When GPT-4 was first announced in March 14th, OpenAI did a livestream showcasing some of its capabilities, and during the livestream Greg Brockman showed how GPT-4 could be used to develop a discord bot to call the OpenAI API.

This leaked enough information for someone to determine what the invite link of that bot was, and when it was online in discord it would actually respond to mentions and answer people’s requests - including images!

I had access for approximately 24 hours.

This screenshot is from the GPT-4 discord bot on March 16th, right after the original announcement. Screenshot showing discord bot identifying Trump correctly.

And this is from October 3rd, after the new release. Screenshot showing GPT-4 refusing to identify Trump. So it now refuses to identify people at all.

What if I, someone it does not know, want to be roasted?

Again, from March 16th. Screenshot of me getting roasted by GPT-4.

And now from October 3rd. Screenshot of GPT-4 refusing to roast me.

So since the time between the announcement back in March and the release for images in September, OpenAI has RLHFed the model to refuse to comment at all about people.