Say you're playing with one of those AI image generators. And you give it a prompt to generate a room with no elephants in it.
I'd like a room that is empty without elephants, please. I said do not put an elephant in the room in this image. What is going on?
Researchers and casual users alike have documented that generative AI tools often struggle with negation, the language we use to express absences, untruths, and opposites.
Gen AI models may be able to show you what you do want to see, but they're not as good at not giving you what you don't want to see.
Just look at that last sentence, and you can understand why.
The ways we use negation in language can be pretty complex, but our brains are built to handle all the nuance of negation, whether we know it or not.
I'm Dr. Erica Brozovsky, and this is "Otherwords." In its simplest form, negation is language's Uno reverse card.
If affirmation expresses that something is true or real or present, "The sun is up," "He wears glasses," "I ate some cookies," negation expresses the opposite.
"The sun is not up," "He doesn't wear glasses," "I ate no cookies." But notice how we add to subtract. Affirmation is basically the default of language.
If I say, "My shirt is blue," I don't have to use any special markers to indicate the thing I'm saying is true.
