Photographs have always been manipulated. Will AI make things worse?
My craft has been unfairly maligned. As a graphic designer, I toy with images to illuminate and distill information, to evoke, represent. After Photoshop became popular in the ‘90s, it was common to see articles blaming the software for destroying trust in the media, for anorexia, for imposing beauty standards — for the end of truth.
But people have been editing photographic images, painting over them, cutting, pasting and doctoring for almost two centuries. Something is different now. Using techniques from AI — artificial intelligence — photographs can be manipulated without creators understanding how it happens.
Everyone who creates images using pre-trained AI generative software needs to understand the process. Greater literacy will not only enhance creativity and temper fake-news paranoia, it will also make it harder to fool people with AI. However magic it may appear, the technology is neither incomprehensible nor untamable. At least not yet.
Generative AI works with a set of images and their matching descriptive words. If you prompt the software to create an image of cutlery on a plate, the model can easily generate it, because pictures of cutlery on plates abound on the internet. But if you try to get a plate balanced on top of a fork, the AI will be confused. As humans, we can imagine a plate on top of a fork. It’s weird, sure, but we can picture it. When the dataset is biased toward a different logic — forks always go on top of the plate — it’s hard to fight it. Other ideas about where things are “supposed” to go or, say, how a given identity, class or group is “supposed” to look, are pervasive in datasets.
These datasets contain not only lots of images, they contain lots of words. The slightest change in wording can have an impact on the image. Art is about decisions. When an individual is manipulating a photo using the darkroom or Photoshop, the outcome depends on the maker’s intent, technique and imagination. But with AI, someone or something else has already made important decisions that greatly influence the outcome.
This is not the first time we’ve had to reckon with the consequences of a technology that looks like magic. In the mid-1800s, William Mumler claimed to be able to photograph spirits and ran a thriving business photographing people with their deceased relatives. Among his most famous pictures is one of Abraham Lincoln’s ghost with his widow, May Todd Lincoln.
Even after the technique was proven fraudulent, some continued to believe. It wasn’t only his technique. It was the culture: Many people had lost their loved ones in the Civil War, and the mid-19th century Spiritualism movement was in vogue. People wanted to believe in ghosts, the technology only exacerbated those beliefs.
Nowadays, even if someone thinks ghosts are real, it’s unlikely they will use a photograph as evidence. We might not be able to replicate the image in the darkroom, but we have a better understanding of how photographic manipulation works. We know it’s possible to artificially create a picture in which it looks like someone is in the company of a ghost. With healthy skepticism and curiosity-driven literacy, we can again close the gap between what we see and what we know to be true.
When photography was invented, most had never seen an image of themselves as young people. Only the rich could hire portrait artists— for most, the past was relegated to imagination. Photography made tangible proofs of the past accessible to everyone. It changed how we create, imagine and remember.
Like AI now, back then photography represented the democratization of image-making. We are going through a similar, exciting moment. But when damaging images are so easily created and distributed at warp speed, the stakes of understanding the basics of how this technology works — its reliance on datasets, and who made those datasets — are higher than ever before.