It appears the API is following links, is this a feature?

If I just sent a url to the API I get back a response explaining what the page is about. Is this a feature that we should expect with the API, I thought it wouldn’t make requests?

Did you url contain any text that described what the endpoint was?

i.e. was the links something like www.shapes.com/draw_a_square.html and then the AI said the page was for drawing squares?

The API is not internet connected so, unless the model guessed what the page does or was trained on the pages contents then it cannot tell what a url contains.

Yes, interesting, I guess that makes sense but it’s pretty accurate so I thought it had scraped the site. Thanks

This is the type of request that now has lots of fine-tuned denials within latest models to avoid this type of hallucination.

Filling the AI with denials just to counter user misconceptions about the nature of AI makes it overall dumber and non-compliant, though.