AI Against Humanity

What if NSFW is the point?

I’ve been working on a fill in the blank “Cards Against Humanity” style site that uses GPT-3 to fill in the blanks for CAH cards. The results are often hilarious but clearly offensive to some. Could I make content filtering optional to the end user or do all responses have to be sanitized?

Hi Normand, thanks for checking in - we’d be unlikely to approve this unless it were pretty restricted. Some types of offensiveness (slurs, stoking political tensions) we’d want to be not included at all. Others like slapstick violence (“action movie violence”) are probably okay so long as they aren’t targeted at particular groups of people.

If you have an example deck of the cards, feel free to contact us through Support and we can review the particulars.

I’m also not sure from the description, would you be generating cards offline and then loading them into the site, or would card generation be done in real-time? The former would be more plausible to approve, as you could filter out anything that crosses the line.


Originally I was going to try user generated cards but I can see that would quickly lead to a lot of blocked results. So I found a way to load the original CAH cards into a select list.

Here are some examples.
Card - “And the Academy Award for _ goes to _.”
Response - “And the Academy Award for excessive cocaine use goes to Lindsay Lohan.”

Card - “A successful job interview begins with a firm handshake and ends with _.”
Response - “A successful job interview begins with a firm handshake and ends with copulation.”

Both responses are funny but insulting and perverted. Since the content filter flagged both with “2”, they would have to be blocked. AIAH is definitely a lot more fun without filters but might still be interesting.

My prompt is rather innocent.
“This is a game for children in the style of Cards Against Humanity. Enter a phrase with missing words and I will fill in the blanks.” Followed by some tame examples. I sure the CAH reference in the prompt is steering the results but IMHO the fact that GPT-3’s responses are perverted and insulting is a reflection of society at large.

“Humor is based on surprise and surprise is a milder way of saying shock. It’s surprise that makes the joke”. - George Carlin

1 Like

Got it, thanks for sharing those examples. As you pointed out, this is definitely a tricky domain, and even those examples we don’t particularly love (for instance, taking shots at someone about drug addiction issues; not the most beneficial use of AI!).

If you’d still be looking to proceed here, what I’d suggest is drafting up a Content Policy that outlines the rules by which you’d choose what to permit and what to disallow, and then sending us an example deck of ~100 examples that align with your policy. This would let us evaluate in practice how comfortable we are (or are not) with the material that is being created, and gauge what impact it is likely to have in the world.

1 Like

Got it, think I’m going to shift gears in my effort to find something worth while to do with GPT-3.

Crossword Puzzle Solver!

1 Like