Laughter. Poetic language. The bodily – language’s excess – is “invoked in language,” and as Caroline Bassett’s text, Feminism, Refusal, Artificial Writing, reminds us, “frustrate[s] linguistic purity.” These embodied exuberances “mock the dreams of one true code” and, “with its refusal to be accounted for, frustrates particular kinds of accounting with.” Though earlier and elsewhere Bassett has argued that these – even silence itself – might be “apt as a response to an expansion in the writing economy,” which comes to us as “AIs learn to create content with a language structure,” here Bassett argues for another mode of refusal. Not withdrawal or silence, but instead, looks to feminisms, bringing together scrutiny and a focus on collaboration; noise and signal.

Cyborg politics is the struggle for language and the struggle against perfect communication.⁠1

The words of Donna Haraway’s 1985s A Cyborg Manifesto echo on, finding new resonance as AIs, experts in processing code, which is their native tongue, augment their capacity for natural language processing, producing new forms of writing and speech by simulating those of humans.

As for the cyborg, it never spoke.

Haraway’s intervention turned on the cyborg’s illegitimate body, unmarked by race, gender, and sexuality, confounding categorisations that have done so much violence; the cyborg lives in a mythical time. Embodiment has been central in feminist struggles engaging in and around language. Hélène Cixous urged women to throw themselves, “unthinkable unthought”,⁠2 into words, but the loud silence of the unspeaking body might itself constitute a form of powerful resistance. A mode of refusal. The bodily, invoked in language, insisted on as language’s excess has frustrated linguistic purity. Laughter mocks the dreams of one true code, and poetic language, with its refusal to be accounted for, frustrates particular kinds of accounting with. Silence, Delirium, Lies. Elsewhere I’ve argued these constitute a response to platform monopoly, a form that gathers up and expresses the social relations of computational capitalism in general.⁠3 But silence may not be apt as a response to an expansion in the writing economy, which comes as AIs learn to create content with a language structure.⁠4

Writing this, I am aware of the quiet outside. Inside is a cacophony. “Life on the screen” is louder than ever.⁠5 And wordier. There’s Zoom, of course, hosting re-mechanised conversational turn-taking, where humans stare each other down, mute as bottles,⁠6 or tinny as preachers, hiding behind images of themselves. There is also the activity of bots, scrappy exchanges, continual small interrogations, a forest of words growing denser all the time. Finally, coming forwards fast are the big new players in these language games, including neural network-based general transformer model AIs, such as OpenAI’s GPT/2 and GPT/3, “learning for themselves”, claiming the capacity to produce sustained writing at a human standard. You can speak with them, write with them, co-produce with them. They’re the new naturals. They claim to be able to talk to the dead.⁠7 And they’re supposed to be dangerous to know.⁠8

Why can’t they break out the machine guns?⁠9

We are in the presence of a new kind of writing. We are beginning to hear new kinds of voices speak. If language is a prison house, the AIs are breaking in.⁠10 Feminists responding to this may wish to ask who might break out and who gets locked down more tightly. They may also ask how or whether the arrival of these alien writers should be welcomed or excoriated, and what of an earlier political project for socialist and intersectional feminism, which was always in part a project of language. Might it be rendered obsolete? Or demand a new configuration? What are the possibilities for generating an authentically hopeful and critical techno-feminist response to these developments?⁠11

A First Refusal

I begin by refusing to set aside what has come before, the recognition of which has been fought for. Hold on to the old grudges; they were justified. Generosity does not stem from forgetfulness.

The excitement around AI and language production subtends a larger field. Visionaries of AI futures (utopian or terminal) often assume that the relation between humans and increasingly intelligent machines is the only one that matters now since technology will render old antagonisms and divisions irrelevant—or fixable. This position, which has popular purchase, does not ask for whom this will be a utopia or a dystopia. Nor whether it will be free.

Interrogating this technologically fuelled vision and disputing its inevitability is a necessary first move, opening the way to make a feminist techno-politics that is affirmative but also critical.

Dissenting from the Automation of Consent

Consider the artificial voices inhabiting the platforms and the networks, the stupid bots and the little scripts, the agents requesting a response (score, accept, assent). The requests are continuous, and the bots are emollient. They care about my privacy; they will help with my searching; they will tailor my news. They ask me to prove I am human. I tick all the right boxes and confirm I am. They speak my language (better and better; that’s the point). They require my consent to do what they do. I give it. Or I don’t. It’s almost compulsory but not quite. The bots will mindlessly keep on asking for it anyway. It’s not that I am thinking much either.

Surfacing these small exchanges and automations matters. The intimate is the scale that I am ordinarily (with extra-ordinary precision) addressed through already, where ‘my’ truth, ‘my’ digital assistant, ‘my’ world is offered to me. That might encourage epistemic complacency—or a form of collusion. Dividuation loves the crowd and hates the collective, whose voice, whose decisions, it cannot predictively pre-empt and direct.⁠12 The infra-ordinary, operating at an oblique scale, can hook out the unlooked-for in the everyday.⁠13

We care about your privacy: I accept

The proliferation of automated exchanges is startling. The sheer volume of those offering control over the degree of data capture and algorithmic going-over undermines the offer itself. But the complexity of algorithmic processing and the length of the extractive chains that will take my ‘donated’ data meant it was always hollow anyway. The strong suggestion, the operating environmental stimulus, administered in words, as scripts, through code, is to abandon conscious decision-making in favour of absent-minded agreement ‘going forwards’—there is no need to write back. We’re already quite a way into the kind of dazed functional sort of behaviour that the philosopher Hannah Arendt feared would result from widespread automation.⁠14

Perhaps this is why letting my writing be automated, giving my voice to machines, doesn’t seem particularly odd. The internet is full of words and writing, says Richard Seymour in The Twittering Machine, and more and more and more of it is generated by AIs.⁠15 We’re taking this on board rather absent-mindedly. Or I am. If there’s a struggle for language on my mobile, the language algorithms are not winning, but most of the time, I’m not fighting either. I mostly prefer not to. I know I could use my own words, but often a predictive text suggestion will do. What might I have said next? The Markov chains oblige. I accept their word over mine. The message you got from me was, in part, a message from an AI. Happy Birthday. It’s lucky one of us has an exteriorised memory that includes a calendar. Next year it will send a self-generated haiku.⁠16 As for me, I’ll be out of the loop. It’s not the thought that counts, these days.

One day I am studying math and notice I have a new key. It is the key to life.⁠17

From fragments of artificial writing and bot scripts, back to the matter of ‘life itself’. At this scale, the dominant discourses shift from the political economy to the ontological. Machines with the capacity to address us in writing (to process natural language at a very high level) may be closer to being ‘like us’ and to being intelligent ‘like us’. The Turing test turns on an AI passing a gendered linguistic exchange.⁠18 Language has been named the holy grail of the AI research community.⁠19

The excitement about synthetic writing partly arises because of the prospect that machines, by crossing one line into human language, might be approaching another. If language is accessible, is life? Ontological questions shade into existential ones (for humans) when intelligence is adumbrated as a fixed economy, with only so much cognitive liveliness available. Language, then, plays a central role in intelligent AI debates because it is regarded as a key marker of human distinction, with writing considered crucial to memory and history, and also as a marker of machine intelligence.

The way (human) language is invoked—at once rarefied and abstracted within these frames—must be contested. Through language, we are thrown into the world.⁠20 A world in which language is fought for and over and is a powerful tool of dominant groups. Struggles around writing are integral to this. Recent work has explored how literacy was at once denied African American slaves and invoked as a measure of the fully human.⁠21 To withhold literacy was at once to reify and withhold one of the tools of memory, continuous with that more general destruction of history and collective memory that produces a profound alienation.⁠22 This work is invoked here because it underscores the epistemic and very real violence of attempts to corral humans into one category, in language, and against them, machines. Language, language technology, communications technology, technical media; the operations of power are continuous and continuously differentiate between humans.

These are propositions feminism has explored extensively. Of particular relevance here are techno-feminist interventions that see, in computational technologies, the possibilities for disrupting gendered relations of speech. Haraway’s attack on the ‘one true code’ in the 1980s had, in its sights, Lacan’s cybernetically inspired account of the master signifier.⁠23 Ironically, the illegitimate cyborg escapes the law of the father by virtue of its (also) cybernetic being. Other interventions have pitched humans clean out of the symbolic order,⁠24 defined the digital as new grounds for a genderqueer performative politics,⁠25 used it to question the autarky of authorship,⁠26 insisted on writing’s intrinsic technicity, and affirmed rationality whilst also invoking the force of laughter (the irrational) to find new ways of writing with machines. Audre Lorde’s famous claim that the master’s tools cannot dismantle the master’s house always had a rider; she used many of them anyway and to explosive effect.⁠27 Others have followed.

Transformers?

Turning back to the new AI text generators, which slip between the existential scales and temporalities invoked around artificial life and the infra-ordinary scales of the bot fragments and scripts already instantiated into everyday life, how are they being adumbrated, and how might their potential be understood? Smarter, by far, than what has come before (but still stupid, as Janelle Shane notes), language processors like GPT-2 and 3 work using pattern recognition.⁠28 Effectively they predict what needs to come next to generate a text. Disrupting some of the claims made for these developments involves undermining a set of circulating assumptions about the degree to which what they write is pristine, unpolluted by human values, and suggesting ways in which their alienness can be explored as generative.

Autonomy/Boloney

These new AI language models (pre-)train through exposure to giant collections of speech and discourse. They may then be tweaked on smaller data sets—like feminist manifestos, early modern literature, and medical data, for instance. Copies of the vast lakes of discourse that constitute the partial, located, Western-biased internet are the AIs’ world. They might be said to operate within a concrete ideology, unmoored by bodies that might check discourse’s normative claims, force a realisation of violence done, or register as absences.⁠29 Developers recognise that their AIs lack “real world grounding”.⁠30 Divorced from a material environment, but as humanly informed as they come, AIs write into the future, buttressed by the crudity of the present, using a learning set that has already condensed a series of biases, and amplifies them. Let’s hear it for the uncommon crawl. Meanwhile, it is tempting to withdraw to a safe corner.⁠31 But there is more to say about synthetic writing.

Oddness/Generation

Janelle Shane’s observations of AIs building scripts reveal a peculiarly oblique process of recognition and construction.⁠32 AIs work at strange scales, catching themselves up in loops for reasons that can’t be discerned, developing texts by way of logics unreadable by humans. This is language knocked together from the outside.⁠33 At times the process is resonant of glossolalia, a form of speech where language is promised but refused; things that almost make sense don’t, the blurred outline of a convincing shape appearing but breaking up. Delirium might be back by another route.

How long will they speak in haiku and apothecary’s delight?

What is revealed, though, is not (only) something that is less than human (language) but rather the surprising workings of an alien logic. Image recognition AIs have been known to join contexts to foregrounds. A ruler becomes part of particular cell culture, reflecting a preponderance of images with scale indicated in training materials. Cells have rulers. And kitchens have women (yet another example). These can be described as training errors, but are in another sense, perfectly reasonable moves, given the source material available.

The insight provided here is not only into AI operations; an alien perspective disturbs naturalised human presumptions about boundaries, scales, edges, limits. But there is more here to be explored than the deconstructive force of these interventions. What is produced is sometimes intriguing for itself, not quite expected; something new. You could call these productions serendipitous. But that reduces computation to chance. And this won’t quite do.

You must be a human to open an account.⁠34

From a computer science point of view, the proposition that something new may be found in artificial speech is heretical. Simulation is possible because language can be processed algorithmically, but there is nothing creative about this. The interest here, however, is not in exploring creativity, with all of its human connotations.⁠35 And I am certainly not claiming that these AIs are creative beings.

What is worth exploring is what simulation entails in relation to these writing processes. To do this, I want to invoke the fictive, using it to supplement the idea that simulation is reducible to copying. It also involves making. The fictive is defined in terms of feigning, but also in terms of an as-if—or a making.

And the sea, And the sea, And the sea, And the sea.

In the Turing test, the criterion of intelligence is the successful simulation of a gendered conversation, but the criterion for passing the test is passing, not a matter of exactly copying, but of being able to pass as. There’s a break here. Katherine Elkins et al. ask whether GPT technologies can even pass the Turing test.⁠36 The answer they find is equivocal—not surprising since the test itself is ambiguous⁠37 because it rests on passing⁠38 and appeals not only to the criterion of simulation as perfection of the copy but on an imitation or performance that isn’t measurable by the same rule.

Do you want to just practice being a good person? Then I am your teacher. Hello, Hello?

The fictive, which combines what is made and what might be created with what is feigned (or simulated), might be adopted as a way to probe synthetic writing—to search for its ‘as if,’ for what is produced that is not ‘copied’ exactly. This might be found by exploring the peculiar processes of AI writing under construction or focusing on those moments when simulation as a reasonable copy fails. If synthetic writing has a discrete aesthetic or partakes of the aesthetics of the discrete,⁠39 then glitches might be places where this alien generativity may be discerned and explored.

StoryAI: “at first a human said”. “an AI continued” … ⁠40 ⁠41

It needs to be taken into account. Not in the interest of declaring a new form of machinic autonomy or independence for the ‘writing machine’—this would be ludicrous when the matter on which these AIs are trained is re-invoked apart from anything else. Instead, scrutinising the stakes of simulation and considering generation in AI writing is necessary to explore and make demands around how synthetic writing can be a co-production. A key part of this concerns identifying the shifting balance between different kinds of agency found on all sides in artificial writing operations. Technique is not on one side and inauguration on the other.⁠42 The sides here are not reducible to human versus machine, nor a matter of two different ontologies of writing. A second necessity is to reach beyond the conception of machine and human intelligence as a closed system—in which machines writing more automatically means humans must think less.

More Noise

Engaging with the emerging politics of AI and machine learning, what is needed is not withdrawal or silence, but scrutiny, on the one hand, and a focus on collaboration, on the other; noise and signal. But there is still a form of refusal here. It is central to a politics of writing that sees potential in a proliferation of speakers and writers, that celebrates and exploits the disruption of an alien intervention into gendered, classed, sexed relations of speech that might produce, but also generate, a critique of the ways these technologies are being instantiated; we need to refuse to give writing away.

I am hugely indebted to fellow Cambridge ghost-writers Anne Alexander, Joe Walton, and Alan Blackwell, and to those taking part in the CDH workshop series on synthetic writing that we jointly organised. Many thanks also to Elise Hunchuck for her incisive editing.

Caroline Bassett writes, researches, and teaches at the University of Cambridge and leads Cambridge Digital Humanities. Her research explores the intersection of technologies and cultural forms and practices. She has written widely about gender and technology, refusal, future technologies, and media archaeology. She is co-author of Furious (Pluto, 2020), amongst other things.