Article begins
When AI-generated images began to flood the internet in autumn 2022, beholders were struck by the naturalism of these algorithmically rendered human figures, objects, and environments. One niggling issue became apparent though: the AIs couldn’t get a handle on the finer details of human anatomy. Hands, ears, teeth, and toes appeared in uncanny forms, exhibiting too few or too many articulations, strange meldings, impossible configurations. Some commentators breathed a sigh of relief at this evidence of AI fallibility. Others reflected on why these glitched details caused (human) viewers so much concern. Almost immediately, developers like Midjourney and Dall-E set out to remedy the flaws by creating datasets that focused on these unruly aspects of the human form. These “errors” made clear that generative-AI models do not know what these body parts are, beyond a collection of patterned pixels. They cannot understand the pragmatics of cuspids and incisors, lobes and helixes, opposable thumbs and triphalangeal fingers. In the final instance, a generative AI cannot know how central these “minor” limbs are to human practices of sensing, living, and making (including meaning-making). It cannot know that the dexterity and versatility of the hand, in particular, is central to our myths of human (or hominid) exceptionality.
What to make of this fleeting moment in the emergence of generative AI? For me, the focalization of “minor” physiognomies recalls historian Carlo Ginzburg’s elaboration of the link between the Italian Renaissance art connoisseur Giovanni Morelli and Sigmund Freud. Ginzburg notes that both scholars’ methods hinge on close attention to minor details—of a painting and of a patient’s discourse—that are seen to illuminate an elusive whole. For Morelli, “insignificant” details become clues as to an image’s authorship; these innocuous tidbits betray the identity of the artist (or forger) where more obvious figural patterns are liable to lie. Likewise, for Freud, a slip of the tongue can disclose the undercurrents of a patient’s consciously produced narrative. Recalling these modes of analysis, we might ask: what do the figurative glitches in AI-generated images reveal about the process of its “author” or “authors”? What happens if we approach these snags as symptoms of a latent artificial unconscious, fed on an aggregation of human-wrought images? Conversely, what might we make of public response to these flaws? Although generative-AI technologies seem to have surpassed their early anatomical hurdles, let’s tarry a little longer in the space of “error.”
First, on the question of generative-AI “authorship.” Rather than rehearse the debates around intellectual property rights for uncredited artists, I’d like to consider what the glitches in early generative-AI images indicate about their in/human author(s). Botched details appear where the algorithms’ manifest programmatic logics falter, yet such glitches are remarkably systematic; one flawed digit or bicuspid could be seen as an aberration, but these recurrent “errors” suggest an underlying organization. For this reason, I suggest, it’s imperative that we attend to these “glitches” and ask what elseis articulated there.
Published explanations offer a few clues. In spring 2023, AI spokespersons, educators, scholars, and journalists echoed two refrains: (1) the training data sets were flawed or incomplete; and (2) generative AIs cannot understand the three-dimensional pragmatics of human anatomy. The first of these refrains accented an easily remedied problem; if faulty data sets had precipitated anatomical “nightmares,” then surely the solution would be more/better data! (And doesn’t this sound familiar?) Commentators cited the obscurity of certain physiognomic details within the typical AI-training datasets. Because most photographs of people do not clearly display the structural complexity of hands, ears, teeth, and toes, the repository of “good” examples is relatively sparse. Such anatomical features only rarely break the surface of automatic perception, instead inhabiting that chthonic realm Walter Benjamin dubs “the optical unconscious.” Artists’ renditions likewise often obscure, distort, or stylize these physiognomies. Accordingly, early generative-AI training datasets were populated by images that introduced informatic noise: images in which a “hand,” for instance, is entangled with other bodies, is partially obscured, is deliberately abstracted, or is so pixelated as to be structurally indistinct. It’s not just that these already-small features are further minimized within human-made images; they also vary wildly across appearances. The lauded flexibility of prehensile hands and idiosyncrasies of depicted ears, toes, and teeth result in a noisy dataset. Faced with this cacophony, we can understand why algorithms would have difficulty inducing a statistical norm (which is how they tend to work).
Given obscure and conflicting images, early generative-AI models needed to ad lib on the basis of the cues that they did have. In short, they needed to improvise rules for depicting handness, toothness, earness, and so on. This brings us to the second refrain: early generative-AI models produced grotesque anatomical features because they were fundamentally unable to grasp the three-dimensional structures and corporeal pragmatics of human hands, toes, ears, and teeth. Having been trained on millions of static, two-dimensional images in which these body parts are often indistinct and pixelated, AI models seem to have arrived at a few qualitative common denominators for each appendage. For instance, if early generative-AI ears appeared as “fleshy whirlpools,” then this phenomenon could be said to reflect an interpretation of “earness.” Across their varied appearances, ears could be distilled into the concatenation of fleshy, concave, and volute. It’s worth emphasizing that such “flesh” reflected a strong skewing towards lighter tones, reinforcing what Sylvia Wynter calls an “over-representation […] of an ethno-class conception of the human.” Rather than interrupt systematic discrimination, these and other AIs have tended to amplify the biases implicit in their datasets. (Such “flaws” are more ominous than glitched fingers.) So we encounter a machine’s-eye view of “earness,” disconnected from the pragmatics of human corporeality and reduced to a bundle of hypostatized qualities: fleshiness, concavity, and volution. Although gen-AI models could approximate these features’ characteristic pixel-patterns, their basic structure eluded the algorithms’ grasp.
Further, “errors” indicate that the algorithms still cannot compute these appendages’ importance for human-bodied beings. They have not (yet) incorporated the practical and figurative significance of such “minor” anatomies. To this point, I’ve suggested that early generative-AI glitches can be read as symptoms of latent processes operating in lieu of statistically determined rules, refracting the biases and obscurities of their human-wrought data feed. So long as we’re lingering in the glitch, we might also consider the symptomatic dimensions of popular response. Notably, despite other anatomical distortions, the majority of commentators focused on AI-generated hands and fingers. Perhaps this focus reflects a feedback loop within media circuits, but I suspect it also evinces the emblematic role of the hand in definitions of “the human as homo faber,” a tool-wielding animal with distinct creative powers. Recall Michelangelo’s depiction of the divine touch granting Adam the spark of life, a dramatization of the relation between creator and created that has been recreated in multiple memes about AI art (unsurprisingly).
Many responses to early gen-AI errors disavowed similarities between algorithmic models and human artists. For instance, the New Yorker’s Kyle Chayka writes, “Hands are a symbol of humanity […] As long as we are the only ones who understand them, perhaps our computers won’t wholly supplant us.” Encyclopaedia Britannica’s Meg Matthias echoes, “Perhaps AI can’t understand human hands […] because it can’t understand what it is like to be human.” Such remarks suggest that “They” cannot understand what it means to be “Us,” finite, fleshly, en-handed beings. In order to achieve “realistic” human anatomy, quipped one anonymous commenter, an AI would need to “build a human from the bones to the muscles to blood to skin.” The crux? Gen-AI models cannot properly depict human bodies because they don’t have them.
Other responses suggested the converse, that gen-AI models are well on their way to becoming “more human than human”; these early wobbles are simply their first (legless) steps. Again, comments honed in on hands and fingers, this time citing human artists’ difficulties as proof of the task’s complexity. In an online discussion thread, for instance, users proposed that generative AIs “fail at hands” for precisely the same reasons human artists do. Further, commenters suggested that children often commit implausible anatomies to paper and that errors are part of the learning curve. Not only do novice artists often struggle to depict convincing hands, but art history is also rife with non-naturalistic appendages. One writer suggests that “only about 0.3% of our 200,000-year-old art history features beautiful hands,” so gen-AI models shouldn’t be faulted for a few misplaced fingers. Another asserts, “artists throughout time have avoided […] drawing hands because of their difficulty […] It wasn’t until the Renaissance period that artists like Leonardo da Vinci started studying and sketching hands.” Setting aside several tenuous assumptions—that naturalism equates to beauty, that “realism” is universal and timeless, and that the history of art comprises a singular, linear “progress” towards ever-greater naturalism—we might note the lamination of individual artistic improvement, world-historical artistic progress, and the projected development of gen-AI models. Algorithmic development recapitulates ontogeny which recapitulates phylogeny. Or something like that. Such responses to early gen-AI errors emphasized the similarity between machine learning and human learning, suggesting that generative AIs can learn to depict human bodies without possessing them. (A reassuring sentiment, from a Black Mirror standpoint.)
A third strand of popular response goes something like this: gen-AI distortions reflect the way hands appear in dreamsand nightmares. Rather than imitating empirical anatomies, these algorithms latch onto submersed realities that human dreamers grasp in shreds and tatters and render them perceptible to the waking eye and mind. What does it mean to frame generative-AI images in this way? It depends on how we understand dreams. Are they “all too human” or the stain of something in/human that eludes conscious perception? Either these media-gorged machines have absorbed a fundamental grammar of the human imagination, or their creations are reminders that something other whispers at the edges of our consciousness. Perhaps it’s a bit of both. Is the perception that generative-AI images are the stuff of human dreams and nightmares a “symptom of a symptom” or its folk diagnosis? So long as the lights are out, so is the jury.
Responding to early anatomical glitches, AI developers created bespoke data sets and began to refine AI models’ 3D pointmapping and modeling capacities. Improvements have been rapid and evident. Most obviously, recent generative-AI images tend to feature subjects with five fingers on each hand instead of the usual proliferation (the irony of a digital technology’s initial inability to depict the right number of anatomical digits seems to have gone unremarked). Artists and designers no longer need to correct AI-generated anatomies using Photoshop, and the software is more able to accommodate all comers. In short, these changes herald a plug-and-play technology for creating convincing, photorealistic images.
Improvements are double-edged, however. Not only do some digital artists lament the lost opportunity to manually tweak glitched images, but such developments have also erased a key insight into the processes that underpin AI image generation. As companies continue to roll out updated gen-AI models, public discourse has increasingly turned to the question of deep-fakes. Critical media audiences appear to have valued early generative-AI glitches on Morellian grounds—as clues that an image is a “fake.” Just as an art forger’s intuitive way of sketching an earlobe can give them away, so too can a digital image’s provenance be revealed by characteristic AI flourishes (a luxuriation or absence of lobes, for instance). Over the past 18 months, many of us became amateur detectives, whipping out skeuomorphic magnifying glasses to scour online images for indications of artificial authorship. Such detective work (like art connoisseurship and psychoanalysis) relies on abductive logics that link scattered indexes to an otherwise obscured whole. I have suggested that early gen-AI glitches offered precisely this sort of index. In this respect, the current wave of generative-AI hands are far poorer manicules (☛) than their malformed predecessors.
By padding the gaps in generative-AI models’ knowledge base, developers have curbed algorithmic creativity in order to offer up more docile (and commercially viable) machines. What is lost in this process of domestication? And what is now obscured, ticking along beneath the glossy surface? If early AI errors exposed a raw nerve cluster at the intersection of human and in/human modes of image making, then perhaps we have relinquished valuable insights by suturing the glitch with more data. I’d like to end by echoing a recent provocation: rather than surround ourselves with pliant and predictable AI models, “we must also make space for unruly algorithms.” I suspect they will accompany us, one way or another, on the surface or beneath it.