The OpenAI API provides direct access to language models.
An AI language model accepts an input (such as messages telling it how to operate, and messages telling it what to produce), and then it generates a predictive language output.
Nothing about that is “the model connects to the internet”. AI models cannot escape from the fact that they simply generate language tokens.
However, recent OpenAI models can also receive a tool specification from you, such as a currency_converter("input_currency":string, "output_currency":string) (in metacode). Offering that tool specification to the AI means it may send an output not to the user, but to invoke the function to get more information.
You must then receive that function in your code, and fulfill the AI’s needs that you promised, such as actually looking up information on an API you pay for, or simply looking up some knowledge in a database. Then you return the new information and the AI can answer the user.
Or offer the AI your own “search_the_internet(search_terms)” function where you perform the searching.
Telling the AI it can do something it cannot will simply produce plausible language not based in fact.
