ChatGPT's 2022 launch post named its own model's limits

Start here

The model behind ChatGPT was not new when the product made headlines. OpenAI's own launch post, Introducing ChatGPT, dated 30 November 2022, states that 'ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022' -- months before the chat product existed. What changed on launch day was the interface: a free, conversational, low-friction window that let people without an API key or technical background try the same underlying capability directly. For anyone judging what 'changed' about AI, this is the first distinction worth making: whether it is a new model, or a new way of reaching an existing one.

What the documents say

The launch post describes ChatGPT as a 'sibling model to InstructGPT' trained with 'Reinforcement Learning from Human Feedback,' using human trainers who played both the user and assistant to build a dialogue dataset, then ranked alternative replies to shape the model further. It states its own limitations directly: the model 'sometimes writes plausible-sounding but incorrect or nonsensical answers,' is 'sensitive to tweaks to the input phrasing,' and is 'often excessively verbose.' OpenAI's help center article on what ChatGPT is, current as retrieved on 16 September 2026, adds practical detail: the assistant 'is not connected to the internet,' has 'limited knowledge of world and events after 2021' in that base model, and states that conversations from individual accounts 'may' be used for training unless a user opts out through the service's privacy controls.

Check this

Two checks come straight from the documents. First, if an answer seems wrong or evasive, rephrase the same question rather than assuming the topic is off-limits -- the launch post's own admission of phrasing sensitivity means a slight rewording can produce a materially different, sometimes better, response. Second, before trusting a time-sensitive answer, check what the tool itself says about its knowledge cutoff and whether it is browsing the live web for that particular reply, since 'not connected to the internet' describes the base model, not necessarily every mode a product built on it offers.

What holds and what fails

This is an editorial reading built on the launch materials rather than a claim either document states outright: the pattern of an interface unlocking an already-existing capability holds up as a general lesson, and it is worth applying to the next product announcement that feels like a leap. It fails as an excuse to ignore what actually changed, since a new interface can also change real things, such as who has access, what the service costs, and what happens to the conversation afterward, all of which the documents themselves address directly.

• When a product 'feels new,' check whether the underlying model's training or release date actually predates it.
• If an answer disappoints, try rephrasing before concluding the assistant cannot help with the topic.
• Check the current data-use and training-opt-out settings rather than assuming last year's defaults still apply.

The launch post's own account, a research preview built to collect feedback rather than a finished product, is a useful description to keep in mind for whatever chat interface you open next.