A couple years later, I was preparing a syllabus on computer-generated poetry and looked up the URL to find that @a_long_drive had been removed from Twitter, as well as the profile of the author. Devastated, I tried to find other traces of it online, but the only result I could find was a single screenshot of the homepage from a bookmark I had made on Are.na. The internet is a vast place, and many things can often be hidden for most if not all users for years while never truly disappearing. @a_long_drive seems lost though, with the remaining evidence of its existence sitting in a file on my Google Drive.

It is to my mind one of the best digital-born literary works of the 21st century, and I find myself bringing it up whenever I’m trying to explain the mechanics of creating engaging work with computer-generated text. But there is nothing to show others now. In this longing, I began to consider: what would it mean to recreate this bot from the partial archive of it I have collected?

@a_long_drive can be slotted in a long history of combinatory text methods that traces back before the widespread adoption of computers. In the 1920s the Romanian artist Tristan Tzara crafted the manifesto How to write a Dadaist Poem, which proposes a method of cutting up and drawing at random scraps of newspaper from a bag. Tzara's Surrealist contemporaries played a similar game in groups but with each participant contributing a word or phrase to be stitched together—the game became known as 'exquisite corpse' purportedly from a sentence generated during one of the first appearances of the game.

What is arguably the most well-known analogue predecessor to generative text bots comes from Raymond Queneau, co-founder of the Oulipo collective who in the 1960s sought to architect a 'potential literature' through procedural methods of writing or editing text. His book A Hundred Thousand Billion Poems (1961) contains ten sonnets with physical flaps which allowed readers to swap each line of the sonnet, for a total of 1014 possible poem combinations. Interestingly, within a year of the book's publication a digital version was created by Dimitri Starynkevitch, a programmer who had met with the Oulipo. The collective's response was tepid, with Queneau querying Starynkevitch's methods in a letter: "we hoped that the choice of the verses was not left to randomness". For Queneau, the potency of the text lay not in the possible combinations but in the reader's agency to select from those combinations.

In contrast, many of the earliest generative texts with computing machines did rely on computational randomness to draw from text templates. Christopher Strachey's Love Letters (1953) is an archetypal example, generating sweet nothings via a simple punch card algorithm. I was amused to learn about Strachey's hand-written word banks, which look just like the way I sketch out my own generative poems on paper. Fluxus artist Alison Knowles collaborated with computer programmer James Tenney to create House of Dust (1968), a Fortran program that would repeatedly describe a house through a randomised grammar including materials, sites, light sources and inhabitants. The generative poem was initially printed in a foot-high stack of perforated printer paper, but Knowles later made physical installations based on the brief of selected stanzas.

Despite the different themes and mediums of these generative texts, all of them use a simple substitution algorithm like childhood ‘mad libs’ games, where words or sentences are swapped into a given template. This is the same underlying structure as @a_long_drive—as a simple example, a 'strange paperback' may alternately be 'old', 'worn', 'secretive'. The number of potential unique outputs of these algorithms increases near-exponentially based on how many words can be drawn from the corpus it uses. In a retrospective of House of Dust, Knowle’s daughter Hannah Higgins writes of ‘fugitive inhabitants’ in its generated stanzas, the sense that from within the randomness of the algorithm emerges combinations of words and phrases that can surprise with every reading.

The rise of Twitter in the late 2000s marked an interesting inflection in public awareness of computer-generated text: it was not only one of the earliest social media platforms to enable bots which could post autonomously through its API interface, but it was also notable as a place where bots were received both as machine agents and yet just like any other account on Twitter. An early example comes from Allison Parrish, whose bot @everyword started in 2007 and spent seven years tweeting every word in the English language in alphabetical order. Followers would interact frequently with the bot, with certain notorious words such as 'sex' or 'weed' widely retweeted and riffed on.

The practice of making creative bots on Twitter reached a critical mass roughly between 2015-2018 with two innovations. The first was Tracery, a software library developed by artist Kate Compton to simplify the process of making work with a kind of substitution process known as a context-free grammar. The second was Cheap Bots Done Quick, an online service set up by V Buckenham which allowed anyone to upload and link a Tracery grammar with a Twitter account, through which the account could then post from that grammar on a periodic basis. @a_long_drive, along with many others, relied on both of these technologies to operate on Twitter.

Though it's difficult to get the exact number of bots on Twitter, Cheap Bots Done Quick hosted about 54,000 bots chattering away, each varied in format, genre, and audience. Harry Josephine Giles provides a useful taxonomy of strategies for bot poetics, showing how some bots recontextualise other tweets on the platform (like @pentametron, which finds rhyming pairs of tweets in perfect pentameter) while others juxtapose texts from other sources (as in @TwoHeadlines which mashes news headlines together). Unfortunately, some time after Twitter (now X) was bought out by billionaire loser Elon Musk, free use of the Twitter API was removed, forcing Cheap Bots Done Quick to be wound down. While the profiles for many of these bots are still on X, they do not post—they are now statues with closed mouths.

Twitter bots emphasise the most interesting aspects of generative text as a material for literary production. With their endless feed, they disrupt the temporal basis of traditional narrative—even poems, often viewed as the most abstract form, contain a start and an end to the text. Instead, the mode of reading a bot's feed feels more spatial than temporal, a sense of exploring in multiple directions for as long or as little as you please.

The spatial metaphor is picked up by many generative artists. Take for example Mike Cook, game designer and author of Twitterbots, who visualises the structure of creative generators through nested circles: the larger 'possibility space' of all possible output, and within it the 'generated space' (or spaces) that encapsulates whatever text the generator is programmed to create. Not all possible text is evocative or interesting, of course, and so the generative artist's endeavour frequently revolves around crafting these generative spaces to overlap with desired areas of possibility space. I picture this like walking through a field, deciding what to plant and where, always forecasting how the text will look together when it blooms.

In this way Twitter bots are a kind of literalisation of poet Lyn Hejinian's idea of the 'open text', whose components are 'maximally excited' through repetition, recomposition, and recontextualisation in contrast to the fixed interpretation of closed texts. As opposed to a finite sequence of static lines, generative text bots grow infinitely, a collection of randomised fragments that offer unique readings with each iteration. At the same time, the best of these bots must maintain an aesthetic coherence that unites each discrete encounter, what generative artist Everest Pipkin describes as 'the aura of the space'. This is the special way that generative texts make sense of infinity, the way that Hejinian's open text is capable of 'opening uncertainty to curiosity, incompleteness to speculation, and turning vastness into plenitude'.

@a_long_drive has always been the most moving realisation of this form for me. In fragments of 1-3 short sentences each, we encounter distinct scenarios that share an underlying exhaustion and dread: strange voices are heard over the radio, difficult questions are avoided, passing towns evoke bittersweet memories. An emotional narrative emerges despite little explicit plot—two people on the run, towards or away from something they refuse to acknowledge despite it consuming their internal worlds. This comes as much through what is implied but left unsaid, a side-effect Hejinian diagnoses from 'building a work out of discrete fields' which creates gaps between outputs that the reader fills in themselves—thus 'what stays in the gaps remains crucial and informative'.

I'm reminded of the great title of Rosmarie Waldrop's collected poems, Gap Gardening. The curatorial aspects of building a corpus for a generative text must also tend to what text should be pruned or left out, which provides a consistency of tone and also leaves space for the reader to make their own paths through the field that is generated. Against common contemporary understandings of computer-generated text, the best generative bots evidence a particular craftsmanship—a word Raymond Queneau also used to describe the carefully constructed procedures of Oulipian poetry.

If this approach to generative text is like crafting a personal garden, generative text harvested from large language models (LLMs) like the GPT line or Google's Gemini is akin to industrial agriculture: optimised for efficiency, mass consumption, and with little regard to the soil from which it was sourced. GPT-3, a now-dated language model from 2020, was trained on roughly 300 billion words scraped from the internet and archives of digitised books. Think about the many voices contained here, personal knowledge and experiences minced and condensed into a linguistic paste suitable for any functional language task. While I have myself benefited from this technology—I frequently seek coding advice and troubleshooting from LLM chatbots—I'm inclined to agree with Allison Parrish's description of these language models as 'a graveyard we are plundering and making puppets out of people's bodies', and am uneasy about any creative text produced through LLMs. These models are reminiscent of Hejinian's 'closed text': 'The (unimaginable) complete text, the text that contains everything, would in fact be a closed text. It would be insufferable.'

In light of these immense data warehouses, the personally constructed Tracery grammars of Twitter bots present an intimate engagement with generative text. Their algorithmic structure bears the mark of the creator—like Tzara's claim for generative Dadaist procedures, 'the poem will resemble you'.

But @a_long_drive didn't just function as a generative text, it was a Twitter bot. Design researchers Cristina Cochior and Manetta Berends make the case for bots as examples of 'digital infrapuncture', a portmanteau of infrastructure and acupuncture coined by Deb Verhoeven. Distinct from other text generators, the impact of Twitter bots came through their intervention in dominant discursive activity on Twitter, acting as what Cochior and Berends describe as 'infrastructural stress relievers, by actively engaging with the norms and values inscribed into computational tools and infrastructures'. Consider for example Mark Sample's @NSA_PRISMbot, which generates tweets to emulate and satirise the data collection activities of the US National Security Agency, or how @TwoHeadlines works to deflate a culture of sensationalised news headlines. Infrapuncture doesn't solely have to offer relief through satirisation either: @selfcare_bot's hourly reminders to drink water or take a break prompts fellow users to resist the churn of the timeline. These are all, as bot-maker Darius Kazemi describes, 'tiny subversions': small injections of weirdness between breaking news, celebrity gossip and heated debates, timeline gremlins that counter the algorithmic smoothness associated with social media platforms under surveillance capitalism.

The conditions that made the platform a fertile ground for creative bots is somewhat lost in the X of today. After Musk took over there was a mass exodus of users, especially journalists. The dismantling of existing moderation safeguards lead to a proliferation of spam bots and an uptick in the frequency of hate speech and misinformation. X is less the public square that Twitter was and more like a strip mall for con artists and neo-nazis. Its weight has deflated, and as a result there is no more 'infra' to 'puncture'. Some bot-makers have migrated their creations to platforms with similar interfaces like Mastodon, or else simply host the bots on their own web pages. Others have instead opted to lay their bots to rest, tacitly understanding that what they had made was not the same outside of unique dynamics of the garden it was grown in.

In my desire to share @a_long_drive with others again, Theseus's paradox emerges: if I could piece the bot back together, would it be the same? I might be able to reverse-engineer the grammar from my sample data, but I'm unlikely to have captured all the words in its corpus. If I try to incorporate words of my own choosing, could I recapture its aura or will its output noticeably feel different? If I hosted it somewhere other than X—Mastodon, or as its own webpage—would it still be as impactful as if those lonely dispatches were coming from within a once-bustling Twitter feed? Where is the lifeblood of this bot: in its outputs, in its grammar, or in its presence on a specific platform?

There is the other issue of provenance. With the accounts for both the bot and its original author gone, I can't properly attribute the bot. The author may have wanted the bot taken down for a specific reason. What right do I have to circumvent that? I risk repeating the sins of LLM-based generative text, of ignoring the warnings Allison Parrish outlined and commandeering the voices of ghosts.

Maybe it is meant to disappear. Despite the old threat 'the internet is forever', the vast majority of stuff published online dies to link rot. Internet art is an ephemeral art. Software and platforms come and go. There are websites I can recall from my childhood that I have not been able to find again after years of searching. Like small town monuments or highway fruit stalls, they disappear into the rear view and are never seen again.

Works cited

Baillehache, J. (2021). The Digital Reception of A Hundred Thousand Billion Poems. Sens public, 1-13.

Berends, M. & Cochior, C. (2020). Bots as Digital Infrapunctures. Varia.

Cook, M. & Veale, T. (2018). Twitterbots. MIT Press.

Emoji Aquarium [@EmojiAquarium]. (n.d.). Posts [X profile].

everyword [@everyword]. (n.d.). Posts [X profile].

HaikuNewsBot [@HaikuNewsBot]. (n.d.). Posts [X profile].

Hejinian, L. (2009). The Rejection of Closure. Poetry Foundation. (Original work published 1983)

Higgins, H. (2016). An Introduction to Alison Knowles’s House of Dust. The House of Dust, exhibition publication, Art By Translation, James Gallery, CUNY, New York.

Knowles, A. & Tenney, J. (1966). House of Dust [Digital art]. ZKM Center for Art and Media, Karlsruhe, Germany.

Magic Realism Bot [@MagicRealismBot]. (n.d.). Posts [X profile].

Pentametron [@pentametron]. (n.d.). Posts [X profile].

Pipkin, E. (2016). A Long History of Generated Poetics: cutups from Dickinson to Melitzah. Medium.

Queneau, R. (1961). Cent mille milliards de poèmes.

Two Headlines [@TwoHeadlines]. (n.d.). Posts [X profile].

Tzara, T. (1921). How To Write a Dadaist Poem.

UNC-CH Digital Innovation Lab. (2021). Intimacy in the Digital Archive—Allison Parrish and Everest Pipkin in Conversation at the DIL [Video]. YouYube.

Waldrop, R. (2016). Gap Gardening: Selected Poems. New Directions.

Recovering A Long Drive