“Artificcio van Intelligo”: Spotting Virtual Fakes
One might be forgiven for assuming that the onslaught of low-quality images generated by AI, littering social media like a graveyard, has not affected real painting scholarship. Unfortunately, that is not entirely true. Certain types of AI styles are good enough to fool the unwary. They occasionally pop up in art history discussion groups online. I’ve even had personal encounters with these fakes myself. In the interest of a warning, and a word to the wise, here’s some pointers on how to spot them.
(NOTE: This article was written before the release of the latest improved iteration of ChatGPT, 4o. Frighteningly, this AI-detecting advice may no longer be helpful in the near future.)
Not long ago, a Reddit user asked for help identifying a particular painting: a generic scene from ancient Roman times. Some wise-guy immediately cited the artist as “Artificcio van Intelligo.” I’ll admit, it took me a second.
So, how can we avoid the work of good old Artificcio?
The primary rule: AI paintings do not make sense. To find the tells, look for details that just don’t match up. Look for mistakes that no sensible artist would make, regardless of skill. Here is the full image in question - see how many you can spot:
If you aren’t looking too closely, one would be forgiven for taking it at face value. But…
What are those things scattered all over the stairs?
What is going on with the mangled figure holding the spear, at the foot of the steps?
Why does the walking dark figure, in the center lane, have two pairs of legs?
Who is the ghostly figure superimposed over the crowd, on the far left?
What is happening with the irregular arches at the top?
Why are there extra limbs in the lower left hand corner?
Why are Caesar’s robes blending into his leg?
What on earth is crawling over the balcony on the right?
And so on, and so forth, et cetera, ad infinitum.
It was created with mindless lack of intention and no craftsmanship - a machine’s mimicking of the human touch. Most AI pictures often have a general sense of wrongness about them. If you know the glossy “AI art look,” this isn’t hard to learn. (This article is a pretty good broad overview about learning the tells.) However, images intended to mimic a certain style, such as art historical genres, are much trickier. They often may look convincing if you squint and step back. After all, they’re supposed to fool you.
These products of “generative AI” are an amalgamation of countless images that the AI has been taught to “recognize” as some particular thing. (That, of course, is a gross oversimplification. But, since I am not a computer scientist, and probably, neither are you, I'll keep it brief.)
For instance, if you feed the AI several thousand pictures of 16th century Renaissance women, and then you ask it to make a new one for you, it might look something like this.
Not my own creation, of course. She was found on Sketchfab.com, and, at a glance, it passes as a perfectly legitimate portrait. The lady is charming in her own right. But, after a second look, I had my doubts. Let’s examine the painting up close, with those AI-identifying principles in mind. There will be failure in the details.
The floral patterns are nonsensical and distorted.
Her necklace pearls are strange blobs.
The central pearl blends into the mangled blue pendant below.
Something ridiculous is going on with her dress fringe.
The bouquet of flowers is floating off to the side for no reason.
Her tiara is a collection of fragments and bits.
The curtain fringe is choppy and irregular.
And so on and so forth… again.
She just doesn’t hold up to scrutiny. Let’s compare her with two very fine and completely real works: Portrait of a Lady, Florence, 16th century, manner of Alessandro Allori; and Portrait of a Lady, 16th century, by a Flemish court painter active in Italy. Both are available in high-resolution and will make a very fine comparison point for the artificial lady.
FACE: The artificial lady’s face is disconcertingly convincing. Although she lacks the direct eye-catching gaze of the two legitimate portraits, her face is “painted” with appeal and charm. This is indisputably the toughest comparison point. The AI is trained to “know” that each portrait needs a face, eyes, and other features. The flawed garment details are much easier to tell: the AI simply does not have any clue what it is supposed to do. We’ll get to that in a moment.
EYES: This is a challenging comparison point. But, by studying the eyes, we can note the irregular lack of roundness in the pupil and iris, nor any attempt at it. Pay close attention to the imperfection of the outer circle of the pupil. There is also a real inconsistency in the style of the inner corner of the eye.
SYMMETRY: The outfits seen in the two real paintings are appealingly symmetrical in design. Not so for the fake. This is a failure of AI specifically: again, it does not know that things ought to be symmetrical in the first place. Note the random strip of red fabric, as well as the irregular tufts of trim, like the pinwheel on her proper right shoulder!
FABRIC: This is one of the most telltale errors of all. The AI’s gold pattern is a nonsensical collection of blurry bits, bearing no resemblance whatsoever to a real floral fabric. Compare it with the beautiful golden embroidered workmanship on the Florentine lady’s garment, and the crossed leaves pattern on the pink dress. Complex as they are, they do make sense.
EARRINGS: Both of the real paintings depict convincing, solidly rendered pieces of jewelry. There is true skill in the craftsmanship. However, the AI fake is nothing more than a nebulous suggestion of brushstrokes. It’s quite inconsistent with the supposedly realistic and meticulous art-style, as seen in the rest of the (non-existent) painting.
DECORATIVE DETAILS: The artificial lady includes multiple attempts at Renaissance fabric, including this blue brocade-type motif with a resemblance to peacock feathers. At a distance, it looks nice, up until you get close! Compared to the ornamental patterns on the real garments, it’s obviously visual nonsense. If you can’t immediately spot the difference, I recommend trying to count the number of decorative details - lines, dots, etc. - and compare with the fake. Most artists do not replicate patterns exactly, but there’s usually a reasonable attempt.
PEARLS: Again we encounter the problem with shapes – here, it’s the lack of consistent roundness. Pay close attention to the way the pearls fade into confusing blobs near the edge of her neck. No such error is seen in the real painting. Although the necklace edge is less precise than the center, the shape of the pearls remains intrinsically recognizable on the right. As for that disastrous pendant, it speaks for itself.
TIARA/HEADBAND: All of the above rules apply: lack of consistency, implausible structure, and so forth. Ask yourself: could this object exist in real life? (In terms of physics, not fanciness or decorative complexity. That wasn’t much of a limitation back in the day, considering the splendor of noble dress.) If it doesn't make any sense, beware.
One final note of caution: some art styles probably do not hold up to the very same scrutiny by these standards. Impressionistic pieces, for instance, are pointless to judge or measure by a lack of precision. However, loose brushstrokes, in those types of real paintings, are purposeful and deliberate. In AI “art,” there can only be a digital failure of mimicry. That’s the core rule of thumb here: always look for signs of the artist’s intentions. If there are none, you’re probably looking at machine-made forgery.
I hope this exercise in 21st century fake-spotting is helpful in some way! It is derived purely from my own experiences and observations while encountering AI fakery in past months - and the wit with the “Artificcio van Intelligo” comment. (The image has since been deleted, but the post lives on.)
- - -
A very final note: I don’t condemn AI in all cases. For instance, I find it perfectly useful as an OCR tool to expedite research document processing, as I’ve written about on this very blog. But it belongs nowhere near the subject of art itself!
P.S. Not all AI art-historical “art” is difficult to spot. The Sketchfab user who uploaded the Renaissance lady also posted this.
3D version of the world famous painting “Der Unicorn Rider” by Wolfgang Pumfelstrasser.
“Der Unicorn Rider” (1897) by the enigmatic German painter Wolfgang Pumfelstrasser stands as a testament to late 19th-century Symbolist romanticism. Pumfelstrasser, known for his eccentric personality and penchant for blending fantastical myth with biting social critique, captures an otherworldly scene of a lone rider astride a luminous unicorn beneath a tempestuous, cerulean sky. The rider, a spectral figure with elongated, almost skeletal features, symbolizes the tragic folly of human ambition tethered to illusion and impossible ideals.
Pumfelstrasser’s audacious use of chiaroscuro, evokes the unsettling tension between fantasy and grotesque reality. The swirling brushstroke imbue the work with a frenetic dynamism that has been compared to a hallucinatory reverie.
That text is all AI-written. There is no “Wolfgang Pumfelstrasser” anywhere on Google, and maybe that’s for the better. But, if this image weren’t so blatantly a modern digital-art concoction, it might fool the unsuspecting viewer. The ease with which academic-sounding material can be faked - visually and verbally - is really quite troubling. Discernment and discretion are still critically important, even for something like that!
