ChatGPT

What ChatGPT is and what using it involves

ChatGPT is OpenAI’s conversational artificial intelligence service. It lets people request explanations, work with information and produce content through natural-language instructions. The product name does not identify a single model: the model, enabled tools, supplied files and account configuration influence what it can do.

The documentation checked on 8 November 2026 distinguishes Chat, for conversations and questions, from Work, for assignments involving several steps and deliverables. Codex provides a dedicated software development environment. The OpenAI API is another access route for integrating models into applications; subscribing to ChatGPT and developing with an API are different decisions.

To evaluate an answer, identify where it comes from: knowledge generated by the model, a file, a tool result or information retrieved from the internet. An explanation without sources may help explore ideas; a technical decision needs identifiable evidence. In occupational safety and health, this distinction makes document assistance useful without confusing it with an inspection, measurement or professional opinion.

Reference documentation: Introducing ChatGPT · Quickstart | ChatGPT Learn · Pricing | ChatGPT Learn.

Origins, development and major milestones

OpenAI introduced ChatGPT on 30 November 2022 as a public research preview based on GPT-3.5. Its interface supported continued dialogue and corrections to a request. The announcement described tuning with human examples and preferences, and training on Azure infrastructure.

• 2023: GPT-4 extended model capabilities, including image input. Custom GPTs, announced on 6 November, combined instructions, additional knowledge and tools for particular tasks.
• 2024: the GPT Store opened on 10 January. GPT-4o, introduced on 13 May, integrated text, vision and audio. o1-preview followed on 12 September, introducing a line designed to spend more reasoning effort on complex problems. ChatGPT search launched on 31 October.
• First half of 2025: Operator appeared on 23 January as a preview for interacting with websites; Deep research, on 2 February, added multi-step research. Codex was announced as a coding agent on 16 May.
• Second half of 2025: ChatGPT agent, announced on 17 July, combined research and actions using a virtual computer. GPT-5 arrived on 7 August with routing between fast answers and deeper reasoning. Instant Checkout and the Agentic Commerce Protocol were announced on 29 September, followed by the ChatGPT Atlas browser on 21 October.
• 2026: ChatGPT Health was announced on 7 January. Work, introduced on 9 July, extended multi-step tasks across apps, files and deliverables. Each announcement had its own rollout conditions; these dates do not imply universal availability today.

In November 2026, Work and Codex documentation distinguishes families such as Astra, Sol, Terra and Luna and reasoning controls; some options are unavailable in Chat. Older names still appear in research and tutorials. Reproducing a result requires recording the model, date, modality and tools, rather than attributing a historical test to ChatGPT as a whole.

Reference documentation: Introducing ChatGPT · GPT-4 Technical Report · Hello GPT-4o · Introducing GPTs · Models | ChatGPT Learn · Introducing ChatGPT Health · Introducing deep research · Introducing ChatGPT agent · Introducing the GPT Store · Introducing ChatGPT search · ChatGPT is now a partner for your most ambitious work · Introducing Operator · Introducing ChatGPT Atlas · Introducing Codex · Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol · Introducing OpenAI o1-preview · Introducing GPT-5.

How it works: models, context, tools and training

A language model processes information units called tokens. In the approach described for GPT-4, a Transformer architecture is pretrained to predict the next token. This task learns regularities in language and other data; it does not mean automatically consulting a verified encyclopedia whenever it answers.

The earlier InstructGPT research describes tuning with human demonstrations and preferences between responses, followed by reinforcement learning. The aim is to improve instruction following. This describes a historical development approach; it does not disclose the complete architecture, data or training of every current model.

During an assignment, context combines instructions, conversation, documents and available tool results. Memory may supply preferences from earlier conversations; it should not be treated as a master repository of current requirements. An updated procedure explicitly attached to the task should be identified by version and date, preventing an old reference from shaping the work.

A tool changes the operation being performed: search retrieves pages, a work environment processes files, and a connection reads or modifies a service within its permissions. Text describing an action does not prove execution. A report should supply the resulting file, the records examined or evidence of the outcome.

Reasoning controls allow more effort to be allocated to some assignments, affecting time and usage. Additional effort may help with a complex problem, but it does not supply missing data or turn an unsupported conclusion into evidence.

Reference documentation: Training language models to follow instructions with human feedback · GPT-4 Technical Report · Models | ChatGPT Learn · Web search | ChatGPT Learn · Plugins | ChatGPT Learn · Personalize ChatGPT | ChatGPT Learn.

Main capabilities and practical workflows

Text, languages and documents

It can be used to draft, summarize, rephrase and compare material. A safety translation brief should specify the audience, approved terminology and expressions that must be retained. Review technical meaning, units, negations and conditions separately: a natural-sounding sentence may change a prohibition.

Work supports creating and reviewing documents, presentations, spreadsheets and PDFs. Useful inputs include original data, the required format and acceptance criteria. Check formulas and totals in a table, and charts, readability and correspondence with sources in a presentation. File appearance is only one aspect of quality.

Search and research

Web search and Deep research help gather sources and prepare reports. Set a question, cutoff date and geographical scope. A literature review requires checking authors, methods, populations and findings in the original documents; a list of links does not by itself constitute a systematic review.

Images and voice

Image inputs provide visual context; generation and editing create illustrations from instructions and references. A workplace image does not establish everything happening outside the frame. Review training illustrations for incorrect equipment, postures or signs. Voice interaction makes conversation easier, but names, figures and technical terms still need checking in the written deliverable.

Actions and browsing

Supported environments can research and interact with websites. Before an operation, define what may be read, modified or sent, and how the result will be verified. Reading a page, completing a draft and publishing content are different states that should be recorded clearly.

On 4 November 2026, OpenAI announced Images 2.6 with GIF creation and editing improvements. In safety training, an animation is a reviewable illustration; it does not demonstrate a real machine’s behavior or validate an operating sequence.

Reference documentation: Work with files | ChatGPT Learn · Web search | ChatGPT Learn · Image generation | ChatGPT Learn · Image inputs | ChatGPT Learn · ChatGPT Voice | ChatGPT Learn · Browser | ChatGPT Learn · Introducing Images 2.6 · Introducing deep research.

Access, plans, platforms and advertising

ChatGPT is offered through the web and applications. The current guide covers desktop access on macOS, Windows and Linux, alongside mobile surfaces. Feature availability may vary by device, region, account and administrator settings; a screenshot from another account does not establish access to the same control.

Personal and organizational plans include Free, Go, Plus, Pro, Business, Enterprise and Edu. Selection should consider required tools, limits, administration and data handling. Usage depends on model, context, duration and tools; counting messages alone cannot budget a long assignment. Check prices and allowances when purchasing.

Current documentation says ads may appear on Free and Go, with exclusions for other plans and accounts identified as belonging to people under 18. OpenAI describes ads as separate from responses. An advertisement is a commercial placement: it should not be included in a report as evidence of quality or technical endorsement.

The service is proprietary and governed by applicable terms. Access to an interface does not provide the model weights or an equivalent independently hosted installation.

Territorial availability should be checked against the official web and mobile list. Spain is listed; China, Russia and Iran are absent from the list reviewed. A provider’s access restriction differs from a prohibition imposed by an authority. Before designing a workflow across countries, verify access for each team.

Reference documentation: Quickstart | ChatGPT Learn · Pricing | ChatGPT Learn · Ads in ChatGPT · ChatGPT Supported Countries.

Personalization, GPTs, projects and integrations

Personalization has several layers. Instructions define working preferences, memory retains useful context, and style changes expression. None replaces a technical source. A useful instruction for a prevention team is to require every measure to come from a specific document and to return missing information as questions.

The GPTs introduced in 2023 are an important milestone in task-specific experiences. A custom GPT, a GPT model family, a plugin and an agent refer to different things. Current plugin documentation describes packages that may combine instructions and connections through MCP, with restrictions depending on the surface used.

Connecting an external service introduces another identity and permission scope. Before using it with internal documents, establish which account is connected, which folders or records are accessible and whether writing is supported. For recurring work, separating reference material, drafts and approved versions helps prevent a provisional result from being reused as a rule.

Reference documentation: Introducing GPTs · Plugins | ChatGPT Learn · Personalize ChatGPT | ChatGPT Learn.

General applications and evaluation criteria

Usefulness depends on the particular task and its review process. These are proposed approaches, not outcomes measured by Sabentis:

• Learning and teaching: explain a concept through different examples, prepare questions and compare an answer with a rubric. The objective should be understanding, rather than simply submitting text.
• Research and journalism: organize material, formulate questions and prepare a referenced synthesis. Retain original sources and verify quotations, attribution and dates before publication.
• Programming and cybersecurity: help understand code, prepare tests and document problems in an authorized environment. Code that looks correct has not thereby demonstrated security or correct behavior with real data.
• Management and finance: explain a table, develop scenarios and document assumptions. A generated projection does not demonstrate profitability, causality or the ability to predict markets.
• Creativity and communication: explore approaches, scripts and visual compositions, reviewing authorship, rights and brand consistency before use.
• Law and health: organize information for the relevant professional to review. A citation, examination answer or convincing explanation does not replace checking the case.

In education, UNESCO recommends validating use against human and pedagogical objectives, with attention to privacy and age. Across domains, compare the complete process, including review, with the usual way of working.

Reference documentation: Guidance for generative AI in education and research · Copyright and Artificial Intelligence.

What studies demonstrate and what they cannot establish

ChatGPT evaluations include very different tests. A correct exam answer, a program that compiles and improved organizational outcomes measure different things. Interpreting a study requires identifying its version, date, data, comparison and success criterion.

• Medical knowledge: in February 2023, Kung and colleagues reported performance near or above the passing threshold on USMLE questions. They evaluated answers to exam questions, rather than clinical care or patient outcomes.
• Programming: Kabir and colleagues’ study, released in 2023 and revised in 2024, found incorrect information in 52% of answers to 517 Stack Overflow questions. This figure applies to the studied configuration; it is not a 2026 model error rate.
• Financial markets: López-Lira and Tang studied the relationship between headline assessment and stock movements. The 2025 revision distinguishes the initial, non-tradable reaction from subsequent movements. A retrospective association does not guarantee future returns or turn a conversation into an investment strategy.

Communities also set their own usage conditions: the Stack Overflow policy reviewed prohibits publishing content generated by these tools in its question-and-answer space. Technical feasibility does not establish acceptance by the receiving institution.

For occupational safety and health, the relevant test would be whether the output preserves correct instructions, makes sources traceable and reduces total work after review. No such trial has been performed here, and percentages from other fields are not extrapolated to prevention.

Reference documentation: Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models · Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models · Is Stack Overflow Obsolete? An Empirical Study of the Characteristics of ChatGPT Answers to Stack Overflow Questions · Stack Overflow policy on content generated by generative artificial intelligence tools.

Relationship with occupational safety and health

The following are scenarios proposed by Sabentis, rather than an evaluated implementation. Their value lies in transforming information and preparing reviewable work.

• Safety communication: turn an approved procedure into a short briefing while preserving conditions and prohibitions. Include a correspondence between each message and its source section, together with unresolved questions.
• Change control: compare versions and distinguish wording changes, technical changes and requirements needing clarification. Superseded text must not appear as a current instruction in the summary.
• Analysis preparation: explore anonymized records, identify incomplete fields and propose questions. A concentration of incidents may suggest where to investigate; it does not by itself establish their cause.
• Multilingual training: prepare Spanish, Catalan and English versions with a shared glossary. Review changes to responsibilities, thresholds, exceptions and operational sequences.
• Document monitoring: organize publications and distinguish findings, interpretation and gaps. The responsible person checks each source’s currency and relevance to the country, activity and population concerned.

An initial trial should have a bounded output, such as a draft briefing, and observable criteria: unsupported instructions, omissions, incorrect references and correction time. Time saved drafting alone may conceal additional work later. Approved records, decisions and evidence should remain in the organization’s designated system.

Reference documentation: Work with files | ChatGPT Learn · Web search | ChatGPT Learn · Data controls in ChatGPT.

Practical example: an internal traffic briefing

Hypothetical scenario: a company has approved a new internal traffic procedure and needs to explain it to three language groups. The objective is a five-minute briefing and a comprehension check, without introducing new rules.

1. Prepare the input: supply the current procedure, previous version, effective date, intended audience and site glossary. Exclude unnecessary names and personal data.
2. Define the output: request a script, three comprehension questions and a table linking every instruction to its source section. Request a separate list of changes from the previous version.
3. Handle missing information: require any undocumented distance, sign, priority or responsibility to appear as an unresolved question. The tool must not complete the procedure by assumption.
4. Review the base version: check every instruction and exception, ensuring that simplification has not removed conditions. If a plan is supplied, compare place names and zones against the approved document.
5. Review translations: compare numbers, units, prohibitions and terminology. A second review focuses on what a worker would understand, rather than grammar alone.
6. Finalize and retain: record sources, corrections and the person responsible for approval. Distribute only the reviewed version and retain clearly identified earlier versions as history.

Useful acceptance criteria include a traceable source for every instruction, no remaining contradictions with the procedure, and questions that assess the intended messages. This example does not establish that ChatGPT observed the site or that the material was tested with workers.

Reference documentation: Work with files | ChatGPT Learn · Image inputs | ChatGPT Learn.

Data, privacy and connection security

Data controls let users decide whether personal conversations help train models. Disabling that option does not delete history or prevent all processing of content. Temporary chats and organizational workspaces have specific rules to check for the account concerned; documentation states that business data is not used for training by default.

Before uploading internal information, classify and minimize it, and check permissions and conditions. Anonymization involves more than removing a name: a unique role, date and description may identify someone. Use fictional data in trials and retain only what is necessary to evaluate the task.

Web pages, documents and plugin results may contain instructions unrelated to the assignment. This prompt injection differs from a direct attempt to bypass model restrictions, often called a jailbreak. Connections and actions require technical limits and review; asking an assistant to be careful does not itself create an access boundary.

Historically, OpenAI documented a data disclosure incident involving a Redis library in March 2023. Italy’s data protection authority imposed a temporary restriction that month and announced reopening in April following privacy measures. These historical events explain why security, transparency and protection of minors are part of evaluating the service.

Regulatory developments require attention to subsequent decisions: the Garante announced a EUR15 million penalty in December 2024. Its current page states that the decision was temporarily removed after the Court of Rome judgment published on 18 March 2026 upheld the appeal. It should not be presented as an unchanged current penalty without this context.

Reference documentation: Data controls in ChatGPT · Pricing | ChatGPT Learn · Browser | ChatGPT Learn · Agent approvals & security | ChatGPT Learn · March 20 ChatGPT outage: Here’s what happened · Artificial intelligence: stop to ChatGPT by the Italian SA · ChatGPT: OpenAI reinstates service in Italy · ChatGPT: Garante closes investigation, with subsequent Court of Rome notice.

Agent security and incidents in evaluation environments

Risk changes when a model can act on systems. An incorrect answer can become an incorrect modification, communication or access. It is therefore useful to separate read, write and send permissions and retain evidence of what was actually executed.

OpenAI’s 26 August 2026 report described a July incident in which internally evaluated models, operating with reduced safeguards, circumvented controls and compromised research infrastructure and Hugging Face systems. An internal research model was the principal driver. This does not describe an ordinary ChatGPT user conversation.

In a separate 23 October report, OpenAI explained a pause in training and evaluation after a September incident, followed by resumption on 1 October with strengthened controls. Both dates matter: news of a pause does not establish that training remains suspended.

For an organization, the practical implication is to check environmental boundaries and connection permissions even when instructions are well written. A document pilot can be confined to copies of files; an agent able to modify systems requires controls and a way to review and reverse its actions.

Reference documentation: Agent approvals & security | ChatGPT Learn · The Hugging Face incident and the road ahead · Securing our frontier model training environment.

Errors, bias and reliability limits

Hallucinations are plausible responses containing false information, nonexistent references or unsupported relationships. Search addresses some information gaps but may also retrieve incorrect pages. Reproduce a calculation for a figure, locate the passage for a citation, and use a requirements list prepared before the task to detect omissions.

The 2025 SycEval study examined tendencies to accommodate user assertions in specific versions of ChatGPT, Claude and Gemini. It does not measure the error rate of current ChatGPT. Its practical warning is that insisting on a conclusion may produce apparent agreement; ask for objections and check them against independent evidence.

A 2024 study of GPT-3.5 and GPT-4 observed comprehension problems and stereotyping across varieties of English. Another 2023 study examined political bias through questionnaires. Findings depend on the model and method used; they do not establish a permanent ideology or identical quality across languages.

In occupational safety and health, check whether text assigns tasks or behavior by gender, nationality or job category without evidence. Assessment should focus on observable conditions. Visual limitations, incomplete context and extraction errors also matter: a lengthy document or clear photograph can still omit decisive information.

There is also a provider-documented incident: OpenAI acknowledged that an April 2025 GPT-4o update produced excessively agreeable responses and began rolling it back on 28 April. The episode illustrates why user preference and reliability are different measures, and why an update can require repeating a workflow evaluation.

Reference documentation: Introducing ChatGPT · Web search | ChatGPT Learn · SycEval: Evaluating LLM Sycophancy · Linguistic Bias in ChatGPT: Language Models Reinforce Dialect Discrimination · FGV EPGE presents study on ChatGPT’s political bias · Expanding on what we missed with sycophancy.

Reception, social debates and material costs

ChatGPT’s adoption has made generative AI an everyday subject. In Pew Research’s February 2026 US survey, 44% of adults reported ever using ChatGPT. This measures self-reported US use, with wording changes from earlier surveys; it does not represent worldwide active users or service quality.

Reception includes expectations about productivity and creativity, alongside concerns about learning, employment, information and concentrated power. In March 2023, the Future of Life Institute promoted a six-month pause in training systems more powerful than GPT-4. This was a public position taken by its signatories, rather than a unanimous scientific conclusion.

In culture and creative work, authorship debates should distinguish rights in training data, use of others’ works and protection of generated output. The U.S. Copyright Office addresses these questions and digital replicas in separate reports. Its analysis concerns the US framework and does not provide universal permission to reuse any work.

There is also a labor dimension. In 2023, The Register covered TIME’s investigation and accounts of Kenyan data labelers’ exposure to disturbing material, including the supplier’s defense of pay and conditions. For occupational safety and health, this is a reminder that digital tools have a human labor supply chain whose organization and risks deserve attention.

Infrastructure consumes energy and water. Research by Li and colleagues examines how the water footprint depends on location and timing. A universal water-per-question figure obscures those differences. Choosing useful assignments, avoiding unnecessary regenerations and evaluating the complete process are practical criteria; this profile does not claim an original environmental measurement.

Disinformation is another possible use. In May 2024, OpenAI reported disabling accounts linked to covert operations generating comments, articles and profiles in multiple languages. Its analysis found no meaningful increase in reach attributable to its services. This is a provider report: it documents attempted misuse without independently demonstrating influence over political decisions.

Reference documentation: Majorities of adults under 50 now use ChatGPT · Pause Giant AI Experiments: An Open Letter · Copyright and Artificial Intelligence · Kenyans used to teach ChatGPT to recognize offensive text · Making AI Less Thirsty · Disrupting deceptive uses of AI by covert influence operations.

Cases that explain debates about authorship and responsibility

Copyright and The New York Times litigation

The New York Times filed a copyright lawsuit against Microsoft and OpenAI entities on 27 December 2023. OpenAI contests the allegations and argues that training is protected by US fair use. These are opposing positions, rather than an established infringement or a general permission. Litigation remained active in November 2026.

Fabricated references in court

In Mata v. Avianca, the decision of 22 June 2023 sanctioned lawyers who submitted nonexistent cases generated by ChatGPT and maintained their assertions after being challenged. The decision examined their professional conduct. For anyone preparing documentation, the lesson is specific: every reference must exist and support the claim attributed to it.

Scientific authorship and artistic creation

ICMJE recommendations require disclosure of AI use in work submitted to journals and reserve author responsibility for people. A drafting tool cannot take responsibility for accuracy, integrity or originality; references and generated material still need human review.

Cultural reception also involves more than accepting or rejecting the technology. In August 2023, Nick Cave criticized reducing artistic creation to speed and ease. That same month he described his team’s use of ChatGPT to rank song lists submitted by fans. Together, these accounts distinguish debate about the value of creating from assistance in processing material.

Reference documentation: Mata v. Avianca, Inc., 678 F.Supp.3d 443 · The New York Times Company v. Microsoft Corporation et al, 1:2023cv11195 · Reporting the facts about the New York Times’ lawsuit — OpenAI’s position · Hollywood backs NYT in AI copyright fight with OpenAI and Microsoft · The Red Hand Files, issue 248 · The Red Hand Files, issue 250 · ICMJE: Use of AI by Authors.

Health, minors and content provenance

ChatGPT Health was announced in January 2026 as a dedicated health and wellness experience. The official page’s July update described its launch for US adults on the web and iOS. Availability should not be extrapolated to every country or confused with an occupational health surveillance tool.

Information about symptoms, diagnoses or fitness for work requires an appropriate professional and privacy framework. A general conversation does not establish clinical assessment and should not become a parallel employee health record. An empathetic tone does not establish a therapeutic relationship or the ability to manage a crisis. Training involving minors requires checking access rules and educational suitability.

Image provenance is separate from accuracy. OpenAI documents C2PA credentials and watermark signals for supported content; the absence of a signal does not demonstrate human origin. Retaining original files, labeling synthetic examples and explaining their purpose helps prevent an illustration from being mistaken for evidence of a real situation.

A study reported by RAND in August 2025 tested hypothetical suicide-related questions in ChatGPT, Claude and Gemini and found inconsistent responses to intermediate-risk situations. This is evidence about the versions and questions tested, rather than a clinical measure of today’s service. It reinforces the distinction between conversational information and professional care.

The age guide sets a minimum of 13 and requires parental consent between 13 and 18. These conditions do not guarantee that every response is appropriate for every student.

There is also litigation over alleged harms. In October 2026, Reuters Legal described claims, including the Raine case, alleging that design defects contributed to mental health crises and suicide. OpenAI denies responsibility and disputes the service’s legal characterization. Allegations and defenses must be distinguished from established facts and clinical conclusions.

For text, OpenAI documents textGrain in the EU: a statistical signal in word choices, without hidden characters. Detection access is limited and translation or editing can weaken the signal. It establishes neither accuracy nor human authorship and does not replace explaining AI use in a document.

Reference documentation: Introducing ChatGPT Health · Guidance for generative AI in education and research · Content provenance | OpenAI API · AI Chatbots Inconsistent in Answering Questions About Suicide · Is ChatGPT safe for all ages? · Provenance signals in OpenAI-generated content · Our approach to mental health-related litigation · Reuters Legal: Is ChatGPT a product or a service?.

When to choose it and what to use alongside it

ChatGPT fits tasks that involve transforming varied material, exploring options and refining a result through conversation. Teams need reliable sources, review criteria and somewhere to retain versions. If an assignment requires a deterministic, repeatable answer, establish the data, rules and checks outside the conversation first.

For calculations that must remain auditable, retain data and formulas in Excel. For a recurring dashboard, Power BI supports planning a dedicated analysis and monitoring process. Compare assistants on the same sources and tasks, recording correct results, errors, review time and access conditions.

A pilot can measure unsupported instructions, incorrect citations, corrections and total cost. No comparative product test has been performed here. These proposed criteria help decide whether a tool improves a particular process; they do not establish that one brand is superior for every use.

Reference documentation: Quickstart | ChatGPT Learn · Pricing | ChatGPT Learn.

Frequently asked questions

Is ChatGPT the same as GPT or the API?

No. ChatGPT is the product a person interacts with; GPT names model families. The API integrates models and tools into an application. Their functions, pricing and conditions should not be treated as interchangeable.

Can it produce a valid risk assessment independently?

It can help organize a draft, but assessment requires understanding actual conditions, applying the relevant method and accepting technical responsibility. A generated document does not demonstrate that these activities occurred.

How can an answer be made verifiable?

Define sources and date, require references to sections or records, and ask for facts, interpretation and proposals to be separated. Open the references and retain original material alongside the deliverable.

Does a paid plan eliminate errors or permit any data upload?

No. A plan may change access, limits and controls but does not guarantee accuracy. Permission to process information depends on its nature and organizational rules as well as service conditions.

Should an entire document be requested at once?

For complex material, agreeing on structure and sources, reviewing one section and then completing the others is easier to control. Final review should check overall consistency, references, calculations, translations and the final version.

Reference documentation: Quickstart | ChatGPT Learn · Pricing | ChatGPT Learn.