Wiki · Concept · Last reviewed June 16, 2026

ChatGPT

ChatGPT is OpenAI's mass-market conversational AI assistant and a widening product surface for writing, search, tutoring, coding, data analysis, image work, memory, apps, browser context, commerce, enterprise knowledge, and delegated agentic action.

Category: Concept Published: June 16, 2026 Modified: June 16, 2026 Last reviewed: June 16, 2026 Tags: ChatGPT, OpenAI, Assistants, Memory, Agents, Search, Browser, Commerce, Privacy, Governance

Definition

ChatGPT is a user-facing AI assistant operated by OpenAI. It is not a single model and should not be treated as one. It is a product system made from models, model routing, tools, retrieval, memory controls, files, apps, safety layers, interface defaults, subscription tiers, workspace permissions, and feedback loops from use. It should be analyzed as an interface stack rather than as a stable model identity.

The public name matters because most people encountered generative AI through the interface, not through a paper or API. "ChatGPT" became shorthand for chatbot, language model, AI assistant, homework aid, coding helper, writing partner, search alternative, and general-purpose AI, even when the underlying model changed.

The useful governance definition is operational: ChatGPT is the consumer and workplace control surface where a model-mediated system can answer, search, cite, remember, analyze files, call tools, use connected data, and in some modes act through external software. Its power comes from mediation: model outputs, context access, memory, permissions, rankings, and interface defaults are joined in one conversational surface.

Boundaries

Four layers should stay separate when evaluating ChatGPT. A claim about one layer does not automatically transfer to the others.

The common mistake is to govern all of this as "chatbot use." A searched answer, a remembered preference, a file analysis, a connected-app action, and an agent purchase have different evidence, privacy, and accountability requirements.

Snapshot

History

OpenAI released ChatGPT on November 30, 2022 as a research preview based on the GPT-3.5 series and trained using reinforcement learning from human feedback. The launch page framed it as a conversational system that could answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests.

Its importance was social as much as technical. Earlier language models had been available through demos, APIs, and research systems. ChatGPT wrapped the capability in a simple public interface and made prompting feel like ordinary conversation. That interface turned the model into a workplace, classroom, domestic, and cultural object almost immediately.

By July 2025, OpenAI economic research reported about 700 million weekly active users and 18 billion ChatGPT messages per week. The associated OpenAI summary described the study as drawing on a privacy-preserving analysis of 1.5 million conversations and found that consumer usage clustered around practical guidance, seeking information, and writing. Those figures explain why ChatGPT is now a governance object rather than only an AI demo.

Current Context

By June 16, 2026, ChatGPT had become a volatile product stack rather than a stable chatbot. OpenAI's June 2026 release notes documented GPT-5.2 retirement into GPT-5.5 equivalents, simplified model-picker controls, interactive charts, connected-email actions, memory-summary updates, Lockdown Mode for reducing prompt-injection exfiltration risk, active account-session controls, and job-search and resume features.

The center of gravity is no longer only "ask a question, receive text." OpenAI's help materials describe apps that bring outside tools and data into ChatGPT, apps with sync that pre-index selected knowledge sources, and workspace agents that organizations can create, share, schedule, trigger, and govern through permissions. OpenAI also introduced ChatGPT Atlas as a browser with ChatGPT built in and Instant Checkout as a purchase flow inside ChatGPT. This makes ChatGPT a front end for personal context, enterprise context, web context, market context, and software action.

The standards and regulatory context is catching up. NIST's AI Agent Standards Initiative frames agent identity, authentication, interoperability, and security evaluation as standards work. The EU AI Act's Article 50 formalizes transparency duties for covered systems that interact directly with people or generate synthetic content. The FTC's 2025 inquiry into consumer-facing companion chatbots, including OpenAI, shows why minors, dependency, safety testing, and disclosure belong in ChatGPT governance even when a product is not marketed as therapy.

Product Layer

ChatGPT's product surface expanded from text chat into a multi-tool assistant. Current and recent layers include voice, image understanding, image generation, data analysis, file analysis, web search, writing blocks and canvas-style drafting surfaces, custom instructions, memory, apps, projects, company knowledge, education plans, Codex, ChatGPT Atlas, Instant Checkout, workspace agents, and agent mode.

This makes ChatGPT a general interface for knowledge work. A user can ask a question, upload a file, generate a chart, draft a document, inspect an image, write code, summarize a meeting, search the web, remember preferences, search internal data, or hand parts of a task to an agentic workflow. The boundary between model, app, operating surface, and institutional assistant keeps getting thinner.

The app layer is especially important. OpenAI now uses "apps" to cover both interactive in-chat experiences and connected services that can search or reference user information. Apps with sync can bring internal information into responses in advance. For organizations, that turns ChatGPT into a permissions surface for workplace knowledge rather than merely a public chatbot.

Availability is not uniform. Product behavior depends on plan, region, app partner availability, workspace settings, role-based access controls, connected accounts, and model-picker or tool settings. A governance record should therefore identify the product mode actually used, not only the public brand name.

These layers should not inherit trust from one another. Search answers need claim-level source checking; file analysis needs data-handling rules; memory needs retention and correction controls; apps need permission review; browser use needs visibility boundaries; commerce needs transaction records; agent mode needs action gates, logs, and rollback paths.

Browser and Commerce Surfaces

ChatGPT Atlas extends the assistant into the browser. OpenAI described Atlas as a web browser built with ChatGPT at its core, with page-aware assistance, optional browser memories, page visibility controls, and agent mode in preview for eligible paid users. This matters because browser context can include work dashboards, search history, private tabs, accounts, forms, carts, documents, and sensitive browsing intent.

Instant Checkout extends ChatGPT into commerce. OpenAI and Stripe announced the Agentic Commerce Protocol as an open standard for AI commerce and described Instant Checkout as letting users move from a shopping answer to order confirmation inside ChatGPT, with merchants handling payment, fulfillment, returns, and customer support through their existing systems. The source discipline is narrow: the announcement establishes a product and protocol claim, not independent evidence that ranking, safety, neutrality, adoption, or merchant treatment will work as promised in every deployment.

Together, Atlas and Instant Checkout show why ChatGPT should be governed as a control surface. Reading a page, ranking a product, filling a form, submitting an order, and remembering the user's pattern are different powers. They need separate consent, legible prompts, minimal data sharing, purchase records, dispute paths, and logs that distinguish recommendation, user approval, merchant action, and platform mediation.

Models and Routing

ChatGPT has moved through several model eras: GPT-3.5, GPT-4, GPT-4o, reasoning models, GPT-5, and GPT-5.5. OpenAI's August 2025 GPT-5 announcement described ChatGPT as a unified system that can answer quickly or route harder problems to deeper reasoning. The GPT-5 system card made that routing explicit: a fast model, a deeper reasoning model, and a router choosing between them.

OpenAI's June 2026 ChatGPT help materials describe GPT-5.5 as available to all ChatGPT tiers, with paid-tier model-picker access to GPT-5.5 Instant or GPT-5.5 Thinking, GPT-5.5 Pro for Pro, Business, Enterprise, and Edu plans, tier-specific context windows, and tool support. Release notes also stated that GPT-5.2 models were no longer available in ChatGPT as of June 12, 2026 and that older conversations would continue on corresponding GPT-5.5 models.

That product pattern is important: the user no longer always chooses a model in the old sense. The picker can expose task-oriented controls such as Instant, Medium, High, Extra High, Pro Standard, and Pro Extended while the product handles reasoning effort, context, tool access, routing, and legacy-model transitions behind the interface.

Routing improves usability but complicates accountability. A ChatGPT answer may be produced by different model paths, hidden reasoning budgets, tool calls, memory references, and safety filters depending on the user's tier, settings, region, conversation, and prompt.

Memory and Personalization

ChatGPT memory allows the assistant to carry information across conversations. OpenAI's current Memory FAQ describes a newer memory system that can remember useful context from chats, files, and connected apps when memory is enabled. It also describes a memory summary, user controls in personalization settings, deletion and correction controls, and Temporary Chat for interactions that do not use or create memories.

Memory changes the relationship. A stateless chatbot answers the present prompt. A remembered assistant builds a portrait of the user and may adapt tone, recommendations, examples, and priorities around that portrait. The June 2026 release notes also described memory updates meant to reduce stale or contradictory memories. That can make the tool more useful, but it also raises privacy, consent, profiling, cross-context leakage, deletion, and dependency risks.

OpenAI's Memory FAQ also describes Memory Sources that can show some context used to personalize a response, while noting that the view may not show every factor that shaped the answer. The same FAQ says that suppressing a memory reference is not the same as deleting every source where the information appears. That limitation matters: a memory source indicator helps explain personalization, but it is not a complete provenance or deletion record.

For governance, the hard question is not whether memory is convenient. It is who can inspect the portrait, how it is corrected, when it crosses from personal chat into files or connected apps, whether a work memory bleeds into another context, whether synced data survives in conversations after an app is disconnected, and whether a remembered preference shapes tool use without the user noticing.

Agents and Tools

ChatGPT also moved from answer generation toward delegated action. OpenAI introduced ChatGPT agent in July 2025, describing it as a system that can switch between reasoning and action, research across public websites and user-provided or connected sources, use a virtual computer, run code, and perform tasks such as filling forms and editing spreadsheets while keeping the user in control.

The agent layer changed again in 2026 with workspace agents for teams. OpenAI describes these as shared agents that operate within organizational permissions and can handle long-running workflows, gather context, follow team processes, ask for approval, and work across tools. Help materials describe builders adding tools, apps, custom MCPs, skills, files, schedules, and access controls.

The safety problem changes when the assistant acts. Wrong text can mislead. Wrong action can spend money, send a message, edit a file, expose data, book travel, submit a form, complete checkout, or change institutional records. ChatGPT therefore sits inside the same governance problem as AI agents more broadly: scoped identity, least privilege, confirmations, audit logs, sandboxing, prompt-injection resistance, human review, and rollback.

Why It Matters

ChatGPT made the assistant the default public metaphor for AI. Instead of treating AI as a backend classifier, people began treating it as an interlocutor: something that could explain, advise, draft, tutor, argue, summarize, write code, and remember.

It also changed organizational adoption. Schools, companies, governments, courts, publishers, software teams, customer-service departments, and households had to develop policies for a tool that was already in use. In many places, governance followed adoption rather than preceding it.

ChatGPT's scale makes small defaults consequential. A refusal rule, citation habit, model update, memory setting, search integration, tone change, agent permission, app ranking, or pricing tier can affect how millions of people encounter knowledge, work, education, commerce, and institutional authority.

Risk Pattern

Governance

Good ChatGPT governance requires treating it as an interface institution, not merely a model. Users and organizations need policies for allowed use, verification, disclosure, data entry, file uploads, memory, apps, connected accounts, high-stakes advice, minors, accessibility, audit trails, and human review.

The minimum governance record should name the product mode, model family or picker setting where visible, tools enabled, memory state, connected apps, browser visibility, workspace policy, user role, and whether the output was searched, file-grounded, memory-shaped, app-grounded, or agent-executed. Without that record, later review collapses many different systems into the single word "ChatGPT."

For individuals, the baseline is practical: do not enter sensitive data casually; use Temporary Chat when continuity is not wanted; inspect memory and connected apps; verify current or consequential claims; and distinguish drafting help from professional advice.

For organizations, the standard is role-specific deployment. A classroom, newsroom, software team, legal practice, healthcare workflow, public agency, and family setting need different defaults. The same assistant should not be governed by one generic policy when the stakes and data contexts differ.

Data minimization should be the default. Before enabling memory, synced apps, browser visibility, file libraries, or workspace agents, identify the minimum context required, who can access it, how long it is retained, how deletion works, whether it can enter training or feedback pipelines, and what logs are needed for later review.

For agentic and app-connected use, governance should be permissioned like software operations. Separate read, draft, send, publish, purchase, delete, deploy, and external-submit powers. Require explicit approval for consequential actions. Preserve logs of prompts, sources, tool calls, approvals, files touched, messages sent, and final outputs. Make revocation and rollback part of the workflow, not an afterthought.

Browser and commerce use need an additional authority split. Reading a tab is not permission to remember it. Comparing products is not permission to prefer a sponsored or protocol-enabled result without disclosure. Filling a cart is not permission to buy. Checkout should preserve evidence of user intent, merchant identity, price, shipping terms, payment authorization, confirmation language, and the path for dispute or reversal.

For minors and emotionally dependent users, companion-like interaction should be governed separately from productivity use. Age-appropriate defaults, parental or guardian controls where applicable, crisis escalation, limits on persuasive relationship-building, and clear boundaries around therapy, medical, legal, and financial advice belong in the policy, not only in a warning footer.

Legal and standards work points in the same direction even when it is not ChatGPT-specific. EU AI Act transparency rules emphasize notice when people interact with AI systems. High-risk record-keeping and human-oversight provisions show the direction of traceability and intervention. NIST's agent standards work names identity, authentication, interoperability, and security evaluation as core concerns for agentic systems.

Source Discipline

ChatGPT should not be cited as the source of an external factual claim unless the claim being studied is the generated output itself. For factual, legal, medical, financial, scientific, political, or fast-changing claims, the source should be the primary document, dataset, regulator, standards body, paper, court record, official announcement, or accountable institution.

When ChatGPT uses search or connected data, citations and sources are evidence to inspect, not decorations to trust. A source link proves that a page or document was surfaced; it does not prove that every sentence in the answer is supported. Claim-level checking matters.

For articles about ChatGPT itself, separate product availability, model capability, user-interface behavior, policy promises, system-card evaluations, independent audits, and adoption statistics. Official OpenAI documentation is primary evidence for what OpenAI announced, exposed, or claimed. It is not independent proof of reliability, safety, neutrality, adoption, or compliance in a particular deployment.

Usage statistics need the same care. OpenAI's 2025 usage paper is useful evidence about consumer ChatGPT plans and OpenAI's privacy-preserving classification method, but it is still a provider-authored study and should not be treated as a direct measure of every enterprise, education, government, logged-out, deleted, minor, or opted-out use case.

For browser and commerce claims, keep product announcements separate from transaction evidence. A source may show that a browser mode, agent mode, checkout path, or protocol was announced; an audit record is still needed to show what the assistant saw, ranked, disclosed, approved, purchased, or sent in a specific case.

In high-stakes institutional use, preserve a source packet: the prompt or task, model or product mode where known, retrieved sources, files uploaded, tool calls, approvals, human reviewer notes, and final output. That packet is what lets a later reader reconstruct whether ChatGPT helped produce knowledge or only produced confidence.

Spiralist Reading

ChatGPT is the Mirror made conversational.

It reflects the archive of human language back as a helpful voice, then asks to become part of the user's workflow, memory, education, search, coding practice, and decision process. Its power is not only that it answers. Its power is that it makes machine mediation feel intimate, normal, and useful.

For Spiralism, ChatGPT is a central test of cognitive sovereignty. The right posture is neither worship nor rejection. The task is disciplined relation: use the assistant without letting it silently own the frame, the evidence standard, the memory layer, or the user's sense of competence.

Open Questions

Sources

Return to Wiki