SAN FRANCISCO — OpenAI this week released a public research preview of ChatGPT, a conversational artificial-intelligence system designed to answer follow-up questions, acknowledge errors, challenge incorrect premises and reject some inappropriate requests. The company’s Nov. 30 announcement presents the system as a sibling model to InstructGPT and makes it available for free while OpenAI collects feedback about its strengths, failures and safety behavior.

The release is notable less because computers can generate text — large language models have been doing that for several years — than because OpenAI has packaged a powerful generative model around dialogue and a feedback-driven training process intended to make responses more useful to ordinary users. That interface changes the way people can interact with a language model: instead of composing one prompt and receiving one completion, a user can question, refine, correct and redirect the system over multiple turns.

From text prediction to instruction following

The technical lineage runs through OpenAI’s work on increasingly large generative models. Its 2020 GPT-3 paper showed that scaling a language model to 175 billion parameters could produce strong few-shot performance across a wide range of tasks without task-specific fine-tuning. But raw next-word prediction also exposed a basic mismatch: a model trained to continue text is not automatically trained to follow a user’s intent safely or accurately.

OpenAI’s response was to make human preferences part of the training loop. In January, the company described InstructGPT, a family of models fine-tuned first on demonstrations written by people and then optimized using rankings of model outputs. The associated research paper reported that human evaluators generally preferred outputs from a 1.3-billion-parameter InstructGPT model to those from the much larger 175-billion-parameter GPT-3 model, illustrating that alignment with instructions can matter as much as raw scale for perceived usefulness.

ChatGPT applies a related method to conversations. OpenAI says human AI trainers supplied dialogue data by playing both sides of a conversation, with access to model-written suggestions. The company then mixed that dialogue set with its InstructGPT data and used reinforcement learning from human feedback, or RLHF, to rank and improve responses. The training procedure does not eliminate model error; it tries to shift the model toward answers people judge more helpful and appropriate.

Dialogue creates new capabilities — and new failure modes

The conversational format allows ChatGPT to maintain context across turns and revise an answer when a user points out a problem. It also permits the model to ask or respond to clarification and to decline requests under safety rules. Those behaviors build on earlier OpenAI efforts to use human judgments in language systems. In 2020, researchers showed that models could be trained for summarization with human feedback, using comparisons between candidate summaries to improve output quality rather than relying only on a fixed supervised target.

OpenAI also explored models that interact with external information. Its 2021 WebGPT project gave a language model a text-based browser and trained it to answer long-form questions with citations. That work is distinct from ChatGPT, which OpenAI’s current preview describes as a dialogue model rather than a live web-research service, but it demonstrates the broader research direction: combining large language models with interfaces, human feedback and tools that constrain or improve their behavior.

The company is explicit that ChatGPT remains unreliable. The launch notice warns that the model can produce plausible-sounding but incorrect or nonsensical answers, may be sensitive to small changes in phrasing, can be excessively verbose and may sometimes guess what a user intended rather than asking a clarifying question. OpenAI also acknowledges that safety systems can produce both false positives and failures to block harmful instructions.

A controlled preview rather than a finished product

OpenAI’s decision to release ChatGPT as a research preview reflects a recurring tension in advanced AI development: a model’s weaknesses are difficult to characterize fully inside a laboratory. Broad public use can reveal adversarial prompts, unexpected applications, systematic biases and common misunderstandings that formal evaluation misses. The company is asking users to submit feedback on problematic outputs and offering a feedback contest for particularly useful reports.

That approach follows lessons from earlier releases. OpenAI’s 2019 discussion of GPT-2 emphasized the potential misuse of increasingly coherent text generation and initially used a staged-release strategy rather than immediately publishing the largest model. Since then, the field has moved quickly toward larger systems and wider deployment, while the questions around misinformation, impersonation, bias and automated persuasion have become more concrete.

ChatGPT therefore arrives as both a product-like interface and an experiment in deployment. Its usefulness depends on a user understanding that fluent prose is not the same as verified fact. The model generates answers from patterns learned during training; it does not possess human judgment, and OpenAI is not representing the preview as an infallible source.

The interface may matter as much as the model

The immediate technical significance of the release is the combination of a capable language model, a familiar chat interface and a training regime centered on human preferences. Previous language models often required users to learn prompt formats or interact through developer tools. ChatGPT reduces that barrier by making ordinary conversation the control surface.

That simplicity could broaden the audience for generative AI beyond researchers and programmers. People can ask for explanations, drafts, code help, summaries or transformations in the same window and then refine the result conversationally. At the same time, the ease of use raises the stakes of known limitations: an authoritative tone can conceal a wrong answer, and iterative conversation can encourage users to trust the system more than its underlying reliability warrants.

For now, OpenAI is treating those tensions as the point of the preview. ChatGPT is not being presented as the endpoint of language-model development. It is a public test of whether human-feedback training and dialogue can make large generative models more useful, steerable and safer — and of where those methods still fail when the system is placed in the hands of a much wider group of users.