Since its debut in 2022, ChatGPT has been the reason for a significant amount of speculation. While some are still dazzled by a tool with such a breadth of uses, it didn’t take long for people to realize that it still makes all sorts of hilarious mistakes, often without the slightest inclination that something is off.
AI researchers call these sorts of mistakes “hallucinations,” as the system has incorrectly put some information together and presented it, quite confidently, as fact. Since ChatGPT “learns” by trawling the internet for text, it makes sense that the nuances of writing data and the very relevant issue that people are sometimes wrong do end up affecting its “judgment.”
This is a bigger issue than one might think at first. It’s also why many artists and writers see AI as mostly just a big “plagiarism” machine since it can only take and modify existing work. If ChatGPT or its image-making counterpoint, Midjourny were students with you at an art school, their work would be quickly dismissed as derivative.
Even worse, people aren’t just often wrong or dishonest, they can be deeply bigoted and prejudiced. These same people will put their thoughts into writing and ChatGPT, without really understanding these things will add these combinations of letters, words, and ideas to their repertoire. To share an example, in one case ChatGPT created song lyrics that insinuated that non-white scientists were inferior to white, male scientists.
Of course, in a way, this is pretty endearing. After all, who among us hasn’t made woefully incorrect statements about anything really? Plus, standing by incorrect information and being confidently incorrect are such deeply human traits that it’s almost quaint to see them in what is supposed to be this piece of science fiction technology.
It’s important to remember that, unlike most humans, chatGPT can never actually understand the context of what it is saying. In that regard, it is like a “stochastic parrot,” a concept coined by American linguist Emily M. Bender. Like a parrot, repeating phrases, an AI simply understands that certain combinations of words should lead to certain answers, without any real analysis. The posts here are a good example of this.
Indeed, some readers might notice how these user conversations with ChatGPT sort of resemble communicating with a pet. It often gets the gist of what you want from it, without knowing why. “Close enough” works most of the time, but there are enough examples of it being quite off, particularly when you ask more serious questions.
