The day ChatGPT launched, it told on itself

The first sample OpenAI showed the world was a bug report. Someone pastes a few lines of Go code and types, all lowercase, that it is not working like they expect and how do they fix it. The bot doesn't fix anything. It says it's hard to tell without more context and asks what the code is supposed to do.

That was November 30, 2022, a Wednesday. OpenAI published a post called "ChatGPT: Optimizing Language Models for Dialogue," opened a page at chat.openai.com, and said usage was free during the research preview. The stated reason for releasing it was to get feedback and learn about its strengths and weaknesses.

What OpenAI actually said it was

The post is short and pretty frank. ChatGPT was described as a sibling model to InstructGPT, fine-tuned from a model in the GPT-3.5 series that finished training in early 2022. It was trained with reinforcement learning from human feedback: human trainers wrote example conversations playing both sides, then ranked alternative answers so a reward model could learn which responses people preferred.

The pitch was the dialogue format. OpenAI said it let the model answer follow-up questions, admit mistakes, challenge incorrect premises and reject inappropriate requests. One sample asked about Christopher Columbus coming to the US in 2015, and ChatGPT pointed out he died in 1506.

The limitations section is the good part

I'd argue the most useful paragraph on launch day was the one listing what was wrong with it. OpenAI wrote that ChatGPT sometimes produced plausible-sounding but incorrect answers. It said the model was sensitive to small changes in phrasing, so a question it claimed not to know could get answered after a slight rephrase. It said the model was often too verbose. And then this:

Ideally, the model would ask clarifying questions when the user provided an ambiguous query. Instead, our current models usually guess what the user intended.

Read that one twice. The company that built it told you, on day one, that it would rather guess than ask. The bug-report sample was the nice case. Most of the time you'd just get a confident answer to the question it assumed you meant.

Models have changed a lot since then, and the newer ones do ask more often. But the habit of guessing never fully went away, and honestly I think it's the right default for a tool that people expect to answer right now. The fix sits with you, and it costs one sentence.