Do you want to see an image of teddy bears mixing sparkling chemicals as mad scientists in a steampunk style? Next question: How many do you want? DALL-E 2, the upgraded artificial intelligence imagery program from OpenAI that can create unique images merely from written descriptions, can provide as many as you need. It can also show those bears doing the same in different styles – say that of a 1990s Saturday morning cartoon. Is it perfect? Take a look:
The answer is clearly no, it’s not perfect.1 But it’s not bad.
DALL-E 2 can also be prompted to imitate images that already exist, like a picture of a shop window or home interiors. Here’s an example of what it comes up with:
Like other artificial intelligence programs released in recent years, like ones that produce images of non-existent people or that write entire newspaper columns on their own, DALL-E 2 brings us close to the uncanny – something we know to be unreal but that could be real. Interestingly, what makes DALL-E 2’s images uncanny, and thus reveal its mechanical nature, are its imperfections.
Unfinished lines colliding with others, doors that we can see wouldn’t actually work, surfaces blending together – these are all evidence that DALL-E 2’s images, good as they are, still remain some distance from rendering a world as we know and live it. Human-made drawings would also contain mistakes, but DALL-E’s mistakes are ones humans wouldn’t make. Inaccuracy is a sign of human touch – but only certain kinds of inaccuracies: intelligent ones. The DALL-E 2 images are good, but the imperfections are a tell. It’s just ‘guessing’ because it doesn’t really ‘know’. This isn’t intelligence after all, it’s just covering its ass.
•••
As it happens, DALL-E 2 can also mimic artistic style, arguably one of the most obvious showcases of human creative touch. Among the examples of art OpenAI used as a prompt from which DALL-E 2 could extrapolate was Georges Seurat’s A Sunday on La Grande Jatte, one of the most famous modernist paintings:
Here are a couple of DALL-E 2’s mimickry results:
Seurat’s massive canvas, which he completed in 1886, is famous for its style as much as its fascinating figures. It’s probably the best-known example of divisionism – more commonly known as pointilism. The logic of the technique was based on the assumption that colour points applied separately to build an image would create maximum luminosity – more, anyway, than mixing the paints together beforehand. The approach was semi-scientific, and gave a nod to the burgeoning mechanical age of the late 19th Century. It was also a general rejection of the dominant Impressionist trend in French art.
Seurat’s art, “was not…about pure sensation, or self-expression, or improvisation,” New York Times art critic Holland Cotter explained in 2004. “It was, instead, an un-Romantic exercise in measurement, objectivity, logic, control, with formal decisions made and conceptually resolved before brush touched canvas.” La Grande Jatte, “was a masterpiece of applied mechanics.” A bit like the product of a computer, you could say.
Seurat’s contemporary critics noted this mechanical trait. “Strip his figures of the coloured fleas with which they are covered, and underneath there is nothing, no soul, no thought, nothing,” French novelist and art critic Joris-Karl Huysmans huffed about La Grande Jatte a year after its completion. “I am decidedly afraid that there is only too much process, too many systems here, and not enough of the flame that ignites, not enough life!”
•••
I don’t bring up these criticisms of Seurat to dismiss them or to suggest that any similar criticisms of computer (or AI)-generated art will one day be seen as silly or wrong. Instead, I think it’s important to remember what they suggest about the hierarchical relationship between human and machine capability: that there is one, and we’re on top.
As artificial intelligence systems get better – or as their outputs become more accurate and/or impressive – we are at risk of flipping for good the relationship between ourselves and the machines, if we’re not most of the way there already.
There’s a general sense in our society that our modern computers verge on being all-powerful and we have become increasingly willing to resign ourselves mostly uncritically to their information and their solutions. I worry this has led us to believe not only that their “intelligence” borders on consciousness (it doesn’t) but, that we equally assume the surpassing of our own intelligence is inevitable. We measure ourselves against the computer more and more using a lower and lower bar to assess what counts as a win: that whatever we can do, the computer can do better. We may think we’re training our machines to imitate us, but things inevitably end up the other way around.
DALL-E 2’s images are interesting and they reflect the instructions they were given. But they’re not art, nor are they even really original. Ultimately, they show us what we have shown them. The computer cannot out-imagine us. So why do we need them? I think this is what Huysmans was getting at all those years ago when he criticised Seurat: Why would you limit yourself to thinking as a machine? After all, while we may not be filled with as much information, we are filled with something better than that: life.
Not only are the images not quite exact, the program itself has inherent biases and worrying implications, outlined both in this Twitter thread and in a report OpenAI released alongside the program.
Create your profile
Only paid subscribers can comment on this post
Check your email
For your security, we need to re-authenticate you.
Click the link we sent to , or click here to sign in.