A recent prank by artist SHLOMS demonstrates just how difficult the struggle can be. He announced on X (formerly Twitter) that, using AI, he had created a fake Monet. He invited people to critique it. Of course it was a painting of waterlilies, a subject common to Monet. It did not take long for the punters to disparage the image, criticizing it as vastly inferior to a real Monet. The brushstrokes, the colour, the composition, everything about the painting just screamed "inferior" to the self-styled art critics.
There was only one problem. AI had nothing to do with the image he posted. It was an actual Monet, one of his most famous waterlilies paintings. The prank has been written about in many places, such as here, here, here, and here.
Ooops. There’s a lesson here for librarians. Actually, there are several. First, labels do not always accurately describe an image. This is often an issue when an image purports to be of an event in the news. It may actually be of an older event or from a different location, purposely mislabelled. But it might also be correct. From the AI literacy perspective, it takes discernment and scepticism.
Second, it’s wise not to jump to conclusions. In our current environment, where deep fakes are getting a lot of attention not only in our professional literature but also in the more general media, it’s way too easy to decide something is AI-generated when it isn’t. Take, for example, declarations that a book or an article published in, say, 1999, is "clearly" AI-generated. This ignores the fact that OpenAI only introduced ChatGPT in late 2022; it confuses AI-generated with AI-trained. The 1999 publication is source material for training LLMs, not the other way round. Understanding this is essential to AI literacy.
Third, merry pranksters will always be around to challenge our serious library mindsets.
As information professionals, we should beware of being over-confident when judging something to be AI-generated. The presence of the word "delve", or copious em-dashes, or sentences with three parallel clauses (yeah, like this one) don’t categorically correlate to AI-created. AI mimics human writing and we humans are the ones who use those writing conventions. AI detectors are far from infallible. Common sense and critical thinking are still crucial components of AI literacy.