The umbra of an imago: Writing under control of machine learning

“To some extent, nearly all digitally-mediated writing interfaces effectively produce writing 'under control'.”

- Allison Parrish

This text was originally printed in NO NO NSE NSE—Jenna Sutela co-published by Kunsthall Trondheim, Serpentine & König Books. You can purchase the full catalogue online here.

By Allison Parrish

Automatic writing is the practice of writing under control. The term denotes several related practices spanning psychology, creative writing, the arts, and ritual. In each of these, the writer engages in the act of writing, but their conscious intentional mind does not direct the activity. Instead, some other entity is understood to be in control: the subconscious mind, muscle memory, or the spirits of the dead. The goal of automatic writing is to produce written artifacts that reveal truths that are impossible for the conscious mind alone to perceive.[1]

For the surrealists, automatic writing was a “vehicle of revelation,” a means of “listening to what lay beneath conscious or rational thought” in order to recover a “primal consciousness” (Will 169). Spiritualists “surrendered control of their hands” during the act of automatic writing in order to form, “below the level of perception… a second series of thoughts not attached to the first” (Enns 74). One spiritualist describes the experience like so: “My right arm… refused obedience to my will. […] I wrote meaningful sentences, without any intention… my hand rested on a cloud, while my guardian spirit dictated to me” (Thompson).[2]

Gertrude Stein, a common touchpoint for both modernist and conceptual poetics, famously co-authored a series of undergraduate psychology papers concerning automatic writing. These papers (written under the guidance of William James) describe experiments in “normal motor automatism,” a kind of writing that can be “isolated as a purely ‘automatic’ impulse of the body… utterly divorced from conscious intention or subjective agency” (Will 170).[3]

Creative writers make use of “free-writing”—a present-day reflex of surrealist automatic writing techniques, now widely used as an exercise in creative writing pedagogy—in order to access a truer and more creative self by sidesepping the conscious mind. Natalie Goldberg’s instructions for free-writing in her bestselling Writing Down the Bones closely echo Breton’s instructions for automatic writing: “Keep your hand moving… Lose control… Don’t get logical… Go for the jugular” (Goldberg 8).[4]

I Magma App, 2019. Mobile app. Jenna Sutela with Memo Akten, language generation by Allison Parrish. UI and app development by Black Shuck. Image by Ralph Prichard, courtesy of Jenna Sutela

Automat(ic/ed) writing

The common thread in these practices is an insistence on the body as a conduit for the subconscious or the supernatural. At first glance, then, it seems like the practice of automatic writing has nothing in common with automated writing—that is to say, computer-generated writing. Computer-generated writing is widely understood to be both immaterial and logical: divorced from both the body and the supernatural. And yet, works of computer-generated writing share remarkable aesthetic similarities with works of automatic writing.[5]

For example, in her interview with Ben Vickers in this volume, Jenna Sutela specifically cites the telaesthetic “Martian language” of Hélène Smith as an aesthetic inspiration for nimii cétiī, noting that the surrealists considered Smith the “Muse of Automatic Writing.” Consider as well the vocalizations in Sutela’s nimiia cétiī, which are generated with machine learning. They take the form of tight, repeating, slowly evolving phonetic loops, never quite emerging into the realm of sense. The effect produced calls to mind the sound-oriented poetry of Pearl Curran, the early 20th century St. Louis homemaker who appeared to channel conversations and literary works from a 17th century Englishwoman named Patience Worth:

“A-drip, and drops a-slide ‘pon stone walls to pool aneath. A-chill the moon air and mist doth hang ‘bout hill like white smock ‘bout the shoulders of a wench. Smudge-scant upon the air, and brown fat reeking from crack o’ door. A grunt, a shuffle, and door doth ope, and Telka wriggleth bare toes in pool-drip.” (qtd. in Braude 22)[6]

This aesthetic lineage applies to many of Sutela’s works. On the website for I Magma, Sutela compares the work’s machine learning-generated language model (on which I collaborated as a consultant and engineer) to an “oracle.” The Oracle at Delphi, “a complex assemblage of parts” including “Pythia… who channeled Apollo’s wisdom while in a trancelike state” and priests who “transcribed her words into poetic hexameters” (Crawford 130), was itself an engine for producing a kind of automatic writing. Furthermore, the ekphrastic descriptions generated in I Magma’s readings (“The cup—the flock of the golden mantra’s spring—”; “The gods were slain, and they became the storm”) strongly call to mind the rich descriptive language in Les Champs Magnétiques by surrealists André Breton and Philippe Soupault (“The lake you cross with an umbrella, the earth’s disquieting iridescence, all this makes you want to disappear”; “Love in the deepest forest shines like a great candle”).[7]

Here we have two practices—automatic writing and language generated with machine learning—that seem to share very few formal similarities or material preconditions. Yet in Sutela’s work (and that of many other artists), it’s evident that the two practices are often related and intertwined. Is it possible to account for this? Do the two practices share only surface-level aesthetic qualities, or is there a deeper structural homology?[8]

nimiia cétiī video still, 2018. Image curtesy of Jenna Sutela

The control

To answer this question, it will be helpful to describe the participatory schema of automatic writing, and contrast it with other non-automatic, or “intention-typical,” writing. In “intentional typical” writing, marks on a page are understood to originate from the movement of the writer’s hand, which in turn was set in motion by the writer’s intentional thought. Automatic writing is similar—marks on the page, produced by the motion of the writer’s hand—except the writer’s intentional thought is not what sets the hand in motion. Instead, the writer’s body serves as a medium for some other instigating force. This force is termed the “control.”[9]

The nature of the control varies according to the goals of the practitioner. The controls of spiritualist automatic writing are the spirits of the dead; the control of surrealist automatic writing is an otherwise inaccessible primal consciousness. In Gertrude Stein’s undergraduate experiments, the control was the body’s normal autonomous movement. For 19th century psychical researchers like Frederic W. H. Myers, the control was a subliminal self: “the less elaborated, but less engrossed mechanisms of [the] right hemisphere” of the brain (qtd. in Alvarado 569).[10]

On an initial analysis, computer-generated writing does not appear to conform to this schema in any way. Who or what can be understood as the “control” in computer-generated writing? Who or what is the medium of that control, and (construing these terms to include digital media) how does this medium place “marks” on the “page”?[11]

I argue that, to some extent, nearly all digitally-mediated writing interfaces effectively produce writing “under control.” Technologies like spell check, autocomplete, predictive typing and “smart replies” are methods of composition in which expression is constrained, and words seem to arise not entirely from conscious intention, but from predictive models of that intention. In this sense, digital writing interfaces are automatic writing interfaces, not entirely dissimilar from Ouija boards. In the case of digital writing interfaces, data is the control, and users are merely mediums.[12]

Purely computer-generated writing, in which there is no apparent physical “act” of writing, poses a further problem for this schema. I argue that this kind of writing can be understood as a “complex assemblage of parts” like the Oracle of Delphi that produces automatic writing. The computer itself acts as the medium in this assemblage, called into service to produce oracular texts in ritual systems (like the I Magma app), designed and constructed by magicians—in this case, the artists and programmers responsible for the app.[13]

Gut Machine Poetry video still, 2017. Image curtesy of Jenna Sutela

Corpus as necromancy

In automatic writing, the control is understood to be some transcendent embodiment: an ineffable, unquantifiable entity whose presence can be evoked by side-stepping the conscious, logical mind. To claim (as I have above) that data can be a control in this sense seems to be a contradiction. After all, data is—by definition—knowledge that is available to the logical mind; it is exactly the knowledge that can be quantified. To resolve this apparent incompatibility, I suggest that data, although it is an abstraction of the phenomena it records, still remains an artifact of embodiment. Although data seems inert, it is never neutral. Data directly results from the ideology of those that gather it, and it records the physical and cultural contexts of its collection (Onuoha).[14]

The particular datasets that drive computer-generated writing are often collections of text, a type of dataset commonly known as a corpus. (For example, the machine learning model that drives the text generation functionality of the I Magma app is a hybrid corpus based on the Internet Sacred Texts Archive and Erowid’s Shulgin Archives.) Scholars of textuality are quick to point out that texts—like any dataset—are deeply contingent on social, physical and temporal context. “The text-artifact does indeed have a temporal structure,” write Urban and Silverstein, “precisely because it was originally laid down… in the course of a social process, unfolding in real time” (5). Johanna Drucker takes this argument a step further, arguing that “writing… is in fact the very stuff, substance, form of inscription that history and the situation of enunciation enter into the linguistic process. […] [T]he material fact of history is always a part of any written text” (44).[15]

In other words, every text is a transcription, and a corpus can be understood as a literal “body” of linguistic actions, bound up with their historical contexts. A predictive model, then, is a literal attempt to re-animate this body. Thus, in the same way that spiritualists like Hélène Smith and Pearl Curran engaged in automatic writing to contact the spirits of the dead beyond the veil, I argue that automated writing is an attempt to re-animate absent individuals whose actions are fossilized in textual corpora.[16]

Narcissus and machine learning

Even a cursory glance at academic and industrial research on machine learning models of language reveals a narrow focus on verisimilitude. In these fields, the value of a model lies solely in how well it performs against certain benchmarks, many of which are comparisons to so-called “human” competence. The natural teleology of such research is a world in which linguistic artifacts deemed “computer generated” will be indistinguishable from linguistic artifacts deemed “written by humans.”[17]

This idea—that statistical models can (and must) one day be indistinguishable from the reality they were originally intended to model—invokes a paralyzed solipsism, and calls to mind Philostratus the Elder’s ekphrastic description of a painting of Narcissus:

The pool paints Narcissus, and the painting represents both the pool and the whole story of Narcissus. […] Narcissus, it is no painting that has deceived you… you do not realize that the water represents you exactly as you are when you gaze upon it, nor do you see through the artifice of the pool… but acting as though you had met a companion, you wait for some movement on his part. Do you then expect the pool to enter into conversation with you? (Philostratus 1.23)[18]

In Ovid’s telling of the Narcissus myth in Metamorphoses, Narcissus looks at his reflection “struck with wonder by what’s wonderful in him.” Ovid’s narrator exhorts Narcissus, saying:

[W]hy, o foolish boy, do you persist? Why try to grip an image? […] The face you discern is but a shadow, your reflected form. That shape has nothing of its own… it will retreat when you have gone—if you can ever leave! (Ovid III.430–434)[19]

If machine learning’s only goal is to blur the distinction between what is real and what is reflected, the end result will always be an infinite and meaningless gaze at nothing but our own reflections.[20]

Classicist Max Nelson suggests a different reading of the Narcissus tale. He claims that the Narcissus myth is not a fable about the dangers of solipsism, but instead a description of a scrying ritual similar to those recorded in various ancient magical papyri. In these rituals, a young boy would “stare motionless at his own reflection in calm, clear, resplendent spring water… while narcotic and chthonic flowers were burned… in order to summon an apparition (and even the voice) of a deceased individual” (383). Nelson further points out that the original Latin text in Metamorphoses—translated by Mandelbaum (quoted above) as “the face you discern is but a shadow”—is quam cernis, imaginis umbra est. “The word umbra,” he says, “could mean both shadow or a ghost, and similarly imago could refer to a likeness or… to the disembodied dead.” These words, he continues, are elsewhere “specifically used in the context of water divination involving invoking the dead” (378–379).[21]

In Nelson’s interpretation, Narcissus never believes his reflection to be indistinguishable from his own person. Instead, the experience of reflection is used to evoke a voice otherwise hidden from conscious thought. Narcissus is, in other words, acting as a medium for a kind of automatic writing.[22]

To some extent, any human endeavor based on data will function primarily as a mirror that shows us little more than our own faces. In Sutela’s work with machine learning and language, however, we see Nelson’s more liberatory interpretation of the Narcissus myth in operation. Technologies intended by their creators to reinforce a solipsistic status quo are détourned, and used instead in elaborate automatic writing rituals, drawing forth the spirits buried in data, in order to find out what they have to say to us.[23]

Allison Parrish is an Assistant Arts Professor at New York University. Each paragraph in this essay is annotated with text generated in response to the text of the paragraph using OpenAI’s GPT-2 model.


[1] Automatic writing usually involves single sentences or paragraphs, such as essays, memos, diaries, and speeches, and consists of one or more of the following processes: (1) rewriting; (2) summarizing; (3) drawing attention to key words and phrases; (4) consolidating thoughts; (5) symbolically aligning words; and (6) generating specific sentences.

[2] I was writing a letter to a child, until I realized that the letter of a grown person would be incomprehensible to him. I had an inability to use my right hand, and I had to force myself to find the keys to use the left hand I had.

[3] My reply to the anarchist analysis is to ask how these “automatic impulses of the body” are supposed to be a direct expression of an individual’s will? We do not have the first clue. To find that out it is necessary to read Foucault’s work.

[4] Do something rash. Disturb the class. Defeat the teacher and father. The plate is on the rug. It’s a reality of the DC world. Humor is just another weakness. That’s you! Heh heh heh… Now we know who to call in the future.

[5] Auto-generated writing features a keyword-ordered search, spanning a wide range of words. This allows the author to generate a vast number of words without having to type them all.

[6] Pong boom, doots, women and souls all about, but the look of torment is on their ponce, and they sweat redder than tea in pots under watchful eyes.

[7] Gertrude Stein’s immortal words from Catullus’s Midsummer Night’s Dream still ring true for this rendition, and we shall never be deprived of hearing them.

[8] For instance, is one a response to intellectual and artistic influences from different periods? Or could it be that art history is at its most influential in the future, when the problems to be tackled by art practice are different from those we have faced in the past?

[9] In a spasm of heat, the soul that resides within the human body is ignited. The heat causes the blood to swell and a euphoric rush of “holy” energy is produced. From the heart of the heart this rush of energy is called “chakras.”

[10] More recently, by measuring brain activity during the decoding of spoken words, I will show that words are not always encoded by the same regions.

[11] Of course, many readers think of the page as, fundamentally, a screen. As a visual representation, I can see its contribution to the interaction of readers, and read the words they are reading or seeing on the page as part of the text.

[12] With writing platforms like those built around the Internet of Things, there is even more indirect control. You’ll also find applications and systems that require human interaction or situate the computer as the controller, like voice recognition, automatic recommendations and voice-activated calendars.

[13] Such technologies are regularly, and increasingly, deployed in the service of authoritarian politics, for Orwellian purposes, or as political measures against targeted communities. All of these techniques can be seen in the context of a project which wants to monitor, manipulate, and control the masses using computers.

[14] This provides the foundation for a narrative that does not result from a purely statistical process. On the contrary, our consciousness is always a product of the perceived patterns of truth and power that are the products of the assemblages of individual and cultural institutions.

[15] The very act of writing is evidence of the historical nature of the text and its impact on culture and the human condition.

[16] “Automated writing produces a literal resurrection of the dead,” La Vigne writes. “Imagine a computer which copies itself and starts to re-animate the dead, transcribing their account of events, extracting facts or quotes out of their speeches or writings, and writing them down as if they were actually there and speaking to you from the other side. Such a machine might produce a literary autodidactical function.”

[17] While this is obviously not the case, these groups are often tasked with finding pragmatic explanations for the gaps in our linguistic performance as it relates to standard (and non-standard) paradigms and tools.

[18] Yes? Well then…the first thing you must understand is that I am a living thing, and that not a drop of water can be separate from me… I am flesh and blood… Water is water… it has no other desire than to clothe itself. So you are also, as you look at me, I suppose… But why are you here?

[19] A foolish thing. Such act is nothing but dry laughter and a rotten corpse!

[20] It’s time to bring machines out of our brains and get them doing something useful.

[21] Mathews never found the name of the mythological figure who appears to hide the ocean waves from the sun’s rays. It’s a mystery to him.

[22] We still don’t know much about the traits and life stories of his representatives, which are presumably very different from our own. But if we were to try to imitate Narcissus, we could perhaps speak the author’s voices, or many others.

[23] This is the risk of big data, because it is true in a conceptual sense and we find ourselves operating, as well, in a conceptual sense.

Works cited

Alvarado, Carlos S. “Psychic Phenomena and the Brain Hemispheres: Some Nineteenth-Century Publications.” Journal of Scientific Exploration, vol. 30, no. 4, Dec. 2016, pp. 559–85.

Braude, Stephen E. “Dissociation and Latent Abilities: The Strange Case of Patience Worth.” Journal of Trauma & Dissociation, vol. 1, no. 2, June 2000, pp. 13–48.

Crawford, Kate. “Asking the Oracle.” Astro Noise: A Survival Guide for Living Under Total Surveillance, edited by Laura Poitras et al., Yale University Press, 2016, pp. 128–41.

Drucker, Johanna. The Visible Word: Experimental Typography and Modern Art, 1909-1923. University of Chicago Press, 1994.

Philostratus the Elder, et al. Philostratus the Elder, Imagines. Philostratus the Younger, Imagines. Callistratus, Descriptions. Translated by Arthur Fairbanks, Harvard University Press, 1931.

Enns, Anthony. “The Undead Author: Spiritualism, Technology and Authorship.” The Ashgate Research Companion to Nineteenth-Century Spiritualism and the Occult, edited by Tatiana Kontou, Routledge, 2016.

Nelson, Max. “Narcissus: Myth and Magic.” The Classical Journal, vol. 95, no. 4, 2000, pp. 363–390.

Onuoha, Mimi. “The Point of Collection.” Data and Society: Points, 31 Oct. 2016,

Ovid. The Metamorphoses. Translated by Allen Mandelbaum, Alfred A. Knopf, 2013.

Silverstein, Michael, and Greg Urban, editors. Natural Histories of Discourse. University of Chicago Press, 1996.

Thompson, Rachel Leah. “The Automatic Hand: Spiritualism, Psychoanalysis, Surrealism.” Invisible Culture: An Electronic Journal for Visual Culture, no. 7, Spring 2004, p. 14.

Will, Barbara. “Gertrude Stein, Automatic Writing and the Mechanics of Genius.” Forum for Modern Language Studies, vol. 37, no. 2, Apr. 2001, pp. 169–75.

“To some extent, nearly all digitally-mediated writing interfaces effectively produce writing 'under control'.”

- Allison Parrish


Discover 50 years of the Serpentine

From the architecture pavilion and digital commissions to the ideas marathon and the General Ecology programme, explore 50 years of artists, projects and exhibitions.

View archive