It’s easy to be smug about the “art” of AI because, hey, isn’t that it? looks like garbage?

But the reality is not so simple. To the untrained eye it appears that AI-generated images are more than passable; in some cases they seem to correspond to the old masters themselves.

At least that is what seems to be the conclusions of a recent blind test made by the blog Astral Star Codexwhich found that readers in the study incorrectly distinguished between AI images and human art 40% of the time.

But perhaps the most striking point is that, overall, participants slightly preferred AI creations over human creations, with six of the ten most liked images being generated by AI, and the top two slots being assigned to AI paintings.

This preference was even the case among participants who said they had a deep distaste for AI illustrations – perhaps demonstrating the technology’s baffling capabilities.

For the record, this was not a scientifically conducted test. But with 11,000 participants, it’s big enough to be interesting. Via an online survey, respondents were asked to browse fifty selected images without details, and on each of them, to indicate whether they were human or AI. The ending included additional questions asking to explain certain decisions, expand participants’ familiarity with the art, and other relevant questions.

The blog’s selection of images, meanwhile, was very varied, including many classic and impressionist paintings, a host of contemporary digital art, and AI facsimiles of all of the above. To make the test difficult, the AI ​​images chosen did not feature obvious cues like sloppy hands or gibberish text, and were often made by experienced engineers (in other words, self-proclaimed AI artists).

Certainly, some images could be recorded by the AI ​​from a mile and a half away, like one of the an anime girland a particularly garish representation of giant cats in a throne room. Others however, like a quasi-impressionist night sceneare much more difficult to pin down.

As we said, things to remember are complicated. Perhaps the preference for AI images, many of which were impressionist in style, simply reflects participants’ taste for paintings from that era, such as Star speculates. It’s also worth noting that significant human work went into selecting only the most compelling AI illustrations, meaning participants only received the cream of the crop.

As for people who hated AI images but seemed to prefer them when all labels were removed, it’s possible that the selection doesn’t reflect what they specifically associate as bad about AI images (like the weird “glow” » that we see in so many images). of them).

The result, however, is that most of us lack the refined visual palette needed to distinguish a human touch from a convincing mechanical touch – which makes sense, since AI models are fine-tuned to mimic human existing works of art by somehow making them. all together and selecting the most pleasant parts. Recently, we have seen people moving towards AI as well. with poetrysomething that very few people read regularly.

At the same time, the test showed that there are indeed aliens who correctly identified almost every image, with five elite readers scoring 49/50.

We will end with the response of an experienced artist who participated in the challenge and showed remarkable talent to be a true art blade runner. A big eye-opener, she says Astralis that the details of human images “have logic”, while those of AI are only “superficially detailed”.

To summarize his point, AI art is like the culinary equivalent of trying to replace food with a tasty, nutritionally adequate protein drink that you could make inexpensively.

Of course, you have enough and the porridge pleases the taste buds. But “imagine people calling this the future of food and saying chefs are obsolete,” she wrote. Astral. It would probably drive you crazy, eespecially if you are an artist.

Learn more about AI: Newspaper fires two AI journalists after bizarre behavior