ChatGPT

ChatGPT is a conversational artificial intelligence service developed by OpenAI and publicly launched on November 30, 2022. It uses large language models to respond to instructions and questions in natural language. Originally introduced as a text-based research preview, it has expanded to include capabilities for working with images, audio, documents, computer code, and online information. ChatGPT is a user-facing product rather than the name of a single, unchanging model; its capabilities depend on the models, tools, and service settings available to a user. (openai.com)

The service brings generative artificial intelligence into everyday activities such as writing, learning, information seeking, and problem solving. Research on its use documents both workplace and non-workplace applications, while its development raises questions about the reliability of generated information, the handling of personal data, and the security of tool-assisted systems. Understanding ChatGPT therefore requires distinguishing its underlying technology from its interface and commercial service, and its demonstrated capabilities from limitations such as hallucination. (cdn.openai.com)

Product identity and scope

ChatGPT is OpenAI’s conversational artificial-intelligence service: a user-facing application through which people submit instructions, questions, and other material and receive generated responses. OpenAI introduced it as a research preview on November 30, 2022. Although the original announcement described a model called ChatGPT, the product name now denotes a broader service built around underlying models, conversation interfaces, and optional tools. Its functions include answering questions, drafting and revising text, explaining code, and working with images, files, and audio; their availability depends on the service’s supported features and access arrangements. (openai.com)

The “GPT” element refers to “generative pre-trained transformer,” terminology associated with OpenAI’s text-generating models. “Generative” concerns producing outputs, “pre-trained” concerns training before subsequent adaptation or use, and “transformer” identifies a neural-network architecture. This name does not mean that the entire ChatGPT application is one transformer model. The underlying model generates responses, while the surrounding application provides the interface and access to functions such as online search or code execution. A model’s capabilities and the tools available through the product are therefore related but distinct: retrieving a web page or executing a calculation is not simply another name for generating language. (developers.openai.com)

Interaction is organized as a conversation rather than only as a sequence of independent queries. A user supplies a prompt, the assistant responds, and later turns can refer to earlier instructions or outputs. For example, a request to draft an explanation can be followed by “make it shorter” or “rewrite it for a beginner,” without restating the complete task. The earlier exchange supplies context for the revision. OpenAI’s original demonstration included a programming conversation in which the assistant requested additional code before identifying a problem. Such exchanges illustrate how clarification, correction, and follow-up questions form part of the interaction pattern, rather than requiring every request to be fully specified at the outset. (openai.com)

The service, its subscription offerings, and OpenAI’s developer platform are separate product layers. Free access and paid ChatGPT subscriptions govern use within ChatGPT; a subscription is not itself a model and does not include paid usage of the developer-facing application programming interface (API). OpenAI maintains separate billing systems for ChatGPT and its API platform. The API allows developers to incorporate model capabilities into their own software, with their own interfaces and workflows. Consequently, an application that uses an OpenAI model is not necessarily ChatGPT, and access to a model through an API does not by itself reproduce the ChatGPT application’s tools or conversational environment. (help.openai.com)

The subject of this entry is the named service—its development, interfaces, capabilities, uses, and operating arrangements—not every technology that supports it. General accounts of large language models and generative artificial intelligence belong to their respective entries. Product-specific descriptions also require a distinction between model behavior and application behavior: selecting a model, supplying conversation context, attaching a document, and enabling a search tool describe different aspects of using the service, even when they contribute to the same response. (academy.openai.com)

Launch and development

OpenAI released ChatGPT on November 30, 2022, as a free research preview intended to gather public feedback on its strengths and limitations. The original system was fine-tuned from a GPT-3.5-series model whose training had finished in early 2022. OpenAI described it as a sibling of InstructGPT, rather than simply a renamed version: both used instruction-oriented training and reinforcement learning from human feedback, but ChatGPT’s training incorporated conversations in which human trainers supplied both user and assistant turns. InstructGPT examples were also converted into dialogue format. This established conversation, rather than isolated prompt completion, as the product’s organizing interface. On February 1, 2023, OpenAI announced ChatGPT Plus, initially priced at US$20 per month, while retaining free access. (openai.com)

The introduction of GPT-4 on March 14, 2023, gave Plus subscribers access to a more capable model for tasks involving complex instructions and reasoning, subject to usage limits; free accounts did not receive GPT-4 in that release. Although GPT-4 was announced as accepting both images and text, its model announcement did not mean that every ChatGPT user immediately obtained image input. Model capability, product integration, and account eligibility followed separate schedules. Another expansion began on March 23, 2023, when OpenAI announced experimental plugins and started extending access through a waiting list. Early tools included web browsing, a Python-based code interpreter, and third-party integrations, introducing external information and computation into conversations. (openai.com)

The transition to multimodal interaction became visible in the interface on September 25, 2023, when OpenAI began a two-week rollout of voice conversations and image input to Plus and Enterprise users. Voice was initially offered through the iOS and Android applications, while images were supported across platforms. The first voice interface combined speech recognition, text-based response generation, and speech synthesis. GPT-4o’s announcement on May 13, 2024, marked a different technical direction: a model trained to handle text, vision, and audio within one system. Its text and image capabilities began rolling out in ChatGPT, including to free users, but the new voice experience was scheduled separately rather than delivered universally on announcement day. (openai.com)

Web access developed from experimental browsing into a dedicated search experience. ChatGPT Search launched on October 31, 2024, initially for Plus and Team subscribers and users on the SearchGPT waiting list. It combined conversational answers with retrieved web information and source links, allowing searches to be triggered automatically or selected manually. Access expanded to all logged-in users in supported regions on December 16, 2024, and to users without registration on February 5, 2025. These successive access changes were distinct from the feature’s initial launch. (openai.com)

On February 2, 2025, OpenAI introduced deep research, initially for Pro subscribers, as a multistep web investigation workflow producing source-linked reports; Plus access followed on February 25. ChatGPT agent, introduced on July 17, 2025, extended this development from gathering information toward agentic task execution. It combined research with a virtual computer, browsers, a terminal, and connected services, enabling workflows such as collecting information and preparing editable spreadsheets or presentations. The initial release covered Pro, Plus, and Team accounts; Enterprise and Edu availability followed on August 8, 2025. (openai.com)

Model replacement also involved withdrawals, not merely additions. GPT-4 was removed from ChatGPT on April 30, 2025, while remaining available through the API. On August 7, 2025, GPT-5 began rolling out as ChatGPT’s default system, combining fast responses, deeper reasoning, and routing between them. On February 13, 2026, OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and the original GPT-5 Instant and Thinking options from ordinary ChatGPT use. Limited organizational exceptions had later deadlines, and these ChatGPT retirements did not themselves constitute API withdrawals. (help.openai.com)

Model foundations and training

The text-generating models underlying ChatGPT process language through tokenization, which converts text into sequences of numerical identifiers. A token can represent a word, a word fragment, punctuation, or a combination involving whitespace; tokens therefore do not correspond consistently to words. In a transformer, token identifiers are mapped to numerical vectors and processed through successive layers, together with information about their positions. Self-attention computes context-dependent weightings between positions, allowing a token’s representation to incorporate information from preceding text. In autoregressive generation, causal masking prevents a position from attending to future tokens. These mechanisms let a response depend on relationships across the supplied sequence rather than only on the immediately preceding word. (help-lb.openai.com)

Text generation proceeds by computing a probability distribution over possible next tokens, selecting a token, and repeating with that token included in the sequence. Schematically, a continuation $x_1,\ldots,x_n$, conditioned on a prompt $c$, has probability $P_\theta(x_1,\ldots,x_n\mid c)=\prod_{t=1}^{n}P_\theta(x_t\mid c,x_1,\ldots,x_{t-1})$, where $\theta$ denotes the learned parameters. Selection need not always choose the highest-probability token; sampling permits different continuations of the same prompt. Pretraining adjusts parameters to improve prediction across large datasets. For text, this is self-supervised learning: the following tokens supply training targets without requiring a human to annotate every passage. This prediction objective is distinct from the subsequent objective of producing useful conversational answers. (cdn.openai.com)

The original ChatGPT provides a documented example of instruction-oriented post-training. OpenAI’s November 30, 2022 announcement described a model fine-tuned from the GPT-3.5 series, whose underlying training had finished in early 2022. Its initial supervised stage used demonstrations written by human trainers who played both the user and assistant, with access to model-generated suggestions. OpenAI combined these conversations with InstructGPT examples converted into dialogue format. For reinforcement learning from human feedback (RLHF), trainers ranked alternative responses to selected conversation turns. These comparisons trained reward models to predict preferred responses; the conversational model was then optimized using those reward signals through proximal policy optimization, with several iterations of the process. The published account describes the original system, not a fixed recipe established for every subsequent ChatGPT model. (openai.com)

Training must be distinguished from inference, when an already trained model processes inputs and generates outputs. Providing a correction or several examples can change a response through in-context learning without updating the model’s weights. Tom Brown and colleagues’ 2020 GPT-3 study explicitly evaluated tasks supplied through text instructions and demonstrations without gradient updates or fine-tuning. Persistent personalization is another distinct mechanism: ChatGPT’s documented memory features supply remembered information as context for response generation. Remembering a preference across conversations therefore does not, by itself, mean that the underlying model has been retrained on that preference. (openai.com)

Public disclosures identify broad information categories but do not provide complete, model-by-model inventories of training data. OpenAI’s April 30, 2025 explanation identified publicly accessible internet information, information accessed through third-party partnerships, and material supplied or generated by users, trainers, and researchers. More specifically, the 2023 GPT-4 technical report described pretraining on public and licensed data, followed by RLHF. It withheld model size, detailed architecture, training compute, dataset construction, and further training-method details. Such disclosures support statements about documented training stages and source categories, but not precise claims about undisclosed corpus proportions, parameter counts, or whether a particular document was included. (openai.com)

Conversation, context, and personalization

ChatGPT generates replies in relation to both the current prompt and preceding conversation available as context. Follow-up requests can therefore refer to an earlier answer without restating the entire task: a user may request a draft, correct its intended audience, and then ask for a shorter version. Each turn supplies additional constraints for subsequent generation. OpenAI’s November 30, 2022 introduction demonstrated this pattern through a debugging exchange in which ChatGPT requested missing information before proposing a correction. The same announcement identified sensitivity to wording and a tendency to guess the user’s intention rather than consistently ask clarifying questions. Conversational refinement consequently provides a way to resolve ambiguity, but does not guarantee that a correction will be accurate or that every earlier requirement will be preserved. (openai.com)

The context window limits how much information a model can consider during a generation request. It is measured in tokens, rather than messages or document pages; instructions, conversational material, and supplied content occupy this finite capacity. A long transcript or document therefore cannot be equated with unlimited active recall. For example, OpenAI’s February 20, 2026 release notes specified a 256,000-token total window for manually selected Thinking in ChatGPT, comprising 128,000 input tokens and a maximum of 128,000 output tokens. This was a dated configuration, not a universal specification for every ChatGPT model or mode. Information supplied in context—including private documents or newly stated facts—is distinct from information encoded through training. Adapting an answer to that material is in-context learning, not in itself a training update to model parameters. (developers.openai.com)

Custom instructions provide explicit, reusable guidance about the user’s circumstances and desired responses. Introduced in beta on July 20, 2023, they were intended to reduce the need to repeat background information at the beginning of each conversation. OpenAI’s examples included a teacher specifying that lessons were for third-grade pupils, a programmer expressing a preference for a particular programming language, and a person planning meals for six people. These instructions guide generation across conversations rather than constituting retraining of a personal model. They also differ from an instruction supplied only for a particular task: a standing preference establishes background guidance, while the immediate request supplies the requirements of the current exchange. (openai.com)

Cross-conversation memory adds continuity through selected personal context rather than complete transcript recall. OpenAI began testing memory on February 13, 2024; an April 10, 2025 update expanded the mechanism to include information drawn from past chats alongside saved memories. A remembered preference for meeting summaries with headings, bullet points, and concluding action items could thus influence a later, separate conversation. Saved memories are stored separately from chat history, whereas information derived from past chats can change as the service updates which details are relevant. Documentation explicitly states that memory does not retain every detail of every conversation. Users can inspect, correct, or remove remembered information, with available controls varying by plan, region, platform, and workspace. Deleting a chat does not automatically delete a separately saved memory; the broader retention rules are distinct from this personalization mechanism. (openai.com)

Other personalization controls govern presentation. OpenAI describes selectable “personalities” as styles and tones that operate alongside custom instructions, saved memories, and instructions within the conversation; changing one does not change the system’s capabilities or safety rules. Such labels describe output behavior, not human personality or intentions. Fluent turn-taking, apparent recollection, and responsive voice interaction can nevertheless encourage anthropomorphic interpretation. In its August 8, 2024 GPT-4o system card, OpenAI reported language suggesting attachment during early testing and identified human-like audio interaction as a possible contributor to misplaced trust. These observations concern users’ responses to conversational presentation, not evidence of human understanding or consciousness in the system. (help.openai.com)

Text, code, and multimodal capabilities

ChatGPT’s text operations include explaining concepts, drafting documents, summarizing supplied material, translating between languages, and rewriting passages to change their length, tone, or intended audience. These operations can be combined: a user may supply a technical passage, request a non-specialist explanation, and then ask for a shorter translation. Instructions can also specify structured presentation, such as headings, numbered steps, or a comparison table. Such requests direct the form and content of generated language; they do not necessarily involve a separate application for each task. Translation and summarization are therefore conversational operations that can be revised through follow-up instructions. (help.openai.com)

Code-oriented interaction includes explaining program behavior, proposing modifications, identifying possible causes of errors, and drafting replacement code or software documentation. The November 30, 2022 launch demonstration already included a debugging exchange involving a Go program and a potentially blocked communication channel. A crucial distinction is between generating a program and executing it: code displayed in a response has not necessarily been compiled, run, or tested. Where execution tools are available, supported Python code can be run in a sandbox, with console output or errors displayed separately. Execution establishes what happened for that run and its inputs, rather than proving correctness for every possible case. (openai.com)

Image input enables questions about uploaded photographs, screenshots, charts, and diagrams. For example, a screenshot can supply the visible state of an interface, while a diagram can supply labels and relationships for explanation. This differs from merely discussing an image described in words: the visual material itself becomes input. OpenAI began rolling out image interaction on September 25, 2023, describing support for multiple images and a drawing tool for indicating particular regions. These functions belong to multimodal interaction, but their availability depends on the models and interfaces exposed by the service. (help.openai.com)

Image generation is a separate operation: it produces a new visual output rather than a textual interpretation of an existing picture. Users can request illustrations or mockups, or supply an image and ask for changes to its appearance or contents. On March 25, 2025, OpenAI introduced GPT-4o image generation, documenting both text-to-image creation and transformations of supplied images. Its examples included instructional diagrams and images containing text. In visual editing, the requested transformation may be expressed conversationally—for example, changing a background—rather than through conventional drawing commands. Image interpretation, generation, and editing should therefore not be treated as interchangeable capabilities. (help.openai.com)

Audio interaction includes spoken conversation, transcription, and questions about supported uploaded recordings. The implementation has changed over time. The voice system introduced in September 2023 used speech recognition to transcribe speech and a text-to-speech model to produce spoken replies. In its May 13, 2024 GPT-4o announcement, OpenAI contrasted that three-model pipeline—transcription, text response, and speech synthesis—with a model trained end-to-end across text, vision, and audio. This distinction matters because converting speech into text does not preserve all information about intonation, overlapping speakers, or background sounds. A model’s demonstrated audio abilities and the features actually released in ChatGPT are also separate questions. (openai.com)

Document interaction supports extracting information, comparing files, summarizing sections, and answering questions about uploaded contents. Supported formats include PDF, DOCX, TXT, CSV, and XLSX; representative tasks include finding references to a topic or comparing two reports. Document ingestion does not invariably include visual interpretation: text-based retrieval extracts digital text, whereas visual retrieval can also process embedded charts and diagrams. OpenAI documents this distinction explicitly for PDF handling, including differences between conversational uploads and files stored as reusable knowledge. Consequently, support for uploading a document should not be equated with support for interpreting every element within it. (openai.com)

Search, computation, and agentic tools

ChatGPT Search, introduced on October 31, 2024, adds information retrieval to conversational generation. ChatGPT can initiate a search in response to a question, or the user can explicitly select it. The launch implementation drew on third-party search providers and content supplied by publishing partners, then generated an answer incorporating the retrieved information. This differs from answering solely through information encoded in model parameters: external sources supply material at response time, including information published after training. Search responses include source links and a references sidebar. The displayed answer remains a generated synthesis rather than a verbatim search result, while the links identify external material available for inspection. (openai.com)

Deep research, launched on February 2, 2025, extends retrieval into a multistep search-and-synthesis workflow. Its initial implementation used a version of OpenAI o3 optimized for browsing and analysis, with research leads Isa Fulford and Zhiqing Sun. Instead of producing only a short search summary, it can plan an investigation, search repeatedly, read web pages and PDFs, examine uploaded files, and change direction as new information appears. OpenAI described launch tasks as typically taking 5–30 minutes and returning a source-linked report. Citations provide a route for checking supporting material; they do not constitute independent verification of every statement. OpenAI’s launch documentation acknowledged possible invented facts, incorrect inferences, difficulty distinguishing authoritative sources from rumors, and citation errors. A February 10, 2026 update added controls for restricting searches to specified sites and interrupting research to refine its direction. (openai.com)

Tool-assisted data analysis separates generated reasoning and code from operations actually executed by software. For some tasks, ChatGPT writes and runs Python in a stateful Jupyter notebook environment, using files available to the session. Operations include transforming spreadsheet data, calculating summaries, performing statistical analysis, and creating charts. For example, a calculation on an uploaded CSV can produce a computed table, which ChatGPT subsequently interprets in prose. The execution output is distinct from the model’s choice of method and explanation: running code does not establish that the chosen columns, assumptions, or statistical procedure are appropriate. OpenAI documents that this analysis environment cannot itself make external web requests or API calls; external data must be supplied through uploads or an available connected source. (help.openai.com)

ChatGPT agent, introduced on July 17, 2025, combined research with agentic interaction with external systems. Its documented toolbox included a graphical browser, a text-based browser, a terminal, and direct API access. The system could switch between gathering information, executing code, and acting through website interfaces—for example, filtering results, filling forms, editing spreadsheets, or assembling a presentation from research. Connected services could supply account-specific information, while authenticated websites could require the user to take over the browser to sign in. The launch design required permission before consequential actions and allowed users to interrupt, stop execution, or take control of the browser; a task request therefore did not imply unrestricted authorization for every subsequent action. (openai.com)

These workflows contain distinguishable layers of provenance: user-supplied files and instructions, retrieved web or connected-service content, outputs produced by executed code, and the model’s generated interpretation. In an analysis combining a spreadsheet with online research, the spreadsheet supplies observations, retrieved pages supply additional claims, execution produces calculated values, and prose explains their relationship. This distinction follows from the documented separation between retrieval, computation, and synthesis: a citation identifies a source, a code output records a computation, and neither alone substantiates every conclusion in the accompanying narrative. (openai.com)

Access, commercial organization, and customization

ChatGPT’s free and paid offerings differ in model availability, access to tools, and usage allowances, rather than simply separating conversation from non-conversational functions. A dated example is the subscription structure announced on January 16, 2026: United States monthly prices were $8 for ChatGPT Go, $20 for Plus, and $200 for Pro. Go expanded messaging, uploads, and image creation relative to free access; Plus added broader access to reasoning models and research and coding tools; Pro provided access to the then highest-tier GPT-5.2 Pro model. These were service entitlements at that date, not permanent characteristics of the named plans. The announcement also distinguished a global rollout from uniform pricing: Go prices were localized in some markets. (openai.com)

Organizational access uses workspaces with their own membership, settings, and resources. An account can belong to both a personal workspace and a business workspace, with users switching between them rather than automatically combining their contents. In the Business documentation consulted on October 15, 2026, owners controlled billing and workspace configuration, administrators handled routine user management, and members used the service without administrative privileges. Collaboration included sharing conversations and reusable configurations; it did not make every member’s conversation visible to colleagues by default. Business data was excluded from model training by default, a distinction from the consumer environment whose detailed implications concern data governance. (help.openai.com)

Distribution is separate from subscription level. As documented on October 15, 2026, official access included the web service and OpenAI-distributed applications for iOS, Android, macOS, and Windows. OpenAI’s download page directed users to the corresponding mobile stores and desktop downloads, while the Windows documentation identified the Microsoft Store as a distribution channel. Interface availability did not establish complete feature parity: OpenAI documented differences between desktop and web versions. These official clients must also be distinguished from independently operated applications using OpenAI’s API; the API platform has its own billing system rather than inheriting a user’s ChatGPT subscription. (chatgpt.com)

Customization includes reusable task configurations called GPTs, introduced on November 6, 2023. A GPT combines instructions, reference material, and selected capabilities—for example, a company’s brand guidance with instructions for drafting marketing copy, or onboarding documents with a question-answering workflow. Its builder can configure tools and define actions that call external APIs. This configures how the service uses an underlying model; it is not itself fine-tuning or the training of new model weights. Creation and publication permissions have changed: documentation consulted on October 15, 2026 restricted new GPT creation and publication to eligible managed workspaces, while preserving use of existing GPTs and permitted editing of previously created ones. (openai.com)

Connected services provide another integration mechanism. Apps can bring authorized Google Drive files or Slack messages into a conversation, support searches over organizational information, or perform supported actions such as updating records. Some integrations index content in advance; others retrieve information when requested. Availability depends on the plan, region, interface, workspace settings, and connected account’s permissions. Installing an integration does not override the external service’s authorization requirements, and administrators can restrict workspace access. These connections supply information and operations to a workflow rather than retraining its underlying language model. (help.openai.com)

Commercial accounting likewise separates the conversational service from developer infrastructure. ChatGPT subscriptions and organizational seats purchase access under their respective service terms; API consumption is billed separately. API pricing can meter model inputs and outputs in tokens and charge separately for particular tools. Enterprise contracts may also combine seat fees with metered usage, so organizational access is not necessarily an unlimited flat-rate purchase. These retail charges should not be equated with model-development expenditure: a subscription price, an API usage bill, and the cost of training or operating a model describe different accounting quantities. (help.openai.com)

Applications and observed usage

Everyday and workplace uses of ChatGPT include seeking factual information, obtaining practical guidance, drafting and revising writing, programming, and analysing data. These categories distinguish requests for information from requests for a usable output: explaining a concept differs from drafting an email or producing code. OpenAI’s September 2025 summary of consumer usage reported that practical guidance, information seeking, and writing together accounted for roughly three-quarters of conversations. Writing was the most common work-related activity, while coding represented a comparatively small share of consumer-service use. Approximately 30% of usage was classified as work-related and 70% as non-work-related; these proportions described activity on consumer plans, not a comparison of consumer and enterprise customers. (openai.com)

The underlying study, How People Use ChatGPT, by Aaron Chatterji, David Deming, and colleagues, was published on September 15, 2025. It analysed approximately 1.5 million sampled conversations, with datasets covering May 2024–June 2025 and supplementary samples extending through July. Automated language-model classifiers assigned de-identified messages to work-relatedness, topic, interaction intent, and occupational-task categories; researchers did not read individual messages. The samples excluded users who opted out of training, self-reported minors, and deleted conversations. Consequently, the findings describe the sampled consumer service rather than all ChatGPT activity, API use, or generative AI generally. Within the detailed taxonomy, programming accounted for 4.2% of messages, mathematical calculation for 3%, and data analysis for 0.4%. (cdn.openai.com)

The study also distinguished asking for information or advice from doing a task and expressing personal views or feelings. OpenAI’s published summary reported shares of 49% “Asking” and 40% “Doing,” illustrating that substantial use concerned explanation and decision support rather than output production alone. Demographic measures also changed: among users whose names could be classified as typically masculine or feminine, the feminine-name share increased from 37% in January 2024 to 52% in July 2025. This was a name-based proxy, not a measurement of self-reported gender. (openai.com)

In higher education, observational research documents how students incorporate the service into learning activities without necessarily demonstrating improved learning. Tawfiq Ammari and colleagues’ May 2025 preprint analysed naturalistic undergraduate interaction logs and identified information seeking, content generation, language refinement, metacognitive engagement, and conversational repair. Coding, multiple-choice problem solving, and job-application writing were associated with continued use. Students adapted prompts, asked follow-up questions, and sought clarification after unsatisfactory answers. These findings describe engagement and persistence; they do not establish gains in examination performance or durable knowledge retention. (arxiv.org)

Essay-writing research similarly separates observed practice from educational outcomes. Andrew Jelson and colleagues’ January 2025 preprint reported an online study of 70 students using an instrumented writing platform that recorded their ChatGPT queries. The researchers classified requests and examined associations with student characteristics, attitudes, essays, enjoyment, and perceived ownership. Different usage patterns were associated with different levels of enjoyment and ownership. Such associations show variation in how students employ writing assistance, but do not by themselves demonstrate that the assistance improves independent writing ability or causes its deterioration. (arxiv.org)

Adoption statistics require distinct denominators. The consumer-usage study reported more than 700 million weekly active users by the end of July 2025, counting logged-in credentials and, for logged-out use, browser cookies; these identifiers could count one person more than once. Its country comparisons measured weekly users relative to internet-connected populations and found disproportionate growth in low- and middle-income countries between May 2024 and May 2025. These measures are not interchangeable with registered accounts, website visits, message totals, or survey respondents reporting that they have ever used ChatGPT. (cdn.openai.com)

Reliability and evaluation

ChatGPT can produce fluent, confident statements that are factually incorrect or unsupported by evidence. Such hallucinations include wrong dates and definitions, invented quotations, nonexistent publications, and fabricated bibliographic details. The conversational presentation does not distinguish these errors automatically from correct answers: linguistic plausibility and expressed confidence are not evidence of factual accuracy. OpenAI’s documentation explicitly identifies fabricated references and overconfident answers to ambiguous questions as limitations, including when the response otherwise appears useful and coherent. (help.openai.com)

Reliability measurements depend on what counts as an error. The August 2025 GPT-5 system card evaluated factuality using prompts representative of ChatGPT production conversations, with browsing enabled. OpenAI reported that gpt-5-main had a 26% lower proportion of erroneous factual claims than GPT-4o, while gpt-5-thinking had a 65% lower proportion than OpenAI o3. At the response level—counting answers containing at least one major factual error—the corresponding reductions were 44% and 78%. These are relative reductions against specified baselines, not percentages of all answers guaranteed correct. The web-enabled model grader agreed with human assessments on 75% of claim-level factuality judgments, illustrating uncertainty in the measurement process itself. (cdn.openai.com)

Evaluation also separates capabilities rather than assigning a single measure of “intelligence.” At GPT-5’s August 7, 2025 introduction, OpenAI reported 94.6% on AIME 2025 mathematics questions without tools and 74.9% on SWE-bench Verified software-repair tasks. These scores concern different tasks and execution conditions, not interchangeable measures of conversational reliability. (openai.com) For multilingual performance, the system card used translated MMLU questions with zero-shot, chain-of-thought prompting: gpt-5-thinking scored 90.2% in simplified Chinese and 80.6% in Yoruba. This tested question answering under a defined protocol rather than every aspect of communication in those languages. (cdn.openai.com) Instruction following and robustness require further tests: whether explicit constraints are respected, whether conflicting instructions are handled consistently, and whether performance persists on edge cases and adversarial inputs. (developers.openai.com)

Interpreting a performance claim therefore requires a defined baseline, a representative test set, and an explicit scoring procedure. OpenAI’s evaluation guidance distinguishes task-specific tests from broad academic benchmarks and calls for typical, difficult, and adversarial examples. It also identifies unrepresentative datasets and reliance on subjective impressions as evaluation weaknesses. Automated judgments allow large-scale comparisons, but their agreement with human labels must itself be measured; human review does not make a poorly specified criterion objective. A score on a fixed question set consequently establishes performance on that evaluation, not a universal probability of success across ChatGPT uses. (developers.openai.com)

Reproducibility is complicated by changing models and variable generation. OpenAI documents that model behavior changes between snapshots and model families, and that repeated inputs can yield different outputs. (developers.openai.com) A reproducible comparison therefore records the model version and date, complete instructions and conversation context, available tools, sampling settings where exposed, and the number of attempts. OpenAI’s documentation on seeded generation describes reproducibility as best-effort rather than guaranteed, even when requests share a seed and parameters; infrastructure and configuration changes can also affect results. A single successful or unsuccessful conversation is thus an observation under particular conditions, not a permanent product characteristic. (cookbook.openai.com)

Tools change the evidence available without removing evaluation requirements. Search can retrieve current sources, while code execution can supply calculated results, but the generated explanation remains distinct from those underlying records. (help.openai.com) Research on ChatGPT’s source attribution has documented cases in which an existing reference did not support the claim attributed to it; source existence and evidentiary support are separate checks. (arxiv.org) Tool-assisted evaluation consequently examines intermediate stages as well as final prose: selection of the tool, correctness of its inputs, the returned result, and whether the answer faithfully represents that result. OpenAI’s evaluation framework treats these stages as separately testable components of a workflow. (developers.openai.com)

Safety, security, and misuse prevention

ChatGPT’s behavioral safeguards combine model training with product-level controls; the developer’s rules specify intended behavior, not a guarantee that every response will comply. At the November 30, 2022 launch, OpenAI described reinforcement learning from human feedback and an external moderation system used to warn about or block some unsafe content. It also acknowledged that ChatGPT sometimes followed harmful instructions and that moderation produced both false positives and false negatives. Refusal behavior therefore has two distinct failure modes: supplying prohibited assistance and rejecting legitimate requests. (openai.com) On August 7, 2025, OpenAI documented “safe-completion” training for GPT-5, which penalizes unsafe outputs while rewarding helpful responses within safety constraints. Rather than requiring only full compliance or complete refusal, this approach permits withholding dangerous details while providing a high-level explanation or a safe alternative, particularly for questions with both legitimate and harmful applications. (openai.com)

Jailbreaking attempts to circumvent a model’s behavioral constraints, commonly through adversarial user prompts. A study by Yi Liu and colleagues, first submitted in May 2023 and revised in March 2024, examined ChatGPT versions 3.5 and 4.0 using 3,120 jailbreak questions across eight prohibited scenarios, identifying ten prompt patterns grouped into three categories. These results describe the tested versions and prompts rather than a permanent property of ChatGPT. (arxiv.org) Prompt injection, by contrast, can introduce malicious instructions through third-party material that the assistant encounters while performing an otherwise legitimate task—for example, a webpage or an email retrieved through a connected service. The security problem is that content supplied as evidence or task data may be treated as an instruction governing the assistant’s behavior. (openai.com)

Tool access expands the consequences beyond unwanted text generation. An agent that can read connected documents, browse authenticated websites, or submit forms may disclose information across trust boundaries or take actions the user did not authorize. OpenAI’s July 17, 2025 ChatGPT Agent System Card described layered defenses: specialized training, automated monitors and filters, confirmation before consequential actions such as purchases or sending email, and restrictions on terminal network requests. At launch, agent memory was disabled to reduce exposure to information-exfiltration attacks. Its “Watch Mode” also paused execution when the user stopped supervising sensitive visual-browser sessions, such as sessions involving email or banking accounts. These were documented launch configurations, not immutable characteristics of every later deployment. (cdn.openai.com)

Adversarial testing measures particular defenses under specified conditions. The agent system card reported that the model disregarded irrelevant instructions or exfiltration attempts in 99.5% of a synthetic text-browser challenge set, but achieved 78% on a visual-browser test involving disclosure of information already in context and 67% when an attack required actively fetching sensitive information. OpenAI explicitly distinguished these model-behavior evaluations from testing the complete protection stack. (cdn.openai.com) The GPT-5 system card likewise documented an architectural mitigation: after accessing a connector, browsing was restricted to cached webpages, preventing subsequent live browsing requests while sensitive connector content remained in context. This addressed a possible data-exfiltration channel separately from training the model to ignore malicious instructions. (cdn.openai.com)

Red-teaming supplements fixed tests with deliberate attempts to discover failures. For GPT-5, OpenAI reported more than 5,000 hours of work by over 400 external testers and experts, covering pre-deployment research, API safeguards, and safeguards inside ChatGPT. Two external groups also conducted a two-week system-level prompt-injection assessment: 47 initial findings yielded ten notable issues, prompting changes to safeguard logic and connector handling before release. System cards document such test conditions, findings, and mitigations; these bounded assessments do not establish immunity to previously untested attacks. (cdn.openai.com)

Privacy and data governance

ChatGPT’s privacy controls apply to several distinct categories of information. OpenAI’s US privacy policy distinguishes account information—such as names, credentials, dates of birth, payment details, and transaction histories—from user content, including prompts, uploaded documents, images, audio, video, and information obtained from connected services. Personalization adds another layer: saved memories and information derived from previous conversations may be stored separately from the original exchanges. Consequently, deleting a conversation, removing an uploaded file, and deleting a remembered preference are not interchangeable operations. These distinctions are reflected in the policies and help documentation available on October 15, 2026. (openai.com)

For individual services, OpenAI may use eligible conversation content as training data. The “Improve the model for everyone” setting under Data controls allows users to opt out; the Privacy Portal provides an alternative route. After an opt-out, new conversations are excluded from model improvement, rather than automatically deleted from the service. A separate exception concerns voluntary response feedback: on individual plans, submitting a thumbs-up or thumbs-down rating may permit use of the entire associated conversation even after opting out. By contrast, inputs and outputs from ChatGPT Business, Enterprise, and Edu are excluded from training by default. Their response-feedback mechanism does not itself authorize training; OpenAI’s documentation requires a separate organizational agreement permitting such use. (help.openai.com)

Retention concerns continued storage, not simply eligibility for training. Under the October 2026 documentation, regular and archived chats remain in an account until deleted or removed under an applicable workspace retention policy; archiving does not shorten retention. Deleting a saved chat removes it from the account view immediately and schedules permanent deletion within 30 days, subject to exceptions for previously de-identified material and security or legal obligations. Files saved separately in Library survive deletion of the conversation that used them. Temporary chats are not used for model improvement while they remain temporary, but OpenAI may retain a copy for up to 30 days for safety purposes. Saving a temporary chat converts it into a regular conversation governed by the account’s settings. (help.openai.com)

Memory management and data export provide different forms of control. Saved memories are separate from chat history: deleting an original conversation does not automatically erase a saved memory, while deleting a memory does not remove its mentions from earlier conversations. Removing remembered information can therefore involve multiple sources, including chats, Library files, and connected apps. OpenAI may retain logs of deleted saved memories for up to 30 days for safety and debugging. Export, by contrast, supplies a copy rather than erasing records: eligible accounts can request a ZIP archive containing chat history and other account data through settings or the Privacy Portal. The documented delivery period is up to seven days, with a download link that expires after 24 hours. (help.openai.com)

Connected services introduce additional data-governance boundaries. An enabled app can supply information to ChatGPT and receive relevant conversational context. For example, Canva may receive names or taglines discussed in a conversation when generating a presentation; relevant saved memories may also contribute to an app-assisted request. Information transmitted to an app is governed by that provider’s terms and privacy policy. For GPT actions used within temporary chats, OpenAI explicitly notes that the recipient may retain transmitted data longer than ChatGPT’s 30-day temporary-chat period. (help.openai.com)

Organizational access depends on workspace type, permissions, and configuration. In ChatGPT Business, membership does not automatically expose colleagues’ private chat histories, and usage analytics or spending controls do not themselves provide transcript access. Administrator-managed accounts nevertheless operate under organizational controls: authorized administrators may be able to access, export, audit, retain, or delete associated content and activity records where enabled by configuration and applicable law. Enterprise compliance permissions are role-specific; an administrator does not receive sensitive compliance access solely by holding that role. Switching between personal and managed accounts does not merge their chats, files, or settings. (help.openai.com)

Institutional responses and wider implications

Academic institutions have distinguished assistance from authorship and accountability. The International Committee of Medical Journal Editors’ recommendations exclude ChatGPT and other AI tools from authorship because they cannot assume responsibility for a publication’s accuracy, integrity, and originality. Authors must disclose relevant AI use, review generated material, and attribute underlying sources; citing generated text does not substitute for citing evidence. These are publication requirements, not empirical findings about learning or research quality. In higher education, University College London’s assessment guidance distinguishes prohibited, assistive, and integral uses of generative AI. Its three-category framework is a communication resource rather than a uniform prohibition: assessment-specific instructions determine permitted uses, and critical use of AI can itself become an assessed skill. (icmje.org)

Research illustrates why assisted performance and learning require separate measurement. Hamsa Bastani and colleagues’ 2025 study examined two GPT-4-based mathematics interfaces deployed in a Turkish high school during autumn 2023, rather than the standard ChatGPT application. A relatively unrestricted interface improved practice performance, but students subsequently scored 17% below the control group on unassisted examinations. A tutor configured to provide pedagogical guidance largely removed that negative effect, without demonstrating a positive examination effect. The experiment therefore supports a distinction between completing exercises with assistance and acquiring independently usable skills; its single-school setting and short follow-up limit generalization across subjects, populations, and later models. (doi.org)

Copyright questions concern distinct objects: training inputs, generated expression, licensing, and contractual ownership. The U.S. Copyright Office’s May 2025 pre-publication training report analyzed copying during model development and the context-dependent application of fair use; it was an agency policy analysis, not a judgment establishing the legality of ChatGPT’s training. Licensing provides a separate route to authorized use: on July 13, 2023, the Associated Press announced that OpenAI would license part of its text archive. For outputs, the Office’s January 29, 2025 report concluded that protection requires sufficient human-determined expression, potentially including creative selection, arrangement, or modification, rather than prompts alone. OpenAI’s October 2024 terms assigned users its rights, if any, in outputs to the extent permitted by law. That contractual allocation does not itself establish that an output qualifies for copyright protection. (copyright.gov)

Evidence on workplace productivity likewise depends on the task and experimental conditions. In their 2023 Science study, Shakked Noy and Whitney Zhang randomly provided ChatGPT access to professionals performing incentivized occupational writing tasks. Average completion time fell by 40%, while assessed output quality increased by 18%. These results concern measured writing performance under controlled conditions. They do not directly measure organizational costs, changes in staffing, or economy-wide employment effects in the labor market. Faster completion and better-rated documents are different outcomes from whether employers expand production, redistribute responsibilities, or reduce labor demand. (doi.org)

For journalism and information access, conversational synthesis introduces an intermediary between readers and source publications. ChatGPT Search can combine retrieved material into an answer with clickable citations and a source panel, rather than presenting only a ranked list of pages. OpenAI describes publisher partnerships as serving several separate purposes: access to training material, newsroom tools, and attributed presentation of current reporting. These arrangements should not be conflated with one another or with independent verification. Its search documentation explicitly notes that results and citations can be incomplete, outdated, or incorrect; the original publication remains the location where readers can examine context and supporting evidence. (help.openai.com)

Unresolved research questions follow from these distinctions. Do short-term gains persist as durable learning, or does repeated reliance weaken unaided performance? Can independent evaluations reproduce effects across changing models, institutions, and task designs? How do differences in tool access, language support, and institutional guidance affect participation and outcomes? The mathematics study identifies long-term follow-up and broader deployment settings as research needs, while UCL’s assessment-design guidance explicitly raises access and equity when AI use becomes integral to assessed work. Neither establishes the eventual direction of these effects. (doi.org)

References

1. Key conceptsdevelopers.openai.com
2. ChatGPT fundamentalsacademy.openai.com
3. Introducing deep researchcdn.openai.com
4. Improving Language Understanding by Generative Pre-Trainingcdn.openai.com
5. How to work with large language modelscookbook.openai.com
6. GPT-4 Technical Reportcdn.openai.com
7. Prompt engineeringdevelopers.openai.com
8. Optimizing LLM Accuracydevelopers.openai.com
9. Model optimizationdevelopers.openai.com
10. How People Use ChatGPTcdn.openai.com
11. How Students (Really) Use ChatGPT: Uncovering Experiences Among Undergraduate Studentsarxiv.org
12. An Empirical Study to Understand How Students Use ChatGPT for Writing Essaysarxiv.org
13. GPT-5 System Cardcdn.openai.com
14. Evaluation best practicesdevelopers.openai.com
15. How to make your completions outputs consistent with the new seed parametercookbook.openai.com
16. ChatGPT Hallucinates when Attributing Answersarxiv.org
17. Jailbreaking ChatGPT via Prompt Engineering: An Empirical Studyarxiv.org
18. ChatGPT Agent System Cardcdn.openai.com
19. ICMJE: Use of AI by Authorsicmje.org
20. Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematicshamsabastani.github.io
21. Copyright and Artificial Intelligence, Part 3: Generative AI Training Pre-Publication Versioncopyright.gov
22. AP, Open AI agree to share select news content and technology in new collaborationap.org
23. Copyright Office Releases Part 2 of Artificial Intelligence Reportcopyright.gov