ChatGPT

ChatGPT is a conversational interface to OpenAI's generative language models, released as a research preview on 30 November 2022. It was neither the first chatbot nor the first large language model, but it was the first to pair a capable model with a simple public interface, and its reception made generative AI a mass-market subject.

The product answers questions, drafts and edits text, writes and explains code, and in later versions works with files, images, and voice. Its outputs are probabilistic: often useful, sometimes wrong, and never guaranteed to be true.

This entry covers the product's release history, the models behind it, how it works, the capabilities the company claims, the limitations that have been documented, and the legal and regulatory disputes that have followed it.

The initial release ran on a version of the GPT-3.5 series, tuned with supervised fine-tuning and reinforcement learning from human feedback so that it would answer as an assistant rather than continue text 1. Access was free and unverified, and the service reached an estimated hundred million users within two months, a figure based on third-party analysis rather than company reporting.

GPT-4 followed in March 2023 for paying subscribers, with a paid tier, a browsing mode, and plugins added during the year 2. Later releases added image understanding, voice conversation, file analysis, image generation, custom instructions, and memory. The model line has since included the GPT-4 family, the o-series reasoning models that spend more computation before answering, and the GPT-5 generation.

OpenAI has not published full technical details of its frontier models, so descriptions of their architecture remain inferential. The company publishes system cards that describe evaluations, known limitations, and safety mitigations, but not the training data or the full method. The product also acquired a memory feature that stores facts between sessions, custom instructions that persist a user's preferences, and project workspaces that group files and conversations — features that shift it from a single chat window toward a persistent workspace.

The original product used a model from the GPT-3.5 series that was trained on text and then tuned on human demonstrations and rankings. The GPT-4 announcement described a large multimodal model that accepted image and text input, and reported results on simulated professional examinations — around the top tenth of test-takers on a simulated bar exam, against the bottom tenth for GPT-3.5 2. Those figures are company-reported and measure the examination, not professional competence.

Later models added reasoning trained with reinforcement learning on verifiable problems, which improves mathematics and coding at the cost of latency. Multimodal versions accept images, audio, and documents. Version names are marketing as much as technical labels: the same underlying model may be served under several product names, with different tools, context limits, and prices.

The model is only part of the product. Retrieval, code execution, image generation, and voice are separate systems coordinated by the model, and each has its own failure modes and its own update schedule.

ChatGPT is a product layer over a model. A conversation is assembled into a prompt that includes a system instruction, earlier turns, and the new message; the model predicts a continuation; the product renders it. The model has no memory of previous sessions unless a memory feature stores a summary. OpenAI described the original model as a sibling of InstructGPT, trained on dialogue data in which people played both the user and the assistant, then ranked alternative replies 3.

Because the underlying model was trained to predict text, its knowledge is statistical and its errors are fluent. Training stops at a cutoff date, so events after that date are unknown unless the product retrieves them through browsing or a connected source.

Tools extend the product beyond text: a code interpreter executes Python, a browser retrieves pages, an image model generates pictures, a speech model transcribes and synthesises. Each tool is a separate system with its own permissions and failure modes, and the model decides when to call one. That decision is itself a prediction, and it is sometimes wrong — a model may answer from memory when it should search, or search when the question is about the conversation itself. The conversation history is also a design decision: because the model sees earlier turns, a wrong statement early in a session can shape later answers, and a correction may or may not propagate depending on how it is phrased.

The product is strongest where output can be checked quickly: drafting, summarising, rewriting, translating, explaining code, and producing structured text from unstructured input. It is weaker where correctness is not visible in the output: arithmetic without a calculator, citation of sources, statistical claims, and multi-step planning.

Hallucination remains the central limitation. The model can invent a reference, misattribute a quotation, or state a plausible falsehood in the same tone it uses for facts. OpenAI's own launch note warned that the model sometimes writes plausible-sounding but incorrect or nonsensical answers, and identified the absence of a source of truth during reinforcement learning as one reason 3. Reasoning models reduce some errors on structured problems at the cost of latency and expense, but they do not remove the underlying tendency.

Independent evaluation is complicated by the product's frequent changes, since results obtained on one version may not hold for the next, and by the fact that published benchmarks can appear in training data. The evaluations OpenAI publishes with each release are the most detailed public evidence, but the company chooses them, and independent replication is rare because the strongest models are available only through the API. Users therefore tend to rely on their own spot checks, which is a reasonable practice and a poor substitute for systematic evaluation.

OpenAI publishes usage policies that prohibit specified categories of use, including fraud, impersonation, non-consensual intimate content, academic dishonesty, and automated high-stakes decisions in domains such as employment, housing, credit, and health without human review 4. Enforcement combines automated detection, human review, and account suspension.

The company also layers behavioural controls into the model: refusals learned from human feedback, classifiers that screen inputs and outputs, and system instructions that set the assistant's scope. A model specification published in 2024 describes the intended behaviour in general terms, and system cards document evaluations and residual risks for each release.

No control is complete. Users have found ways around refusals since the first release, and the policy documents themselves describe the limits of enforcement. NIST's risk framework treats these measures as risk mitigation to be measured rather than as guarantees, and recommends monitoring after deployment as well as testing before it 5.

The release prompted an unusually fast commercial response. Within months, competitors had shipped comparable products, and OpenAI's model became the reference point for a wave of applications built on its API. By 2024, surveys cited in the AI Index found 78 per cent of organisations reporting some use of AI, up from 55 per cent a year earlier, with generative tools a large part of the increase 6.

The effects on education, publishing, and software development were immediate and contested. Schools and universities moved from bans to guidance; coding assistants changed how some developers work; publishers and authors sued over the use of copyrighted text in training, and the litigation has not resolved the underlying questions. Customer service, marketing, and translation absorbed the tool quickly because output could be reviewed before use.

Claims about productivity gains are widespread and unevenly evidenced. The better studies find large gains on some writing and coding tasks and negligible or negative effects on others, depending on the task, the worker's skill, and whether the output is checked. The AI Index summarises a growing body of such research without resolving the variance 6.

The product also became infrastructure. Companies built support desks, document pipelines, and coding tools on the API, and a market of consultants grew up around deploying it. That indirect adoption is harder to measure than the consumer app but is probably larger in economic terms, and it is where most of the employment effects would appear.

The product is sold as a subscription with a free tier, and as an API billed per token. The paid tiers differ in model access, usage limits, and features rather than in the underlying assistant, and the free tier has been the main route by which the product reached a mass audience.

The cost structure is unusual for software. Every query consumes computation, and longer conversations and reasoning models consume much more. Inference at scale is therefore a variable cost that grows with usage, which is why providers impose rate limits, offer cheaper small models, and cache repeated work. The AI Index's finding that inference at a fixed capability level became dramatically cheaper between 2022 and 2024 is the economic reason the product could be offered free at all 6.

Access also depends on infrastructure. The service has experienced outages under demand, and features have rolled out by region and by tier. Researchers and developers have complained that access to the most capable models is gated by price, which shapes who can study them.

The first major regulatory action came from Italy. In March 2023 the Garante ordered a temporary limitation on ChatGPT's processing of Italian users' data, citing the absence of a legal basis for training on personal data, the lack of age verification, and inaccurate outputs 7. The service was restored after OpenAI added an information notice, an opt-out, and an age gate, and the episode established a template for European data-protection scrutiny of generative products.

Other disputes followed. Data-protection authorities in several countries opened inquiries, the European Union's AI Act added transparency and copyright-related duties that apply to the models behind the product 8, and copyright owners sued in the United States over training data. OpenAI has published policies and blog posts defending its practices, and the legal questions about training on published works remain unresolved in most jurisdictions.

The governance of the product is therefore split between the company's own rules, national regulators, and courts, with no single authority setting the terms. That fragmentation is itself a subject of criticism from providers, who argue that compliance costs favour large firms, and from regulators, who argue that the firms have not supplied the information needed for effective oversight 4.

Critics make five broad arguments. The first is about truthfulness: a system optimised for plausible text will produce confident errors, and the interface encourages users to treat fluent prose as authoritative. The second is about consent: the model was trained on text whose authors did not agree to its use, and the resulting product competes with some of them.

The third is about labour and education. Automated writing and coding change the demand for the work that produced the training data, and schools have had to decide whether the tool is a calculator or a shortcut. The fourth is about concentration: a handful of firms control the most capable models, the compute to train them, and the distribution channels through which they reach users. A fifth concerns the environment: every query consumes computation and cooling water, and the aggregate load of hundreds of millions of users is a material addition to data-centre demand, though providers dispute the magnitude.

On copyright, the legal position is being tested in courts and, in Europe, through the AI Act's requirement that providers of general-purpose models adopt a copyright-compliance policy and publish a summary of training content 8. OpenAI has argued that training on publicly available text is transformative and lawful; rights holders argue that it reproduces their work and competes with it. The outcome will shape what future models can be trained on.

Defenders reply that the product is useful, that its errors are visible to anyone who checks, that access is broad and cheap relative to the alternative, and that the criticisms describe problems of governance rather than reasons to withhold the technology. They also note that the same concerns apply to earlier general-purpose technologies, which were eventually regulated rather than banned. Differing accounts extend to the product's significance. One view treats ChatGPT as an interface innovation that made existing models usable; another treats it as a step change in capability; a third treats it as a demonstration whose limitations will define the next decade of work. The evidence supports parts of all three, which is why assessments of its impact diverge so widely.