OpenAI launched ChatGPT on 30 November 2022, describing it as a conversational AI model fine-tuned from GPT-3.5 using reinforcement learning from human feedback [source]. The company positioned the release as a research preview to gather user feedback on strengths and limitations.

The announcement acknowledged known failure modes. ChatGPT may generate plausible-sounding but incorrect or nonsensical answers, a behaviour OpenAI attributed to the difficulty of distinguishing truth from model confidence during training [source]. The model was noted to be sensitive to input phrasing, sometimes claiming ignorance on one formulation but answering a rephrased version of the same question.

OpenAI reported the model tends toward verbose responses and overuses certain phrases, behaviours traced to training biases and known over-optimisation issues in the reinforcement learning process [source]. The system was designed with content moderation guardrails to refuse inappropriate requests, though OpenAI stated it may occasionally respond to harmful instructions or exhibit biased behaviour.

The release followed a moderation-focused approach. ChatGPT was trained using Moderation API guidance and human review to filter out unsafe training prompts [source]. OpenAI indicated the model would ask clarifying questions when user intent was ambiguous, rather than guessing, though this behaviour was described as inconsistent.

The company framed the launch as an iterative deployment, stating that real-world use would surface failure cases not captured in controlled testing [source]. Users were invited to provide feedback through an interface integrated into the chat system. OpenAI did not specify performance benchmarks or error rates at launch.