Psychicones: Hauntologies of Neural Decoding AI Systems

In a rapid burst, a succession of still images parades speedily across the screen. They appear veiled, as if seen through a semi-opaque layer of obfuscation, overlapping to produce a kind of gray matter—visual noise. At certain moments, the legibility of one or another sharpens and we gain greater clarity, the veil momentarily lifts. Shapes emerge: first, what seem to be buildings, then signs, letters from various alphabets, fragments of text, symbols, even Sesame Street’s Cookie Monster. At the bottom of the screen, a cluster of terms vibrates in a shifting text cloud: “book,” “building,” “car,” “computer-screen,” “character.” At one point, the word “character” enlarges and remains dominant, asserting itself over the others.

The online video I am describing is titled Dream decoding from human brain, and it documents the decoding of a dream’s semantic content from brain activity recorded in the visual cortex of a sleeping subject.¹ The visual cortex is the region of the brain that selectively responds to specific visual semantic categories—like “face,” for instance. The images correspond to the eighteen semantic categories available to the trained decoding system—those words appearing at the bottom of the screen—whose probability scores fluctuate, as reflected in the changing size of each term. This is why “character” became larger than the rest. In the verbal report that follows, the dreamer indeed recounts having seen “characters” in the context of writing an essay in their dream. Still, one might wonder whether these images simply triggered neural activations in the visual cortex to other, similar categories (which could perhaps account for the Cookie Monster?).

In another video uploaded to the same YouTube account, Deep image reconstruction: natural images, the screen is split in two.² On the left side, pixelated photographs of animals—an owl, a duck, a swan, a crab, a leopard—are successively displayed. On the right, a hazy image seems to struggle into existence with each new photograph presented on the left. Vibrant, almost alive, the frames flicker in a not-quite-fluid motion that gradually stabilizes toward the end, as if the image were congealing onto a substrate. The images on the right are non figurative. In fact, I tried watching them while covering the reference images with my hand, and it was impossible to identify what the forms were meant to represent. They appeared as nothing more than abstract shapes, gradients of color, mobile lines that never quite coalesced into recognizable objects. Representation slipped away, elusive, like sand through my fingers.

Yet when shown alongside the reference images—the photographs the subject was looking at while undergoing fMRI scanning—clear correspondences begin to emerge: a distant resemblance, a vague echo of shared color palettes, or patterns of light and shadow in similar regions of the frame. It is as if one image enables the reading of the other. Still, it remains unclear how much of this resemblance is genuinely embedded in the image, and how much is projected by the viewer—filling in the gaps of legibility with desire, memory, or inference. Separately, the images are meaningless, but paired together, you can “see” in them.

In 2013, the journal Science published a paper that has since been quoted hundreds of times: A group of Japanese researchers had developed a novel method for decoding mental imagery during sleep.³ The first video corresponds to this publication, and the second one to another paper, published in 2017 by the same researchers.⁴ At the time, it felt as if they had pioneered a new scientific frontier: the neural reconstruction of dream images. Through an innovative synthesis of functional MRI (magnetic resonance imaging) and machine learning, they captured dream content by awakening subjects during the hypnagogic state, asking them to recall their dream images and then training decoders on the fMRI data correlated with visual stimuli from lexical image databases such as WordNet and ImageNet.

In a third YouTube video, showing the protocol of the experiment, the subjects are lying inside an fMRI scanner in a dark and quiet room, their heads stabilized to minimize motion, while simultaneous polysomnography (PSG), using electrodes placed on the scalp and face, monitors sleep stages.⁵ They wear headphones through which the researchers can wake them by voice, and a microphone captures their spoken dream reports immediately after awakening—this process is repeated around two hundred times per subject without any invasive procedures.

By using machine learning algorithms to analyze patterns of brain activity in the visual cortex to identify what visual information a person is perceiving, the researchers achieved statistically significant accuracy in predicting dream content categories. Most notably, in visual areas, the decoded patterns aligned closely with the participants’ verbal reports, offering compelling evidence and laying the foundation for the future mapping of mental imagery during sleep.

Image reconstruction techniques have advanced from simple decoding and mapping brain activity in specific visual areas to sophisticated methods that use deep neural networks and generative models. Today’s methods translate brain signals into latent representations, producing images through tools such as generative adversarial networks and diffusion models. These techniques inch ever closer to externalizing subjective perception, promising profound applications in psychological assessment and brain-computer interfaces. Yet, despite advances in fidelity, these systems remain vulnerable to biases—shaped not only by data, but also by the inherently non-neutral perspectives of the humans who design and train them—and ethical dilemmas, particularly those concerning privacy and consent.⁶

Today, a broader, more international and experimental body of scientific work —extending from China to the Netherlands, Iran to the United States—is driven by a shared ambition: to render visible the elusive images of our minds, and to construct a direct audiovisual conduit between brain and machine. It is a revival of an age-old aspiration that found early expression in the photographic experiments inspired by the invention of X-rays, and that was later fictionalized with striking clairvoyance in films such as Kathryn Bigelow’s Strange Days (1995). The history of techno-science is not linear or progressive, but recursive, redundant, marked by returns and reanimations. Our present is never fully present, as its very condition is haunted by the repetition of unresolved pasts and the ghosts of lost futures. We can foreground the circular and kaleidoscopic nature of technological imagination—the human capacity to envision and project new speculative possibilities through innovations in techno-science—as it seems to return again and again, over the centuries, to certain fantasies, which act as magnets generating their own gravitational fields.

In the spectral logic of history, what was once dismissed as pseudoscience or magical thinking may return as the architecture of tomorrow, as if science were nothing but an ouroboros endlessly eating its own tail. The fictions that organize our present may be deconstructed by refusing to dismiss the ghost, by refusing to claim closure and declaring the past settled, by remaining attentive to what has not been fulfilled. We must then regard contemporary neural decoding experiments not merely as groundbreaking and disruptive feats of computation, but as echoes, resonances of an enduring historical desire that has not been resolved. Indeed, at the turn of the twentieth century, a proliferation of scientists, mystics, and eccentrics attempted to imprint their thoughts upon light-sensitive plates. All sought to capture the immaterial and invisible with the aid of emerging photographic techniques and/or other elaborate apparatuses.

In France, two fervent believers, Hippolyte Baraduc and Louis Darget, devoted themselves to the photographic capture of invisible “V-rays” with “V” standing for “vital fluids.” They invented the “portable radiograph”, a case with a sensitive plate attached to the forehead with a headband, designed to receive “psychicones”—a neologism born from the combination of psyché and icone—luminous and vivid images of thought. Darget, a French army officer and amateur photographer, sought to photograph the emanations of thought itself. To him, the mind radiated a tangible, nearly material force. “When the human soul emits a thought,” he wrote, “it vibrates the brain, which radiates the phosphor it contains, and the rays are projected outwards.”⁷ In one experiment, he attempted to project the mental image of a bottle onto a photographic plate. “It seems,” he recorded, “that the bottle shape I deliberately held in my brain was projected onto the plate, luminous, crossing the cranial cavity like X-rays.”⁸

In 1896, Darget turned his eye toward sleep. By placing a photographic plate on his wife’s forehead as she slumbered, he obtained an image under which he inscribed: “Photograph of the dream. The Eagle.” A 1913 article in The New York Times⁹ recounted how this discovery had stirred fascination in scientific circles in Vienna and Berlin.

Baraduc, a specialist in electromagnetic therapies at Salpêtrière Hospital and disciple of the neurologist Jean-Martin Charcot, Darget’s collaborator and spiritual kin, would later publish The Human Soul: Its Movements, Its Lights (1913). This book detailed both his techniques and the numerous photographs and experimentations they produced. He cautioned against reducing existence to what the human eye could perceive, insisting that true knowledge required penetrating to a deeper, “supra-terrestrial plane.” To photograph the invisible was, for him, a philosophical and spiritual pursuit—a way to glimpse life in its entirety, to reconcile science and mysticism.

These photographs—ghostly, magnetic, obscure—defy empirical scrutiny, yet captivate the imagination. Their significance for us lies not in their scientific accuracy, but in the questions they raise. Their ambiguity, their resistance to definitive interpretation, granted them a speculative power: They enacted the dream of making the mental visible. Were the abstract images of the Japanese scientists already beating in the heart of the “psychicones,” in a virtual state waiting for its actualization? As Franco “Bifo” Berardi reminds us with his concept of “futurability,” the present vibrates with latent futures: “Futures are inscribed in the present as immanent possibilities, not as necessary evolutions of a code.”¹⁰ The ambition of the two Frenchmen is now being technologically updated through artificial intelligence: There exists a spectral continuity between Baraduc and Darget’s “psychicones” and the outputs of today’s neural decoders. Is this “oniroscopy”, once conjecture, a poetic science at the margins, now edging toward material feasibility? And what does this obsessive scientific utopia, spanning decades and geographies, tell us of our relationship with vision and our reliance on it?

Seen in this light, AI neuro-technological breakthroughs and Baraduc and Darget’s luminous plates are not so distant. Both are animated by the same ambition: to breach the boundary between the visible and the invisible, to give form to what escapes representation. Where does this (almost paranoid) obsession come from? This desire to capture everything in images, to eradicate anything outside, or beyond, the visible field of view? Is this not the first step toward a future without imagination, without privacy, without even abstract thought?

Contemporary research in neural decoding could lead us to wonder whether technology is, as the sociologist Max Weber believed, merely an instrument for disenchanting the world and emptying it of fiction and fantasy. Indeed, in the case of Darget and Baraduc, the photographic plate was still a techno-magical instrument that opened the door to some kind of mystical ecstasy or meaningful ontological experience, proving the existence of a different spiritual reality, hidden to our eyes. Today, artificial intelligence could help us completely erase the invisible—that is, provide every thought and idea with a visible copy, completely fulfilling the tyranny of transparency and literality we already live in, and closing off the infinite possibilities of imagination to finite possibilities of vision. The dream of total visibility risks collapsing the fertile ambiguity of thought into the tyranny of representation. But some domains demand not exposure, but opacity: To preserve the possibility of dreaming, we must defend the right not to be seen.

If, however, there is something that crosses and unites the psychicones with their contemporary update, daughter of AI, and which could maybe give us some hope, it is how this corpus of images shows the limits of legibility and visual interpretation. The relationship between information and noise in the images obtained by the experiments invites us to consider the role of the “will to see” in all this. How much is a projection of the eye that looks, and how much is really visible? How does the willingness to believe in technological breakthroughs operate in the representational capacity of these images? Perhaps it is precisely in what escapes these visions—in the noise, in the blur, in the abstraction—that the last vestiges of imagination, and freedom, reside.