So... The question is, how on earth an AI model can get today's date? The model should just try to predict what the next token should be expected most. But any date might be possible. The model and openAI claim that the model doesn't have an internet connection, and do not have any time source.
Then how?
@skyblond
I think it's probably had a lot of fine-tuning training on examples like:
Q: [question it shouldn't try to answer]
A: As an AI model, I can't help you with that.
which raises the probability that it'll do that significantly.
On knowing the date, there's been work on getting LLMs to call appropriate APIs; see for instance https://arxiv.org/abs/2302.04761
I don't know anything specific about ChatGPT (or any other LLM for that matter), but that's my best theory.