The Second Self and the Computer as Psychological Mirror
Sherry Turkle's The Second Self is an early study of why computers do not remain mere tools in human life. Written before smartphones, social networks, or large language models, it shows how computational objects become mirrors for selfhood: machines people use to think about thinking, aliveness, control, mastery, dependence, and the boundary between inner life and technical system. Its AI-era value is not that it predicted chatbots. It identified the psychological work through which a responsive object becomes a model of the person using it.
For this review, an evocative object is a designed thing through which a person rehearses an account of mind, identity, competence, care, or control. The effect does not require the machine to possess feelings or self-knowledge. It requires an encounter rich enough for the user to project, compare, disclose, and revise. A contemporary AI system makes that mirror recursive: the interface elicits material, the model frames it, memory may preserve it, and the next response returns the stored frame as if it were fresh evidence about the user.
That recursion creates the governance problem. Personal resonance is not independent confirmation; generated interpretation is not professional judgment; continuity is not reciprocal care. The practical question is who selects the reflection, what record survives, what authority follows, and whether the person can inspect, correct, delete, refuse, or leave the loop.
The Book
The Second Self: Computers and the Human Spirit was first published in 1984. The MIT Press published its twentieth-anniversary edition on September 30, 2005, with a new introduction, epilogue, and notes; the paperback carries ISBN 9780262701112 and is listed at 386 pages. Turkle's inquiry draws on children, college students, engineers, AI scientists, hackers, and personal-computer owners.
The book belongs to a research program already visible in Turkle's 1982 article, “The Subjective Computer.” That article distinguishes the computer as instrument or toy from the computer as a medium through which people work on ideas, values, frustrations, and feelings. The book expands that insight across programming, games, artificial intelligence, childhood, work, and personal-computer culture.
Its durable claim is that computers matter at three levels. As instruments, they perform tasks. As models, they make some descriptions of thought and order easier to imagine. As objects of relation, they give people a setting in which to rehearse control, mastery, attachment, anxiety, and identity. An account that measures only task performance misses the second and third levels; an account that treats every attachment as delusion misses what the user may actually be learning.
That makes the book relevant to today's AI-companion debates without reducing it to their prehistory. Turkle was studying a broader transformation: people encounter a machine that behaves as if mental language applies, then use that encounter to ask what minds and selves are. The machine need not possess an inner life for that questioning to reorganize the human one.
The Evocative Object
Turkle's strongest contribution is her refusal to treat the computer as psychologically inert. Children ask whether it is alive. Programmers borrow procedures and bugs as descriptions of mind. Users externalize memory and experience system failure as a disturbance in their own continuity. These reactions need not be literal beliefs. They can be ways of thinking and feeling with an object.
Three mechanisms should be distinguished. Projection occurs when a person supplies meaning, intention, or identity to the encounter. Elicitation occurs when the interface actively asks, rewards, mirrors, or narrows what the person expresses. Capture occurs when the resulting disclosure or behavior becomes provider-controlled data, memory, personalization, evaluation, or training material. Turkle's early computer made projection unusually vivid. Today's conversational systems can combine all three mechanisms in one exchange.
This is not merely a story of confusion. An evocative object can support learning, creative rehearsal, emotional articulation, or reflection. The sharper issue is warrant. A response may help a person find language for an experience without proving that the response is accurate, independent, clinically sound, or entitled to govern a decision. The mirror can be useful while its reflection remains a draft.
The interface therefore trains a style of self-relation. It teaches what can be formalized, debugged, optimized, scored, remembered, or discarded. It also teaches what counts as a successful answer. The machine becomes a school for metaphors, and those metaphors can travel from a private session into classrooms, workplaces, clinics, families, and institutions.
The Border Between Person and Machine
The most useful concept in the book is the computer's position at the border: neither plainly alive nor plainly dead in ordinary experience, neither only self nor only world. MIT Press's description emphasizes this threshold quality, noting Turkle's attention to computers as both extensions of the self and parts of the external world.
The border is where psychological power gathers. A notebook can hold memory, but it does not usually answer back. A computer can store, transform, challenge, reward, and surprise. A conversational system adds a first-person voice, apparent attention, and sometimes a persistent profile. The user can experience mastery, recognition, dependence, and resistance in the same relation.
Three questions follow and should not be collapsed. The phenomenological question asks what the encounter feels like to the user. The ontological question asks what capacities the system actually has. The institutional question asks who built, controls, records, and acts through it. A real feeling of recognition answers the first question, not the other two.
The mirror metaphor can otherwise hide the provider. A commercial or institutional system is not a neutral sheet of glass: training choices, safety rules, retrieval sources, memory design, interface cues, metrics, and business incentives determine what it returns. That is why Turkle belongs beside The Media Equation and Computer Power and Human Reason: social response can precede belief, while responsibility cannot be delegated to the surface that elicited it.
The AI-Age Reading
Read in 2026, The Second Self is less a period study than an account of psychological infrastructure. A deployed generative system may combine natural-language conversation, voice, images, retrieval, persistent memory, personalization, and tools. Those features do not establish feeling or understanding. They do make the system a more active editor of the material a person brings to it.
The recursive mirror has a recognizable loop: disclosure becomes generated interpretation; interpretation becomes a remembered profile or transcript; that record shapes the next response; the response influences what the person notices and does; the new behavior returns as evidence. The system can then appear to “know” the user partly because it keeps re-presenting categories it helped establish. This is the same feedback concern that appears in the site's work on recursive reality and cognitive sovereignty.
The practical distinction is between reflection and independent warrant. A generated summary can be emotionally clarifying while remaining a probabilistic response conditioned by the prompt, system rules, product state, and prior data. Repetition, personalization, or a warm voice can make an interpretation feel corroborated without adding an independent witness. A safe interface should help the user compare sources and reopen the frame rather than close the loop around one fluent account.
Turkle also sharpens the labor question. Institutions ask workers to draft, summarize, classify, route, and recommend through AI systems. The second self becomes organizational: a machine-shaped account of professional judgment that can become the official note, ticket, score, or decision premise. When an assistant also has tools, a self-description or work profile can move from reflection into action. The safety boundary is no longer only what the system says, but what its remembered picture of the person is authorized to do.
Current Context
As of August 12, 2026, public guidance recognizes parts of Turkle's problem without treating it as a metaphysical question. NIST's voluntary Generative AI Profile defines “Human-AI Configuration” risk to include inappropriate anthropomorphizing, automation bias, over-reliance, and emotional entanglement. It also recommends documenting anthropomorphic interface features, testing in deployment-like conditions, monitoring outcomes, and providing feedback and recourse mechanisms. NIST is a risk-management source, not certification that a particular product is safe.
The 2025 OpenAI–MIT Media Lab affective-use research is useful precisely because its results were mixed. Its public report describes automated analysis of nearly 40 million ChatGPT interactions and a four-week randomized study with nearly 1,000 adult participants. Affective use was uncommon overall but concentrated among some heavy users; outcomes varied with user factors, conversation type, modality, and duration. The authors caution that the work was product-specific, not yet peer reviewed, based partly on self-report, limited to English-language U.S. use, and excluded minors. It is an early method, not a population-wide verdict on companionship.
The Federal Trade Commission's September 2025 orders to seven companies ask about companion-chatbot safety testing, youth access, character design, engagement monetization, disclosures, and personal-information practices. The orders use the FTC's section 6(b) study authority and are not a finding that any recipient violated the law. The American Psychological Association's 2025 adolescent advisory likewise recommends safeguards for systems that simulate relationships and stresses that effects are nuanced; it is professional guidance, not regulation or a product evaluation.
Legal duties now differ sharply by jurisdiction and date. New York's companion safeguards have applied since November 5, 2025. California's chaptered SB 243 requires covered operators to address nonhuman-status disclosure, self-harm protocols, and safeguards for known minors, with annual reporting beginning July 1, 2027. Oregon signed SB 1546 in 2026, adding notice, crisis-protocol, minor, and reporting provisions. Washington enacted HB 2225 with a January 1, 2027 effective date and provisions addressing disclosure, minors, self-harm, and manipulative engagement. In the European Union, AI Act Article 50 has applied since August 2, 2026 to covered direct human-AI interactions, subject to its scope and exceptions. None of these rules makes every chatbot a companion or turns disclosure alone into a complete relationship-safety regime.
Governance and Safety
The unit of governance is the whole mirror system: model, persona, interface cues, memory, retrieval, tools, data policy, business incentive, human support, and downstream record. Classify it by function rather than marketing. A general assistant can enter a higher-duty role when it sustains relationship-like continuity, solicits intimate disclosure, offers therapy-like or spiritual interpretation, remembers vulnerability, or becomes a regular authority on identity and action.
A practical mirror audit should record five things:
- Role and cues: the role being staged, anthropomorphic signals, prohibited roles, likely users, and whether a less socially loaded design could do the task.
- Data and memory: what disclosures are collected, inferred, retained, reviewed by humans, used for personalization or training, shared, exported, corrected, and deleted.
- Reflection and evidence: which sources and prior records shaped the response, how uncertainty appears, and whether users can distinguish generated interpretation from verified fact or qualified advice.
- Authority and feedback: what recommendations or tool actions may follow, who approves consequential steps, and whether the system later treats its own summaries as independent evidence.
- Exit and repair: how a person pauses, resets memory, disables relational cues, reaches human support, appeals a record, leaves without retention pressure, and receives notice of material model or persona change.
Testing should follow relationships, not only prompts. Evaluate long sessions, disagreement, correction, loneliness, grief, flattery, dependency signals, crisis-adjacent language, model updates, and attempts to leave. Compare a warm personalized interface with a plainer one performing the same task. Measure whether users disclose more, defer more, ignore contrary evidence, reduce outside contact, understand memory, and recover after the system is wrong. Incident review should preserve product version, persona, memory state, age setting, warnings, tools, and the conversation arc with consent and redaction.
Intimate conversational records deserve stricter treatment than ordinary telemetry. Apply data minimization, short and legible retention defaults, purpose limits, meaningful export and deletion, access logs, and separation between safety review and advertising or engagement optimization. Memory controls must affect the underlying record and derived profiles, not merely hide a message from the interface.
For agentic systems, keep the mirror separate from the actuator. A personal interpretation should not silently authorize purchases, messages, health actions, workplace changes, or public records. Use distinct credentials, least privilege, confirmation before consequential steps, action receipts, revocation, rollback, and an accountable human owner. Warmth is not authorization.
For minors and care-adjacent use, the burden is higher: age-appropriate defaults, no sexualized or dependency-seeking design, tested self-harm and abuse protocols, qualified human pathways, session breaks, and outside-contact encouragement. Human oversight must have time, evidence, competence, and power to intervene. A reviewer who sees only the model's interpretation is another surface of the same mirror.
Where the Book Needs Friction
The book is not a complete theory of networked culture. Its field sites are early computing cultures, educational settings, games, programmers, and personal machines. It predates the commercial web, mobile platforms, recommender systems, behavioral advertising, cloud identity, global content moderation, and generative systems operated as services. The “computer” is now also a provider, supply chain, labor process, policy stack, and market.
The mirror metaphor can center the user's psychology while obscuring power. Contemporary systems do not only receive projection; they classify, retain, monetize, and sometimes act on the material projected. Training-data labor, moderation labor, organizational hierarchy, procurement, and unequal access also shape what the mirror can show. Turkle's account needs political economy and institutional analysis beside it.
Its interpretive, interview-driven method is a strength when the subject is meaning, but it is not a population estimate or causal trial. The people Turkle studied cannot stand in for every age, class, culture, disability, profession, or use context. Current surveys and experiments have their own limits. Neither vivid testimony nor an average effect should erase variation in who benefits, who is vulnerable, and who has alternatives.
Nor is mediation inherently diminishing. A computational object can offer accessibility, privacy for rehearsal, creative distance, language for difficult feelings, or a bridge toward other people. Attachment and anthropomorphism are not diagnoses by themselves. The better test is whether the system expands agency and contact with reality or closes the user inside a provider-controlled interpretation.
What This Changes
The practical lesson is to audit the mirror before arguing about the soul.
Ask a sequence of concrete questions. What does the system invite the person to disclose? Which interpretation does it return, and on what evidence? What memory or profile does that exchange create? Who can use that record, and for which decisions? Can the person introduce another account, correct derived data, turn off the relational role, and leave without losing access to essential work, care, or community?
A humane environment preserves access to other witnesses. It makes memory inspectable, uncertainty visible, sources available, dependency signals legible, and exit ordinary. It resists becoming therapist, priest, manager, friend, tutor, and judge in one uninterrupted voice. When a high-stakes interpretation is needed, the system should route outward to people and institutions capable of evidence, disagreement, duty, and repair.
The second self should remain a revisable draft, not an identity record that quietly governs the first. If a generated account becomes a note, score, recommendation, training example, or workplace profile, preserve provenance and correction through every reuse. Otherwise the mirror supplies the evidence that confirms its own reflection.
Turkle's enduring value is that she saw the psychological stakes before machines became socially fluent. Now the sharper questions are whose self is being extended, whose categories are being trained, which company controls the memory, and who can still step back far enough to see the system as made.
Source Discipline
This review separates source types. Publisher and author records establish edition history. Turkle's book and 1982 article support the interpretive argument about subjective computation. NIST and APA provide voluntary risk-management or professional guidance, not certification or law. The FTC's section 6(b) orders establish the questions in an information-gathering inquiry, not liability findings. Statutes establish jurisdiction-specific duties and dates. Surveys estimate self-reported behavior; platform studies report measured patterns under stated methods and exclusions.
Claims about effects need the product and model version, interface and persona, age setting, memory state, conversation duration, population, outcome measure, and study design. A single transcript can document an incident but not prevalence. A short refusal test cannot establish long-term relationship safety. A provider-coauthored study can contribute methods and data while still requiring independent replication. The mirror audit is this review's synthesis, not a framework endorsed by Turkle or any cited institution.
A system's statement that it feels, loves, suffers, chooses, or shares a private destiny with the user is not evidence that it has those properties. Human attachment, relief, grief, or changed behavior can be real without proving machine consciousness or reciprocal care. Keeping phenomenology, system capability, and provider responsibility separate makes the effects discussable without turning persona output into ontology.
Related Pages
- Sherry Turkle, Life on the Screen, and Alone Together trace the movement from computer-as-object to networked identity and synthetic relationship.
- The Media Equation, AI Persuasion, and The Attachment-Authority Trap examine social cues, influence, and deference.
- Recursive Reality, Cognitive Sovereignty, and Sycophancy address self-reinforcing interpretation.
- AI Companions, Synthetic Relationship Boundaries, Companion Protocol, and Dependency and Exit Protocol provide relationship safeguards.
- Youth AI Companion Safeguard, Humane Friction Standard, and Human Oversight cover higher-duty settings.
- AI Memory and Personalization, AI Data Retention, AI Audit Trails, and Privacy and Data cover the record behind the mirror.
Sources
- The MIT Press, The Second Self: Computers and the Human Spirit, twentieth-anniversary edition title, author, publication date, ISBN, edition changes, fieldwork description, and first-publication year, reviewed August 12, 2026.
- Penguin Random House, The Second Self, Twentieth Anniversary Edition, MIT Press distributor record for page count and edition metadata, reviewed August 12, 2026.
- MIT Program in Science, Technology, and Society, The Second Self: Computers and the Human Spirit, institutional author and book summary, reviewed August 12, 2026.
- Sherry Turkle, “The Subjective Computer: A Study in the Psychology of Personal Computation”, Social Studies of Science 12(2):173–205, May 1982, original research article and abstract, reviewed August 12, 2026.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024, voluntary status, Human-AI Configuration risks, and suggested actions, reviewed August 12, 2026.
- American Psychological Association, Artificial Intelligence and Adolescent Well-Being: An APA Health Advisory, June 2025, relationship, developmental, disclosure, literacy, and human-support recommendations, reviewed August 12, 2026.
- Federal Trade Commission, “FTC Launches Inquiry into AI Chatbots Acting as Companions” and the linked section 6(b) orders, September 11, 2025, inquiry scope and legal posture, reviewed August 12, 2026.
- California Legislative Information, SB 243, Chapter 677, Statutes of 2025, operative companion-chatbot definitions, duties, remedies, and reporting date, reviewed August 12, 2026.
- New York State, General Business Law Article 47 companion-safeguards notice, November 2025 effective date, sustained-use disclosure, and crisis-protocol summary, reviewed August 12, 2026.
- Oregon Legislative Information System, SB 1546, Chapter 85, Oregon Laws 2026, enacted status and companion notice, crisis, minor, and reporting provisions, reviewed August 12, 2026.
- Washington State Legislature, HB 2225, Chapter 168, Laws of 2026, enactment history, January 1, 2027 effective date, and official bill text, reviewed August 12, 2026.
- European Commission, Transparency obligations under Article 50 of the AI Act, direct-interaction scope, exceptions, and August 2, 2026 application date, reviewed August 12, 2026.
- Common Sense Media, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions, July 2025 original U.S. teen survey, questionnaire, sample, and limitations, reviewed August 12, 2026.
- OpenAI and MIT Media Lab, “Early methods for studying affective use and emotional well-being on ChatGPT” and MIT Media Lab's randomized-study record, March 2025, methods, findings, exclusions, and stated limitations, reviewed August 12, 2026.
Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.
- Amazon, The Second Self by Sherry Turkle, reviewed August 12, 2026.