This project draws upon several areas of theoretical discussion. The first is a growing body of work gathered under the title ‘material feminisms’,9 work that argues we should not divide human corporeality from a wider material world, but should instead “submerse the human within the material flows, exchanges, and interactions of substances, habitats, places and environments”.10 This thinking is redefining (hegemonic, Western) understandings of culture and nature, and human and more-than-human worlds. Rethinking the nature of being in this way, (new)11 material feminists argue, is an ethical imperative; we have ample evidence that the existence we’ve created is not ethical — climate change, species extinction, wealth inequality, etc. — and the piling up of that evidence forces us to imagine a different existence.12

The second area of discussion deals with questions about AI, what it means to imitate human cognition or to construct a new intelligence, and where its limitations are. The third surrounds the fanfiction community and its reuse of existing texts: who forms this community and why do they write. The fourth is a technique that we use as a reference point in our own reuse of texts: the cut-up technique, dating back to the 1950s. Each of these contexts will be given brief consideration in this review. Together, these areas orient our ontological, epistemological, and methodological choices in this project.

Material Feminisms

Material feminisms thinkers include Karen Barad, Donna Haraway, and Nancy Tuana, who collectively work to disassemble dichotomies between the human and the non-human, the natural and the cultural, or the material and the discursive, drawing our attention instead to the relations that exist between and within these entities.

Barad’s concept of ‘intra-action’, for example, describes the world not as made up of pre-existing individual entities (objects, subjects, bodies) but of ‘intra-acting phenomena’ that only become determinate, material, and meaningful through relations.13 This mode of thinking challenges the dominant Western idea of the coherently embodied individual subject and related epistemologies that focus on discrete objects of consideration, and instead draws our attention to the margins within and between bodies and the rich muddle of intra-actions that happen there. Haraway makes a similar point through her discussion of “material-discursive” practices,14 as does Tuana with “interactionist ontology”.15 Each of these concepts calls into question the nature of agency and its presumed localisation within individuals (whether human or non-human), moving us away from anthropocentrism and human exceptionalism and helping us to consider the complex and ever-emergent relations that exist between and within all things.

Engaging with this thinking has helped us to work productively and sensitively with collaboration as both a concept and a practice. When the world is understood as fundamentally entangled, collaboration is not the product of inter-acting entities but is the ever-emergent process of the world-becoming. This has played out in our research approach, which we treated not as a process of gathering data but as a process of “becoming entangled in relations”,16 focusing on what anthropologist Tim Ingold would describe as “knowing as we go, not before we go”.17

Further, with intra-actions drawing our attention to the margins within and between bodies (human, non-human, inorganic), emphasis is placed on “the entangled materializations of which we are [all] a part” and the topologies of power these reveal.18 For example, “If humans have no separate existence, if we are completely entangled with the world…then we are completely responsible to and for the world and all our relations of becoming with it”.19 However, we are also keen to pay attention to existing power asymmetries that exist and persist within and between human communities, not understanding ‘anthropos’ as a homogenous mass, but as intra-acting phenomena that have their own specific material configurations influencing their ability to act and respond — what Donna Haraway neatly terms “response-ability”.20

Through the process of this project, our aim is to make sense-able some of the sites of sticky friction, resistance, and opposition that are inevitably produced in this context; we aim to represent “viscosity” as opposed to “fluidity”,21 to both participate within and describe these sticky and contested relations.

Art practice research is uniquely positioned to engage with challenges such as these, as observed by Graeme Sullivan:

“Artistic practice undertaken in a digital environment is giving rise to research that is no longer challenged by questions about the human condition but is challenged by the need to revise what it is to be human… Cyberspace is radically altering these notions of individuality as modernist conceptions of identity grounded in traditional psychological perspectives are being replaced by a reflexive and decentered sense of self.”22

Granting a computer permission to dream allows us to experiment with creatively representing the perceived alterations taking place in the socio-cultural and ecological networks and assemblages that we are entangled within. It also allows us to encounter the outcomes of these negotiations in a form that lays itself open to interpretation and aesthetic consideration, inviting human minds to experience and construct a relationship to something that is often carefully camouflaged to fit into everyday experience.

AI, computers, and thought

The idea that the human mind might easily be modelled through simple decision chains is long dead, and questions still abound about where the limits of artificial intelligence lie (as in Hubert Dreyfus’ two wittily titled texts, What Computers Can’t Do, and two decades later, What Computers Still Can’t Do).23^,24 The legendary Turing test, the test of whether a machine is able to exhibit behaviour indistinguishable from that of a human, still towers over popular understandings of AI (the film Ex Machina (2014) is a particularly salient example25), though John Searle’s famous Chinese Room thought experiment26 powerfully made the case that even producing appropriately ‘human’ responses does not necessarily entail understanding, or a mind.27 An exciting possibility is that there is a real benefit to be had by focusing on the ways that computers diverge from human minds and find paths of their own. The possible advantages of adopting alternative tests in the realm of Computational Creativity is interestingly discussed in Pease and Colton, 201128).

Creative experiments with automatically-generated texts abound; an example relevant to the speculative nature of this project is SunSpring,29 the screenplay of which was written by an AI using an open-source algorithm that applies Markov models, efficient algorithms that work well with sequence data. The project has superficial similarities to our own, but they diverge in ways that demonstrate varying attitudes to AI writing. We use stories, rather than film scripts, and, importantly, work written by amateur writers rather than paid professionals, as an aim of this project is to bring something of these writers' methods into our own uses of the texts. Rather than using a play’s convention for emotive human acting to impose socially coherent meaning on an AI generated text, we embrace the incoherence of our outputs, allowing imagery and atmosphere to come to the fore.

This conscious attitude toward machine-made texts allows and examines its otherness. This has also been achieved in other applications, such as the Cybernetic Serendipity exhibition in 1968, which featured several text-generation projects. Most of these were presented as poems, an effective way to make sense of their strangeness. While much research in the field of Natural Language Processing focuses on attempting to better imitate human writing through different text-generation systems, presenting these texts as poetry means that the points at which imitation was imperfect, strange, and imaginatively fruitful becomes advantageous — the jumping-off point of the work rather than its limitation. Margaret Masterton and Robert McKinnon Wood’s computer-generated haikus are a particularly strange and exploratory example.30 Traditionally consisting of a set of objective statements in the present tense, human-written haikus produce a profound effect on the reader, requiring an enormously active and creative engagement. This means that the strangeness of the computer-written haikus becomes provocation for exciting interpretive work.

1 Poem

eons deep in the ice

I paint all time in a whorl

bang the sludge has cracked

2 Poem

eons deep in the ice

I see gelled time in a whorl

pffft the sludge has cracked

3 Poem

all green in the leaves

I smell dark pools in the trees

crash the moon has fled

4 Poem

all white in the buds

I flash snow peaks in the spring

bang the sun has fogged

Computerised Japanese Haiku by Margaret Masterton and Robert McKinnon Wood

This raises a question: why should computers be expected to write in a way that is comfortable and straightforward for us to read? Why is it that the most famous evaluative criterion for an artificial intelligence, the Turing test, is that it should fool us? Should we be placing value only on a computer’s ability to slide undetected into our lives above its ability to show us new things? The Turing test assumes that the appearance of humanness is all-important, while, in a material feminist way of thinking, this anthropocentrism is precisely the thing we want to get away from.

In the case of Sunspring, production work like set design, costumes, props, and human acting allows a basically very strange screenplay narrative to slide comfortably into the expected format of silver-wrapped sci-fi. So much work is done to overcome the nonsensical nature of the text that the end result misses an opportunity. If we are to work with non-human elements, surely there is a value in allowing them to show us things that we do not expect to see. What can be learnt from the experience of otherness if we no longer try to overcome it?

Fanfiction

We make use of fanfiction as source material in this project, and so here we attempt to summarise our understanding of that world and how we position ourselves in relation to it. In general terms, fanfiction is “writing that continues, interrupts, or just riffs on stories and characters other people have al­ready written about”.31 It has a long history that far precedes the internet age (see Coppa, 2006 for examples32), but it is its contemporary online manifestation that is of interest to us here.

In Code: Version 2.0,33 Lawrence Lessig identifies several features of the architecture of online communities that influence the kind of identity and credibility that can be constructed in that world, including anonymity, the kind of identity constructed, the kind of commentary and feedback that is allowed, and so on. Its supportive and collaborative nature means that fanfiction lies not on the sharp edges of internet culture but in a constructive place, produced by people sharing their own work and offering supportive feedback on others’ creative endeavours. We can see the effect of architecture on the workings of this community, just as Lessig describes; the written-only feedback system and lack of connection to real-world identity create an environment in which writers are free to dream and can expect positive responses from their peers.34

It is difficult to give a precise demographic of an online community, with thousands of individuals choosing to share limited information about themselves, but the fanfiction community is often characterised as “feminine, feminist, and /or queer”,35 “written by women, or if not by women, then by people who are willing to be (mis)taken for women”36 (for some demographic information, see Charles Sendlor's data on FFN research37). Also of importance is that fanfiction is usually not published for profit.38 These stories are published on public, searchable websites, making it an ideal place to find a multitude of marginal voices covering an abundance of topics. 

As Henry Jenkins would have it, fanfiction “is a way of appropriating media texts and rereading them in a fashion that serves different interests”.39 Beth Bonstetter and Brian Ott, in their narrative performance text on écriture féminine in fanfiction, claim that the rewriting is a “means for marginalized groups and especially female fans to construct a discursive space within hegemonic culture to express themselves in meaningful and personally fulfilling ways”.40 Though fanfiction writers draw upon popular published and produced texts that flood our culture, the stories told are given new twists and priorities, and the process "permits women to write alternative versions of existing narratives and characters that speak more directly to their feminine, feminist, and/or queer investments and desires".41 

Further, fan writers ignore the devices that the publishing world says make writing “serious” or “professional”.42 As Jamison describes, fanfiction “feeds on its predecessors and its contemporaries, interacts with them, makes them new. It is in a constant state of conversation and exchange”.43 The result is a rich, subversive, shapeshifting and interactive archive of narratives that both speak to and against their ‘master’ (narratives).

The fanfiction community has its own rich language, culture and classification system — Mary Sues, Canadian shack stories, high school alternate universes, mirror universe fics — that organises their engagements and describes typical forms of writing.44 While the community has its own abundant and often participatory structure around these texts, and though the community is engaged in the production of texts, it is also a site for discussion, co-creation, and construction of a worldview. As Jamison reminds us (and by that we really do mean us, the ‘academic authors’ of this work), “Academics can tend to emphasize fanfictions potential for collaboration, non-hierarchical relations, dissent and resistance […] Fic may be known for creating worlds in which anything goes […], but communities and individuals can police these worlds and their boundaries with tremendous vigilance.”45

Fanfiction writers are Henry Jenkins’ "Textual Poachers", fans that pro-actively construct alternative cultures using elements “poached” and reworked from the popular media.46 In this project, we extend this tradition, and try to write through these writers themselves; the poachers become the poached.

Our own experience of working with and through fanfiction is further discussed in Source Material.

The Cut-up Technique

Finally, we introduce a way of making that is a reference point for the interactions that we had with these texts. The cut-up technique was a reference point which grew in significance as the project developed. It is a technique made famous by the beat writer William Burroughs in the 1950s, but dates back further to the Dadaists in the 1920s. Cut-up is a literary technique based on chance, in which existing texts are cut up to create new texts, often poetry.47

When it was first developed, the intention of the cut-up technique was to create new poetic connections and find unexpected (often subversive) meanings in the abundance of texts found in mid-twentieth century lives. In the modern world, texts are more abundant than ever, and with internet communities meeting in forums online we have a rich resource of grassroots voices, offering a view into the psyche of the internet-using community that could not be found in the institutionally sanctioned writings of journalists and published authors. Our technique, therefore, makes use of super-abundant source data and a super-charged, machine-driven cut-up process, making new worlds and new imaginings from the texts of others.

We realised that what had begun as a metaphorical understanding of Artificial Intelligence programming became something we were playing with throughout the project. The films, with their identity scavenged from obscure corners of YouTube, cut and spliced together to create new stories, are another version of the cut-up technique, creating surprises and subversions through the juxtaposition of found material. This thread extends even further, into the way we chose to present our research. The use of different colours for each of our voices evokes a reversed cut-up aesthetic, where meaning that emerges from these juxtapositions is intentional. The cutting is a record of our approaches to that meaning. With the words on this website ‘cut-up’ at their inception, it is interesting to consider how different the communication of this project would be if our words, entangled and undisciplined as they currently are, were ‘un-cut’ and placed in neat blocks of red, blue, and green.

The intention of the cut-up technique is, and always was, to problematise authorship, to produce meanings while exposing their fractured origins, and to both invite and confound the reader’s process of interpretation; interpretation that expects — but will not find — human intentions on the other side. We instead set this process of interpretation against a cut-up process of mixed origin: of a human writing in collaboration with a computer, an algorithm to be executed by a computer on human-written texts, and again, invite an interpretative process on the part of the reader that is bound to be productive, frustrated, and a test of the process itself.

In accordance with the participatory and open-ended spirit of the project, we have made our code available for use and modification. It can be accessed here and here. Below we detail the basic workings of the text-generation algorithm used.

After collecting a suitable number of stories in each of the four categories, all stories were saved into text files. (Suitable in this case refers to a balance between the amount we could feasibly collect in a limited time frame and the amount required to successfully run the program so that it generates random stories from diverse sources.) As a pre-processing step, all punctuation marks were removed and all letters were changed to lower case.

All story texts were read by the program, then all trigrams (sequences of consecutive three words) in the dataset were extracted and saved for later use. We had also tested the program using 4-grams (sequences of consecutive four words), but eventually chose to use trigrams to comply both with the requirements that we felt ethical use of the dataset placed upon us (see [ethics] section) and our vision for the project (see [aims]). Because of the limited size of the dataset, using 4-grams tended to result in the reproduction of small chunks of text from one particular story, which we considered a problem in terms of fair use and of producing a result that could be said to have been written in full, albeit derivatively, by the algorithm.

The algorithm we used to generate the stories is as follows.

A random trigram (w1, w2, w3) is selected from a set of trigrams that start with specific selected seed words (single and pairs of words shown below). Then all trigrams that start with the last two words of the first trigram (w2, w3) are retrieved and a random trigram amongst them is selected. This means the first two words of the new trigram are the same as the last two words in the previous trigram.

For example, in one of the stories we generated, the story began with the randomly selected trigram “before i saw”, which was picked from a set of trigrams that start with the seed word “before” (also chosen randomly from the pool of seed words we input into the program). Then all trigrams that start with “i saw” were retrieved. For this example, the randomly selected trigram that the program picked was “i saw him”. This process continues until a story is generated. The length of the stories can be set by the user. Different stories were generated for each of the four different categories.

We needed to choose seed words for the algorithm to use to begin its search: an initial word or pair of words to form the beginning of the phrase that the algorithm then searched the dataset to complete. We used neutral seed words and seed pairs that would suit any category, like ‘if’, ‘when’, ‘in the’ and so on, plus additional seed words that we felt were suitable for specific categories.

Table 1 shows some of the seed words we used:

We found that, as the stories quickly moved away from the initial phrase, the seed words did not seem to strongly affect the content of the story written.

Python programming languages were used to develop the program and the Natural Language ToolKit (NLTK) was mainly used in pre-processing and to generate the trigrams.

In this process, the thematic content was stripped out and regurgitated, producing a new fanfiction: a computer’s fan writing, taking the writings of others and reanimating them, transforming the original content and adding something different of its own.

With the possibility of producing an infinite number of stories per category, we ran the algorithm 100 times for each of the four categories choosing one algorithmically produced story per category to work with. Choosing was difficult; every story had moments of intriguing lucidity and long, confusing sections with little sense to be made. But we chose and, adding a little punctuation, polished them into short stories.

Just as our harvesting process sidestepped the fandoms that organised the texts for authors, we found that our AI’s stories sidestepped the categories we had given to them. The four categories did not seem to imbue the sets of stories with four distinct characters, or at least not one that rose above the distinctively strange aesthetic that the stories had in common. What we found in their place was that the spectrum of imaginings that we had encountered orbited two main poles: one, a focus on disaster, decay, and environmental catastrophe; the other represented a sinister kind of progress that was hi-tech, robotic, militaristic and highly ordered. These opposing forces of entropy and order, and combinations thereof,
seemed to shape the imaginings of the future we found, and also seemed to divide the categories we developed in two: Apocalypse and Changed Environment roughly align with themes of entropy, and Another World and New World Order with themes of order and technology. This is reminiscent of a theme explored by one of the greats of speculative fiction, Philip K Dick. In his novel Ubik, the protagonists experience similar parallel strands, explained as opposing forces acting to either save or to wither them in their half-life, with the life-giving force embodied by the ever-present product Ubik. Here, we give no such explanation, except that it is common to imagine things as either progressing or degrading, but never staying still.

We chose two stories which particularly represented these two poles, and began to develop them into films.

With our strange and beguiling output stories in hand, we enlisted the help of another algorithm. This was instructed to search Google Images for each trigram in each story, using the setting that gives images that are free to use and modify,60 and select an image from each set. You can see the full set of images chosen by the algorithm in the margins of these pages.The algorithm can be found, free to use and modify, here. We have also made the original outputs of the AI available to the reader in full, both the 100 output stories (in [the stories] section) and the entire range of images selected (in the margins of these pages), to give a sense of the potential and breadth of the material selected by algorithmic means in contrast to our curation of it.

Interested in how our own searches would differ from the work of the algorithm, we also began exploring YouTube ourselves for clips that we felt fit the atmosphere of the stories. Our own fragmentary searches combine with the work of the algorithm to produce the final films.

The algorithm had only three-word phrases to work from, whereas we found ourselves scanning sentences, picking up on implied meanings, and dredging up memories and references to find images that might feel connected to the text. To suit our human judgements on how the text should be interpreted, we found that we had to go beyond the words on the page, to creatively diverge from the original wording and use our knowledge of online posting communities to find search terms that would yield YouTube results that felt, to a human viewer, relevant or appropriate to the subject matter. For example, just searching for ‘dusty patterns cover the wooden furniture’ would yield furniture-cleaning videos, but a search for ‘abandoned building exploration’ gave outputs from a whole community posting videos of such explorations online that evoke just the atmosphere of abandonment that the words suggest. Noting the layers of inference and contextual knowledge put to use in these searches made clear just how much work an algorithm would have to do to simulate the searching human mind.

We worked together to splice these images and videos into short films for two of the stories, with the original story narrated in each case by Ellan Parry. At the point of editing, the strengths of the algorithm became apparent; scrolling through the hundreds of images selected by the algorithm, snatches of text would become manifest through strings of thematically linked images. These images bore traces of the process that selected them, behaving almost like words in that they made a strange kind of sense by virtue of their sequential arrangement rather than through the complexities of content and representation that are usually the province of the image.

The videos reflect this fragmented process and the fractured memories from a computer dreaming, a research output that allows us to present both our data and the process it has been through, with imagery and atmosphere as well as text.

Like dreams, the films produced are strange, disjointed objects, yet still somewhat familiar. Placing images from many different sources with many different purposes side by side, strung together on the thread of a human voice doing what it can to make sense, through prosody and expression, they disrupt narrative norms and create space for speculation. While sitting as complete objects, they do not appear whole. While running from start to end, they do not pretend to be linear. While telling a tale, they do not try to make sense.

This otherness is in large part due to the details of the algorithm we developed. In terms of creating a convincing textual output, as research in natural language processing aims to do, it is a naive construction, producing texts that are not convincingly human-made, that lack a narrative arc, that, unlike most stories with a human author, don't give the impression of attempting to show us something about the world. The algorithm could, however, be improved to produce more cohesive and coherent stories by increasing the data set. A greater number of stories in each of the categories would give the machine more variety of text chunks to choose from. In this, we are embracing what digital theorist Peter Lunenfeld, calls “an aesthetic of unfinish”;64 we leave the question unresolved but in a state of discomfort between very different collaborators. “To celebrate the unfinished in this era of digital ubiquity is to laud process rather than goal — to open up a third thing that is not a resolution, but rather a state of suspension.”65 We leave these texts in a state of suspension, in a state that keeps them from concealing their machine origin among human actors. Further, this “aesthetic of unfinish” is one that has shaped fanfiction from the beginning; “while fans might urge each other on to bring a story to its climax, it is undoubtedly the case that continuity is preferred over closure”.66 The idea of continuity over closure reverberates through popular understandings of plot, through John Fiske’s “infinitely extended middle”67 and J. Yellowlees Douglas’ “inexhaustible story”.68 Both these concepts pose a challenge to models of narrative that insist on defining the story text as a stable and finite thing, instead focusing on their ever-emergent becoming, a mode of being that goes some way towards representing material feminist thinking.

Additionally, in Jamison’s words:

“no fic pretends to be an autonomous work of art. Fic makes no claim to ‘stand on its own.’ It doesn’t need anyone to points out its props and sources because it doesn’t hide them; it celebrates them […] Fic can be uncomfortable for writers who believe they create autonomously in a void. Fic lets its seams show in ways other works that also build from sources and predecessors may be at pains to hide”.69

The ‘discomfort’ that Jamison refers to is something we have chosen to work with, and even exacerbate, in this project, understanding it as another way of revealing sticky frictions — and in revealing them, offering the possibility for response.

Much is lost from the texts in their processing through the progress of this project: their origin in published fiction, the community around them, and the narrative that makes sense of them. The computer engages with the texts as artefacts alone, encountering them as would an alien discovering remnants of our civilisation. The texts are peeled away from their origins, and the stories written by the computer peel away from the vast database we amassed. In this way, what we understand to be human meaning is lost. This is a process of disrupting master narratives and seeing what can be found in the places between.

It is tempting to say that in this losing, we discover the experience of a computer voice, one that we could listen to sustainedly and without attempting to correct its divergences from a human one. This may be a romanticisation: claiming to have given the computer a true kind of agency may seem a stretch when it remains so truly under our command. A truer conclusion might be that, through the stages of this project, we are presenting an ongoing negotiation between humans and machines and, hopefully, a means to see those negotiations as such. We have placed emphasis on process throughout and reported its difficulties and complexities. We have tried to represent the different voices in a collaboration: our own, those in the fanfiction community, and those of the algorithms, recognising their contributions and divergences along the way. Also, importantly, we have made available the algorithms we used for future research, so that our project maintains its own “aesthetic of unfinish”.

Human natural intelligence has, until recent times, lain beyond machine capabilities. But despite the advances of AI, especially in specific domains such as automatic drivers, human creativity cannot be easily mimicked. AI may allow a machine to make informed decisions from what it has learnt, to carry out complex surgery, or do complicated calculations that a human cannot do, yet it lacks a human’s ability to draw on broad contextual knowledge, meaning a computer approaches even simple tasks, like the image-searching, in a very naive way.

Nevertheless, the generated stories provoke a way of thinking that could not be provoked by a human-written text, that offers more in the way of freedom, speculation and uncertainty. That naivety represents an effective (and affective) way of approaching such intractable subject matter as the intimidating and complex future we are facing. Rather than trying to make sense of it all and becoming tangled in the complexity, the AI only repeats what it hears, leaving us with a text that embraces multiplicity and interpretability. We see a melee of voices and images in the stories and films, like a TV flicking wildly between hundreds of channels.

We chose to make two films — one representing destruction, chaos, and environmental disaster, and the other development, technological progress, military progress, and human-machine combinations. These two forces interact and intervene in the filmmaking processes.The output, or stories, could be seen as seeds in themselves, analogous to the seed-words used by the AI algorithm, but using humans’ creativity. Following this direction, we could refer to these as semi-automatic generated stories. On the one hand, we have the machine, the work of the AI algorithm, using humans’ creative outputs as its source material, operating according to strict and orderly directions but creating an output that is disordered, hysterical, confused; on the other, the organic input, its human editors, operating according to elusive, baffling criteria that often appear to defy rationality, but who organise and shape the material in a way that our ever-mysterious minds can make sense of.

And in this way, FanFutures — as a whole and in its parts — can be said to represent the material feminist thinking that we have found so fruitful throughout this project. Describing reality as a fundamental entanglement of entities (anything from a gust of wind to a fleeting thought, a rock, a computer, a wave) engaging in co-creative relations through an ever-emergent process of mutual becoming, material feminisms offer ways to think with the world in all its multiplicity, asking questions about what it means to be human (and non-human) and understanding how we might take response-ability for our actions. Making available such stories of uncomfortable co-creation, of sticky frictions, and experiences that are aesthetic as well as narrative, might help us to pay more attention to the multiplicities with which we are engaged in our interconnected world.