fragile because an operator can patch its properties, suspend the account, merge the economy, or close the service. Reality makes those losses matter; it does not make the arrangement fair.

Simulation Without Nihilism

Nick Bostrom's 2003 simulation argument is a philosophical disjunction about the future of civilizations and ancestor simulations, not empirical evidence that this world is simulated. Chalmers asks a different downstream question: if we were in a sufficiently rich simulation, would that make ordinary objects unreal and knowledge impossible? His answer is largely no. Simulated inhabitants could still know a world whose objects are implemented digitally.

That answer should discipline rather than inflate simulation talk. Uncertainty about substrate is not a revelation, and it does not dissolve duties of evidence, consent, care, or repair. Nor does it make all appearances equally trustworthy. Three claims require different tests:

A simulated red door can genuinely exist in a world while the reconstruction of an ancient city is historically wrong and a policy trained there fails on deployment. Chalmers's metaphysics supports the first claim. It cannot be borrowed as evidence for the other two.

This is also why ordinary interfaces deserve world analysis without being called complete simulations. Feeds, dashboards, rankings, maps, games, markets, search systems, and agent workspaces define visible objects, permitted actions, incentives, and feedback. They become real in a narrower sense when people adapt to their rules and institutions treat the resulting records as facts.

The AI-Age Reading

Read in 2026, Reality+ is useful less as a prediction about headsets than as a framework for systems that generate scenes, personas, records, and actions. A synthetic image can affect reputation; a companion's stored memory can shape attachment; a model summary can enter a case file; an agent can send a message, spend money, or change a database. None of those consequences depends on the output being physical or the model being conscious.

A world model, in the relevant engineering sense, represents or generates how an environment changes from one state to another, sometimes in response to an action. That label alone does not show that the model has learned the right causes. An interactive scene may be visually coherent while getting object permanence, geography, rare events, or the consequences of an action wrong. The site's world-model coverage test and structural-certification proposal turn that gap into three questions: which conditions were covered, which relations were preserved, and which claims transfer outside the generated environment?

Google DeepMind's August 2025 Genie 3 post is a useful bounded example. The company announced real-time, text-prompted interactive environments lasting a few minutes, while also listing limited action space, imperfect rendering of real locations, difficulty with multiple independent agents, and other constraints. The post establishes what the developer announced and the limitations it disclosed. It is not independent evidence that the generated environments reproduce real-world causal structure or are fit for safety-critical training.

The same discipline applies to social simulation. A companion exchange can become emotionally consequential for a user without establishing any inner experience in the software. The present duties arise from the human side of the relation: disclosure, intimate-data handling, age-appropriate design, dependency risk, escalation, and a non-punitive way to disengage. The site's synthetic relationship boundaries make role and authority explicit without pretending to settle machine consciousness.

Agentic systems add a different boundary. A generated plan becomes an external act only when authority, credentials, and tools let it cross into another system. That makes permission scope, confirmation, durable action logs, rollback, and appeal more useful than arguments over whether the agent is a person. Agent log receipts preserve the transition from proposed action to authorized action.

Current Context

As of August 12, 2026, the applicable governance is layered rather than metaverse-specific. Under the EU Digital Services Act, duties depend first on whether a service falls within the regulation's intermediary and platform categories. Online platforms have recommender-system transparency duties; services formally designated as very large online platforms or search engines because they reach at least 45 million average monthly recipients in the EU have additional systemic-risk assessment, mitigation, audit, data-access, and non-profiling recommender obligations. A virtual environment is not automatically a VLOP merely because it has avatars or an immersive interface.

The EU AI Act's Article 50 transparency duties began applying on August 2, 2026. In scope, they include disclosure when people directly interact with an AI system unless that fact is obvious, technically feasible machine-readable marking of certain synthetic outputs by providers, and deployer disclosure for deepfakes and some AI-generated public-interest text, subject to the article's stated exceptions. Regulation (EU) 2026/1744, the AI Omnibus in force since July 27, preserves that general start but gives providers of synthetic-content systems placed on the market before August 2 until December 2, 2026 to comply with Article 50(2)'s marking duty. The Commission's final Code of Practice on Transparency of AI-Generated Content, published June 10, is a voluntary compliance route; the legal duties come from the amended Act, not from signing the code.

C2PA's current 2.4 specification addresses a narrower problem: authenticated records about an asset's source and editing history. That can help a reviewer inspect custody and transformation, but it does not prove that a depicted event happened, that a caption is fair, or that an output is safe. Provenance is a map of claims and handling, not a truth oracle.

In the United States, the FTC's September 2025 orders to seven companion-chatbot companies opened a Section 6(b) study of testing, youth effects, disclosures, monetization, age controls, and conversation data. A 6(b) study has no specific law-enforcement purpose, so the inquiry should not be reported as an agency finding that a named service violated the law. The amended Children's Online Privacy Protection Rule separately strengthens protections for personal information collected by covered online services from children under 13, including limits on retention and separate parental consent for certain third-party disclosures; it is not a general safety code for all teenagers or all virtual worlds.

NIST's AI Risk Management Framework and Generative AI Profile offer lifecycle methods for mapping, measuring, managing, and governing risk. They remain voluntary guidance, and NIST says the framework is being revised. The categories are useful for an internal control program, but citing NIST does not by itself demonstrate compliance, safety, or effectiveness.

The Politics of Worldmaking

Chalmers's own book page raises the danger of a corporation-dominated metaverse. The concern follows directly from virtual realism: if a world can host genuine projects, relationships, possessions, and institutions, then control of its runtime is a form of political power. An operator can define names, visibility, movement, exchange, memory, moderation, and the conditions of continued existence inside the service.

That power is often divided among less visible layers. A platform may control the account namespace; a payment provider the economy; an app store access to the client; a model provider the behavior of synthetic characters; and a cloud host persistence of the state. Responsibility should follow control. Calling the visible experience a “world” must not hide the vendors that can change its rules or end it.

Portability exposes the difference between reality and rights. It has at least four layers: data export, the ability to retrieve records in a usable format; identity continuity, the ability to remain recognizably oneself; social continuity, the ability to preserve relationships or reach contacts; and functional interoperability, the ability to use an asset or capability elsewhere. A download may satisfy the first while doing nothing for the other three. Data portability should therefore never be used as shorthand for universal compatibility or a complete right of exit.

The same structure appears without a headset. A workplace dashboard, school platform, benefits portal, hiring system, feed, map, or agent workspace defines visible objects, recognized identities, permitted actions, evidence, and appeal. People adapt to those rules, and institutions act on the resulting records. The political question is not how immersive the interface feels. It is how much life must pass through it and how costly refusal becomes.

Nor does “virtual” mean immaterial. These worlds depend on hardware, energy, networks, content and moderation labor, physical bodies, and often spatial, voice, gaze, motion, or biometric data. A philosophy of digital objects should keep that supply chain visible. Otherwise users are invited to treat a service as transcendent while its operator treats their bodies and behavior as inputs.

Governance and Safety

The governing principle is proportionality to dependency and consequence. A private game session, a clinical simulator, a public-benefits portal, and an agent rehearsal environment are not the same system. Their controls should follow who can be harmed, which decisions can escape the environment, how difficult correction is, and what an operator can change unilaterally.

A useful audit separates four records:

  1. World-state register: rules, versions, model and asset changes, resets, retention, known failure modes, and continuity plans.
  2. Identity-and-authority register: human accounts, avatars, automated agents, authentication, delegated permissions, impersonation controls, recovery, and revocation.
  3. Evidence register: source claims, generated status, transformations, corroboration, human review, and the person or institution accepting publication responsibility.
  4. Consequence register: purchases, transfers, messages, moderation and access decisions, tool calls, external side effects, incidents, reversals, and appeal outcomes.

Keeping those records distinct prevents category laundering. A provider should not be able to describe a system as entertainment when defending weak controls, then treat the same system's scores, behavior, identity, or simulated results as authoritative when selling decisions. The evidentiary claim must be stated at the point of use.

Simulation assurance should likewise name the claim being tested. A coverage record states which conditions and edge cases were represented. A fidelity record states which causal or structural relationships were compared with external evidence. A transfer record states what happened when a policy, skill, or agent moved to the target environment. A polished demonstration supplies none of these automatically.

Safety review should cover at least four layers. Sensory and embodied safety includes motion, spatial boundaries, harassment, accessibility, and collection of gaze, voice, movement, and biometric data. Social safety includes impersonation, synthetic-role disclosure, dependency, minors, moderation, and crisis escalation. Economic safety includes pricing, scarcity, fraud, asset terms, dark patterns, and shutdown risk. Automated-action safety includes permission scope, confirmation, monitoring, external side effects, rollback, and incident response.

Accessibility belongs in the architecture, not in a late content review. W3C's XR Accessibility User Requirements discusses multimodal input and output, captions, adjustable pace, orientation, and clear control of experiences. It is a Working Group Note rather than a binding law or conformance standard, but it is useful design guidance for avoiding worlds that are operationally real only for users with one assumed body or sensory profile.

Meaningful exit requires more than an account-delete button. Users need notice of material rule or model changes, enough time and documentation to retrieve usable records, safe revocation of agents and credentials, clarity about retained data, continuity planning for paid or institutionally relied-on services, and notice and appeal when an automated or moderation decision affects access. Where interoperability is infeasible, the limitation and its consequences should be stated before dependence forms.

The result is not a demand to govern every fantasy as a state. It is a demand to stop letting presentation determine accountability. The more a system holds identity, evidence, property-like interests, intimate disclosure, or authority to act, the less plausible “only virtual” becomes as an answer to harm.

Where the Book Needs Pressure

The book's generosity toward virtual reality is philosophically productive, but ontology can outrun political economy. A world may be meaningful while surveilling its inhabitants, host genuine relationships while optimizing them for retention, or contain genuine digital objects while granting users only revocable access. Chalmers does address corporate rule; the harder task is to connect that warning to contracts, infrastructure, labor, competition, and enforceable remedy.

The argument also needs a firmer account of the material substrate. Digital worlds are maintained through data centers, devices, networks, energy, mineral supply chains, paid and unpaid creation, content moderation, and care work. Implementation as digital objects does not make those dependencies incidental. When a service fails, the loss is distributed according to who owns the infrastructure and who merely lived through it.

Most importantly, reality does not settle rights. Ontology cannot tell us whether a virtual item is property, whether an avatar is protected identity, whether an operator owes continuity, or when a simulated result is admissible evidence. Those are legal and institutional judgments. The book clears away the claim that digital things are necessarily unreal; it does not supply the missing title, jurisdiction, due process, or public legitimacy.

The 2022 book also precedes today's ordinary combination of generated media, persistent model memory, companion interfaces, and agents with tools. Its framework travels well, but the risk unit has changed from a bounded virtual world to a chain of models, vendors, records, and external actions. That shift makes provenance and authority transitions at least as important as immersion.

Finally, moral-patient thought experiments should not blur present evidence. This review makes no claim that current AI systems are conscious, divine, or AGI. A future digital subject, if one were established, would raise welfare and political questions. Today's platforms already raise human safety, privacy, labor, consumer-protection, identity, and accountability questions without that premise.

Source Discipline

Claims about Reality+ come from Chalmers's book page and teaching guide, the publishers' edition records, his primary paper on virtual realism, and an academic review. These sources support the attribution and bibliographic record; they do not show that any current product satisfies the book's philosophical criteria. Bostrom's original article establishes the form of the simulation argument, not that our world is simulated.

Current obligations are sourced to enacted EU regulations, including the July 2026 AI Omnibus amendment, and official Commission explanations. The Commission's Article 50 code is identified as voluntary even though the underlying statutory duties are binding, and the transition for certain pre-August systems is stated separately. The FTC companion inquiry is identified as a Section 6(b) study, not an enforcement result. C2PA supplies a provenance specification; NIST supplies voluntary risk guidance under revision; the W3C document supplies non-binding accessibility requirements and considerations. None is presented as a universal certification.

DeepMind's post supports only what the developer announced about Genie 3 and the limitations it disclosed. It does not establish independent fidelity, safety, or effectiveness. The four senses of real, dependency-under-power threshold, portability layers, registers, and safety layers are this review's synthesis. They are tests proposed for scrutiny, not claims that Chalmers or any cited body endorsed this site's framework.

What This Changes

The durable lesson is to stop using unreality as an excuse, while refusing to let “real” do more work than it can.

The recurring sequence is concrete: a representation becomes a shared convention; the convention becomes a dependency; an operator attaches rules and incentives; an institution treats the resulting state as evidence; that action enters a record; and later models learn from the record as though it were a fresh observation. At each transition, mediation acquires power. At none does existence alone prove accuracy, ownership, fairness, or legitimacy.

A practical review therefore asks: What persists? Who relies on it? Who can change it? Which actions escape the interface? What evidence supports those actions? How can a person inspect, challenge, move, or revoke the relevant state? What survives a shutdown? Those questions connect the site's concerns with provenance, memory, identity, authority, feedback, and repair without requiring a single answer about the ultimate substrate of reality.

Reality+ is most useful when its metaphysics becomes an anti-dismissal rule. Digital composition cannot by itself erase a harm, relationship, possession, or duty. It also cannot confer truth or justice. The work begins in the gap between a world being real enough to matter and its institutions being accountable enough to trust.

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