thearchivist.funChatGPT makes iterative prompting a public interface
OpenAI introduced ChatGPT as a public research preview on 30 November 2022. The release presented a conversational interface in which a person could ask follow-up questions, revise a request and continue working within a dialogue. The launch post illustrated tasks such as debugging code and explained that the model had been trained using human feedback. The interface changed the practical experience of using a language model. Instead of treating every request as a single isolated completion, users could give feedback in ordinary language and ask for another attempt. The visible unit of interaction became a conversation. That made prompting, response evaluation and revision part of an activity that many people could try without writing an application around a model. The announcement also described limitations. It warned that plausible-sounding answers could be incorrect, that wording changes could affect results, and that the model could be excessively verbose or make assumptions about ambiguous requests. These qualifications belong in the historical record alongside demonstrations of capability. A fluent answer was not evidence that the system had verified the underlying claim. For an internet-culture archive, the durable significance is the emergence of prompts and conversational outputs as artifacts that require context. A screenshot of one answer omits the model version, prior messages, settings and the possibility of repeated attempts. Preserving a responsible example means recording the prompt, relevant preceding context, date and product version where known, while avoiding private information. The launch did not establish a timeless set of abilities: later products and models changed. This record concerns the 2022 research preview and the interaction pattern it made widely accessible, not a claim about the capabilities of every subsequent system.
Context and provenance
The launch source is both a primary historical document and a product announcement. Read its examples as selected demonstrations and its limitations as part of the same evidence. The related InstructGPT research explains a training approach involving human preferences, but a paper's experimental results should not be treated as a universal guarantee for a deployed chat product. When comparing AI tools across time, keep the task and evaluation method explicit. A memorable output can reveal an interface's cultural appeal while remaining a poor measure of accuracy, reliability or the experience of other users.
Why it matters
Conversational iteration became a mainstream way to request, inspect and revise machine-generated text and code.
Sources and evidence
Editorial review: 2026-09-23. New evidence may change this account.