When text-to-image generation became big in 2022, many people reacted in shock. I heard a lot of people saying that nothing like this had ever happened before.

They said that “AI” was going to “kill art.” But people said that back when photography was invented too. They said that it’s not art because it’s just “pressing a button.” But people also said that about photography for over a century. They said that “AI” can’t make art because it doesn’t have soul. But people also said that about recorded music. People said that “AI” will take jobs from artists, which also happened with recorded music, computer animation, and many other past art technologies. People said that it is “just copying”… but all art involves some level of copying and inspiration, and especially artforms explicitly based on copying, like photomontage, sampled music including much hip-hop and turntablism), and appropriation art. (The “just copying” claim is factually not true for visual generation, with rare exceptions.) People say that most art with the new technology is bad, but most art made at the dawn of any new art technology is not “good”. And so on, and so on.

Even if all new technologies have a lot in common, they also must have differences. But what are the relevant ones for “AI”?

Certainly the fact that many popular “AI” models are trained from scraped datasets is important in some way, but that alone does not indicate its impact on the arts. Internet search engines are also built and trained from large scraped datasets in ways that can affect the arts.

The value of historical analogies and trends

In my work, I have argued that new artistic technologies tend to appear in consistent ways, good and bad, including the backlash. Teasing apart these historical analogies is important for a few reasons.

First, the historical analogies help us overcome cognitive biases. Our gut reactions to change are often wrong, and historical examples help us see this, and help us better understand the real dangers. For example, a natural response is to try to protect existing artists is to expand copyright, but this requires caution since overbroad copyright has a long history of hurting artists.

Second, if you want to convince people, then you need defensible arguments. For example, Ted Chiang’s argument against “AI” as art seems not to have been informed by any knowledge of the history of photography or conceptual art or computer graphics or the way algorithms make choices or current developments in interactive tools. His piece will not persuade people who know about these things. His piece could have been a good argument that text-to-image is not a good artistic tool, something I agree with, but that’s not the argument he claimed to be making.

One possible difference is that the tools became effective much faster than with previous technologies. Portrait photography took decades to become widespread, and consumer photography took even longer. But it’s hard to judge speed in the moment, amongst all the breathless hype and controversy; some claims of the immediacy of “AI”’s actual impact may be overstated.

The biggest difference

After a recent talk that I gave, a student in the audience pointed out a way to phrase the difference that I think may be important.

Nissim Maruani wrote to me that: “for the first time in history, human and machines are actually creating using the same art medium … what does art become if the spectator cannot distinguish generated and human-created content? More literally, what would be the incentive for a human artist to publish on Spotify if the public cannot distinguish their work from free, on demand, and even maybe personalized music?”

In short, the new tools are different in that, not long after invention: it is often impossible to tell how generated imagery, video, audio was created.

Was it hand-drawn, digitally painted on a tablet, or generated by algorithm? There are some telltale signals (weird physics, extra fingers appearing), and various stylistic clues but it’s often very hard to say for sure. To some extent, telling the difference may be possible with expertise and connoisseurship, although this can often be misled

In contrast, early photography was never mistaken for painting. You would rarely listen to recorded music and think that it’s a live performance (exceptions include lip sync). Early computer animation looked nothing like live action film or hand-drawn animation.

The move from analog to digital techniques might be the closest analogy here, and each of these had some controversies in their time, e.g., digital cinema versus analog photography, digital image editing. Nowadays one cannot tell the difference between computer graphics and live-action cinematography.

Implications

So, unlike a lot of the gut-response objections, I believe that this Indistinguishability one is valid, and worth considering the implications of.

I am not claiming that the idea is new, but that this is a defensible way to describe how these technologies differ. People have been discussing the implications of this fact, even if they might have phrased it differently. And, I think this is something that was missing from my previous discussions of the topic.

Here are some possible takes.

One take is “AI slop” “crushes creativity”, by making all work meaningless when we can’t tell any of it apart.

Culture theorist W. David Marx has argued that cultural value has already stagnated for the past several decades, and that “AI” merely accelerates this trend: “less value is created when all cultural artifacts are procurable with enough money, can be made anywhere by anyone, and offer no useful social distinctions between philistine and aesthete,” fulfilling an “extinction-level destruction of cultural value” predicted by the Postmodern philosophers decades ago. He further predicts “a re-evaluation of real-life experiences”… maybe we’ll get away from our screens more, if online cultural activities offer less status or meaning?

I do believe that history shows that the best art requires human authorship, experience and connection; we care about art because it is made by humans. Improved tools raised the bar for the best art. But there are many areas where we may not care so much, and these are the areas that “AI slop” inhabit.

I also think the analogy to photography remains useful: photographs didn’t look identical to paintings, but they were close enough to radically challenge the existing notions of the value of creating realistic pictures. Once photography grew widespread, making realistic pictures was no longer valued in the way it had been when it required a painter’s skill. The subsequent Modern Art movements reversed the understanding of artistic skill and talent in the contemporary art world; in Van Gogh’s words, recreating reality was “just photography”.

While skill may still be valued in contemporary art, it is nothing without a good origin story and ideas behind it. In response to the idea of connoisseurship, contemporary artist Jason Salavon likes to quote the study showing that wine enjoyment is affected by beliefs about the price of the wine in a way that can be measured neurologically.

Mass culture still values skill in a way that the contemporary art world does not. Perhaps mass culture will go through a similar transition.

Thanks to Kabir Ahuja for comments and pointing out the link to digital cinema.