Jon D. Lee | February 2022

Jon D. Lee

In 1974, German poet and translator Hans Magnus Enzensberger wrote “Invitation to a Poemat,” a farcical manifesto in which he imagined a machine capable of generating poems. This machine, he theorized, would produce “a not quite endless but indeed quite enormous, perhaps too enormous quantity of poems [exceeding] everything that humanity has brought forth up till now, even if only from a quantitative perspective,” and, according to his calculations, could readily be made with such sophistication that it would not produce a repeated text for several million years.¹

Such a machine might have seemed science fiction at the time, but the post-Sputnik era has demonstrated that computers are changing the world more rapidly than humans are capable of comprehending. And in that light it might not be a surprise that a more-or-less functional version of Enzensberger’s poemat was constructed for a German turn-of-the-millennium poetry festival, and was reused at the 2006 FIFA World Cup, after which it was moved to the Literaturmuseum der Moderne in in Marbach am Neckar, Germany, where it is still displayed.²

Not a decade later, the first of the now-ubiquitous “academic sentence generators” appeared on the internet. Powered by simple algorithms that draw vocabularies from preselected jargon, the websites can quickly create sentences such as “The fragmentation of injustice is sociological in its reinforcement” and “We will extend inquiry-centered outcomes through cognitive disequilibrium,” which were generated for this paragraph by phrasegenerator.com/academic and www.sciencegeek.net/lingo.html, respectively. Using a different preselected vocabulary, the former of those sites is also capable of generating “Shakespearean monologue” such as “O mildewed merit! Thou art the earth’s glorious wretchedness.”

These efforts would not fool a serious scholar, but they evidence the power of even the simple algorithms and vocabularies that drive these generators. That the results might make “sense” to the reader is meaningless to the generator; any connections made are arbitrary, and result from something akin to illusory pattern perception, a concept Jennifer A. Whitson and Adam D. Galinsky define as “the identification of a coherent and meaningful interrelationship among a set of random or unrelated stimuli.”³ Given this, it is not difficult to imagine that a different vocabulary—replace “paradigmatic” with “pear tree”—would generate multiple random poetic sentences, and a more sophisticated algorithm capable of inserting line and stanza breaks could lend the appearance of a poem. And thanks to illusory pattern perception, the reader’s mind would find ways to connect what would be little more than a series of dissociative and fragmented images, creating meaning out of randomness.

The aforementioned generators are limited enough, however, that they are largely unsuitable for poetic constructions. Even a small sampling of sentences generated by phrasegenerator.com and sciencegeek.net reveal that their algorithms operate via simple fill-in-the-blank processes, and the outputs become quickly similar: phrasegenerator.com, for example, inserts random jargon into the sentence “The _____ of _____ is _____ in its _____,” and while sciencegeek.net has a slightly more sophisticated set of preselected phrases, every sentence still begins with “We will,” and then moves predictably through a verb, adjective, noun, and prepositional phrase.

But then enters Erica T. Carter, which is to these generators what a Tesla Model S is to a 1970s AMC Gremlin. Created by Jim Carpenter in 2003, Carter uses a more sophisticated lexicalized tree-adjoining grammar (LTAG) wherein, following the metaphor of a tree, at given points on the “trunk” of the sentence the algorithm can insert and follow various “branches,” which themselves can have further “branches,” thus creating sentences of varying length and considerable grammatical complexity.⁴ When Carter was given a more poetic vocabulary, the results were striking: on his blog at theprostheticimagination.blogspot.com, Carpenter details that Carter’s first poem came out in The Shattered Wig Review #22, and afterward, “on average every eighth submission garnered an acceptance.”⁵

Far more controversially, in the Fall of 2008, Carpenter published Issue 1,⁶ a 3,785-page “anthology” that claimed to feature poems from 3,165 different poets, all of which were actually generated by Carter, including the opening poem, “credited” to Nada Gordon:

A wooden danger

A numb dead ocean peers
               from an other
                            bead at a wooden stimulus of velvet
You could lie

Everyone opens safety
                and despair, where
                   hullabaloos and dangers and hullabaloos
                                   unfold excitement
You are always unspeakable
               for everything that is untrammelled
The bouquet of safety converts to cheerfulness
              in the cemetery
The risk is rather venerable; the
              hopeless snow opens your excitement
Might you not
              open as we
                                open?

Reaction to Issue 1 was unsurprisingly mixed, ranging from the real Andrea Lawlor (whose “poem” appears on pages 3,450-1) noting in her Goodreads.com review that Issue 1 was a “funny smart prank,” to Rachel Mallino’s (p. 3,505) opining on the same review page that she “didn’t like it because ‘my’ poem sucked,” even as she gave credit to the editors for being “marketing geniuses. All the poetic-philosophical vomiting going on about is ha-ha funny.”⁷ Ron Silliman (p. 1,849), however, took umbrage with his “contribution,” stating on his website, “the last time I felt ripped off by an on-line stunt, I sued—as a lead plaintiff in a class-action case brought by the National Writers Union. And while I can’t discuss the suit, as a condition of the subsequent settlement, I will note that we could have gotten a pretty good major league middle infielder for the final amount. Play with other people’s reps at your own risk.”⁸

An evaluation of the literary quality of the poems in Issue 1 could only be idiosyncratic. Gregory Betts concludes in part of his own “entry” (p. 3,435) that it presents “the idea of a language that discloses that which it conceals, a language that speaks absence and thereby abandons the possibility of (human) truth and imitation. This is a poignant confession for a machine”⁹ —a critique which, despite (or perhaps because of) its anthropomorphizing an algorithm into being able to “disclose,” “conceal,” “abandon,” and “confess,” nevertheless values the poem and judges the entirety of Issue 1 as an “important intervention.”¹⁰ But beyond this, it is fair to say that if a student brought “A wooden danger” without explanation into an MFA workshop, it would be received as the product of human effort, and the discussions over the poem’s muscularity and closing three lines would be lengthy.

However, Erica T. Carter is, at almost two decades old, ancient in the world of computers. Newer algorithms have not necessarily moved beyond the fractured obliquity of Carter’s work, but they have become more ubiquitous, with hundreds of websites now featuring generators of greater or lesser quality. This abundance has made possible several informal investigations into the limitations and possibilities of computer-generated poetry, and among the more interesting of their questions concerns how successful readers are at discerning whether or not a poem was written by a human.

Benjamin Laird and Oscar Schwartz have run such a study on their website botpoet.com since 2014. Visitors are presented with the title and text of a poem (all other identifying information has been removed), and asked to judge the piece’s author. The ongoing results show that, for the most part, visitors judge the poems’ origins correctly: as of this writing, 78% of respondents have identified William Blake’s “The Fly” as having been written by a human, and 86% have identified this piece as generated by an algorithm:

Poem

Published on desserts and from pink. Symptoms
Start, 2013 as other poetry does anyone
Word in thailand and write reading one mother
Order they deserve. Well, 2013 recently released
My pants and serve throw from is a beautiful
Insane surreal once hours playing once.

Responding to these results on his website The Hexacoto, journalist, associate editor, and poet Shuan Sim expressed a lack of surprise that many of the computer-generated poems had not fooled visitors: “As can be seen,” he opined, “the above poems are mostly verbiage, and make no sense.”¹¹

More revelatory, however, are those botpoet.com poems that have been incorrectly identified—poems that, in short, make clearer the line between what could be considered human and what could be considered machine. The computer-generated poem most heavily misidentified is currently this one, with 62% of visitors believing it written by a human:

#6

you
  are
               inscribed
                    in the
                             lines on the
            ceiling
              you
are
  inscribed in
                the depths
of
               the
storm

In contrast, 75% of viewers think Deanna Ferguson’s “Cut Opinions” the product of an algorithm:

cut opinions tear tasteful
hungers huge ground swell
partisan have-not thought
green opinions hidden slide
hub from sprung in
weather yah
bold erect tender
perfect term transparent till
I two minute topless formed
A necessarily sorry sloppy strands
hot opinions oh like an apple
a lie, a liar kick back
filial oh well hybrid opinions happen
not stopped

On The Hexacoto, Shuan Sim pondered these results: what were the assumed qualities of human- and machine-generated poetry that this study had uncovered? Discarding the argument that “[prosodic and auditory cues] such as stress, meter, and rhyme give poems a sense of human effort,” since a computer program could take such details into account, he concluded instead that:

human-like poetry trumps computer-like poetry [in its] coherence of ideas. In the poems that read human, most of them have ideas that agree with either a general theme, or the lines preceding and following them. The ideas contained in each line also display a progression, where there is something being explored or developed. The computer-like poems tend to show disjointedness of ideas.¹²

Sim’s conclusions are compelling in their logical simplicity, even if questionable in their development. At the very least, it should be added that both “#6” and Blake’s “The Fly” provide readers with a sense of a definite speaker addressing a specific audience—a sense that is questionable in “Cut Opinions”—and so a fuller conclusion must take such considerations into account to test their universality.

But the botpoet.com results are furthermore problematic in their reliability. Neither Laird nor Schwartz responded to requests to release the raw data of their research, and so key factors such as the number and identity of respondents remain unknown. A few dozen test-takers might produce drastically different results than a few hundred; equally, a group consisting entirely of poets might vote differently than a group of nonpoets, or even established writers in other creative genres. Perception being key, a better study should take such factors into account.

Such a study—though designed and implemented as a contest—is found in Dartmouth College’s annual Literary Creative Turing Tests competitions, named after British computer scientist Alan Turing’s wondering whether a computer could convincingly create output that could pass as having been created by a human. Announced in 2015, Dartmouth’s contest encourages contestants to create and submit programs capable of creating short stories, sonnets, and dance music DJ sets (and, as of 2018, limericks, free-form poems, children’s stories, and machine-human musical improvisations). The created works of art are then mixed with human-generated samples of the same form, and given to a panel of judges who attempt to determine the origin of each sample. Relevantly, as of 2017, no program has yet been capable of writing a sonnet that has fooled the majority of judges.¹³

But even this outcome must be questioned: does Dartmouth’s competition convincingly conclude that computers are not capable of writing believable poems (a conclusion contradicted by Erica T. Carter’s publication record), or does it merely demonstrate that they cannot (yet?) write believable sonnets—a form whose rules and restrictions make it among the more difficult of fixed-form poems? Similarly: would a different panel of judges have concluded differently about the entries? Would a setup wherein unwitting judges were only given computer-generated poems, but asked to determine the human- or computer-generated origin of each entry, have reached the same conclusions? And a host of further matters should be of importance to the poet: what are we to make of these competitions? What are the intentions of machine-generated poetry, and what are its limitations? Is the very nature of poetry threatened?

Some of these questions are more easily answered; others may never be definitively resolved. As detailed by Avery Slater, the roots of machine-generated poetry lie in post-WWI efforts to create machines capable of translating languages, with MIT. granting its first dedicated research chair to machine translation in 1955, and the U.S. government spending “$20 million to fund [machine translation] projects at seventeen different institutions between 1956 and 1965.”¹⁴ The ultimate goal of these projects was to create an “existential monolingualism,” and Slater notes that researchers such as John McCarthy, head of Stanford University’s Artificial Intelligence Laboratory, were so convinced that computers could recreate reality that McCarthy claimed in 1973, “The only reason we have not yet succeeded in simulating every aspect of the real world is that we have been lacking a sufficiently powerful logical calculus.”¹⁵ The following year, Enzensberger wrote his aforementioned “Invitation to a Poemat,” and as is evident, Erica T. Carter and similar programs are increasingly capable of filling the functions of such a poemat, at least for a narrow enough definition of poetry.

The ramifications of these capabilities are both theoretically and concretely profound, beginning with whether or not computer-generated poems actually constitute works of poetry. The fraught nature of the question “what is a poem?” however, quickly leads down a wormhole of refracted semantics. A better question may be “Does a computer-generated poem have meaning?” One answer to this is certainly “no,” since a poem-generating algorithm lacks the capacity to intend poetic meaning, much less “understand” what it has generated; it simply produces texts according to its programming.

Certain linguisticians might agree with this answer. Ignatius G. Mattingly has argued, in “Reading, the Linguistic Process, and Linguistic Awareness,” that while both reading a book and listening to someone speak are processes that result in the transmission, reception, and interpretation of information, there are differences between those activities in terms of how that information is transmitted, received, and interpreted. Speech is, for example, a natural process, and the brain is hardwired to acquire the ability to speak. Reading, in comparison, is “a deliberately acquired, language-based skill, dependent upon the speaker-hearer’s awareness of certain aspects of primary [i.e. speech-based] linguistic activity,” and so “By virtue of this linguistic awareness, written text initiates the synthetic linguistic process common to both reading and speech, enabling the reader to get the writer’s message and so to recognize what has been written.”¹⁶ The act of reading, in other words, is possible because of the reader’s depending on the written text to conform to the same linguistic processes that are present in speech: that information is being transmitted; and that it is possible to receive and interpret that information.

But problems arise when these expectations are violated. Deirdre Wilson argues that the process of interpreting speech (and thus writing, given Mattingly’s connections) involves two disparate processes: decoding and inference. These processes work in synchrony, the listener having to decode the content words and attitudinal expressions in a sentence, and then infer what those words and expressions mean, given the context in which they are presented.¹⁷ But in the case of machine-generated poetry, how is a reader to decode or infer meaning if the “author” of the poem was not poetically aware of what was being coded and inferred?

Another possible answer is “yes”—or, at least, “maybe,” for much of the validation of an object as art relies upon the mind interpreting it as such, and in this sense the computer-generated poem could, for example, be received as a “found poem.” To this end, Katharine M. Wilson argued in 1928 that poetry and music share the same sense of “aesthetic meaning,” which she defined as “the whole content of the mind [being] filled by the thing that gives meaning, the meaning of a sonata or of a poem being everything the mind felt or saw or thought under their guidance, every consciousness modified or induced by them.”¹⁸ Wilson’s definition is not one that would be offered by a dictionary, but instead moves the term towards a sense of all-encompassing awe or involvement in the poem or piece of music: that upon hearing or reading it, one is so overwhelmed in participation that one is almost incapable of being anything but overwhelmed in participation. In this sense, Wilson seems to be echoing Emily Dickinson’s oft-quoted definition of poetry: “If I feel physically as if the top of my head were taken off, I know that is poetry.”

Adding to this sense of “meaning” is Amittai F. Aviram, who in Telling Rhythm: Body and Meaning in Poetry, argues that 20th-century Western poetry has exhibited a movement toward the “meaningful” as free verse overtook fixed form in popularity and changed the poet’s conception of the nature of rhythm and meter. Under this change, Aviram argues that “meaning in poetry [represents], allegorically, aspects of the power of the poem’s own rhythm to bring about a physical response—to engage the reader’s or listener’s body, and thus to disrupt the orderly process of meaning.”¹⁹ True, this definition examines the nature of rhythm in poetry, and is not intended to provide an all-inclusive definition of “meaning,” but it is still notable for its connections to Wilson’s sense of how the poem fills the mind.

Of critical importance is that these discussions all seem to imply that intended meaning is less important than received meaning. None of the authors were of course anticipating their definitions to be applied to machine-generated poetry, but they still imply that any poem, regardless of origin, that creates an emotional or physical response has fulfilled a necessary quality of poetry. Such a definition seems concordant with many of the doctrines of early- to mid-20th-century New Criticism, wherein, as summarized by Art Berman, “[t]he poem is an object, open to analysis that can be confined…to the poem itself,” and the act of “Designating the poem as an autonomous object to be studied entails rejecting critical methods that would understand the poem through history, biography, sociology, or other non-literary disciplines.”²⁰ Such a stripping of externalities could extend even to considering the author of machine-generated poetry a superfluous bit of information, thus rendering the poem open to consideration as literature.

But as New Criticism’s textual isolation ultimately failed to take into consideration the cultural forces that create and engage writers and readers of literature, it was replaced by more reader-centered theories. Even these, however, would seem to allow for machine-generated poetry to be considered meaning-laden literature, since much of their focus is on the interactions between reader and text. Bruce Chadwick notes that in these theories:

No longer was the text to be viewed spatially, as a whole, separated from the experience of reading, but was to be part of the reader’s temporal process of self-discovery in the act of reading. Readers make meaning; the text becomes part of a phenomenological transaction where the text becomes part of the transactional process, the phenomenon of discovery through reading.²¹

Many of the seminal arguments in these reader-centered theories came from Wolfgang Iser and Stanley Fish. Iser saw the text as being indeterminate, with the reader filling in the “blanks” of the text, “the filling of which brings about an interaction of textual patterns,” resulting in the reader being an active maker of meaning.²² Fish, however, even while agreeing that “interpretation is the only game in town,”²³ argued that only an academic “interpretive community” could make meaning: “The discovery of ‘the real point’ is always what is claimed whenever a new interpretation is advanced, but the claim makes sense only in relation to a point (or points) that had previously been considered the real one.”²⁴ But again, and noting that neither Iser nor Fish intended their arguments to be so applied, it seems difficult to discount literature generated by algorithms as having meaning, provided that a reader (or group of readers) find meaning in that literature.

But accepting that machine-generated poetry has meaning—to agree with Andrew Cecil Bradley, who noted, “if we insist on asking for the meaning of…a poem, we can only be answered ‘It means itself,’”²⁵ and so unwittingly allow for machine-generated poetry to be taken seriously—creates profound consequences for literary theory. Among other details, it forces a rethinking of the role of human creativity, and indeed the purpose of literature itself. Dorothy Hall wrote in 1941 that literature “belongs with religion and philosophy. It is an exploration for a set of values to live by…literature is our most easily available guide to the values that sensitive and discriminating men and women have found in living, and hence fertile in suggestions for our own living.”²⁶ But what if humans did not create those guides and values? Would they still be our own? Cofounder of botpoet.com, Oscar Schwartz, noted in a 2015 TEDx talk that his research had led him to conclude that “the category of the human is unstable…the human is not a cold, hard fact. Rather, it is something that is constructed with our opinions, and something that changes over time…the computer more or less works like a mirror that reflects any idea of the human that we show it.”²⁷ So if machine-generated literature does reflect us, is it merely ourselves as seen through the hazy looking glass of computer code? And how much of what it means to be human is lost in translation—the equivalent of an audiophile forever disappointed by digital music downloads?

One answer that partially relieves this pressure comes from Robert E. Probst, who in discussing the role of reader-response theory agrees with theorists such as Robert Scholes that “Meaning lies in that shared ground where the reader and text meet—it isn’t resident within the text, to be extracted like a nut from its shell.”²⁸ But in examining the classroom consequences of such a stance, Probst acknowledges its limitations, noting “This is not, of course, to say that texts mean anything we want them to mean—it’s obviously possible to misread, to misunderstand a word, or miss a point. When we argue that a writer holds a certain belief, that a character has certain values or goals, we obligate ourselves to offering evidence and logic that sustain our position.”²⁹ Still, even while admitting that a text does have an intended meaning that should be taken into account—a stance that seems to remove computer-generated works from consideration as literature—Probst’s arguments only apply in those cases where the true author of the text is known. But in such cases as Erica T. Carter, where the algorithmic nature of the author was not disclosed until after multiple publications, the line becomes less clear. And what if that link were never revealed? What if Carter’s creator had simply published the poems under his own name, and kept their origins a secret?

This latter possibility is perhaps the largest consequence of all. For, were a fully-functional Enzensberger poemat created, the repercussions would go beyond the merely philosophical, such as “Is a poem created by a machine actually a poem?” and “If a machine-generated poem creates an emotional response in a human, was that response intended by the computer or simply an accident of programming?” Instead, the consequences would become worryingly material. Imagine even a single poemat capable of creating a poem in any style or form that could convincingly pass as having been written by a human. Such a machine would not be bound by sleep, food, or rest, so could produce poems ad inifinitum. Given even a generous rate of one poem every ten minutes, that machine would produce in excess of 50,000 poems per year; at a rate of one poem per minute (which is more realistic, given the production rate of current algorithms), that figure jumps to more than half a million poems per year.

Now imagine that poemat also equipped with a program capable of generating cover letters with random author names, and sending out electronic submission packets. Any literary journal or book publisher who did not require a reading fee would be instantly inundated with more material than they could possibly read, and the sheer scale alone would make it impossible to separate the machine-generated packets from the ones submitted by humans. Not even CAPTCHA tests would fully deal with these automatic packets, since a determined human in charge of the poemat could still sit at his desk for an hour or two a day and enter the codes, or simply move to printing out the packets and sending them manually. And again, this scenario only involves a single poemat. But the existence of one would quickly lead to many, possibly forcing the entire industry to adopt submission fees to counter what would quickly become spam. Perhaps worse, however, if the purpose were self-promotion, a resolutely unethical person with deep enough pockets could theoretically establish an award-winning career as a poet without having himself written a single line. What then happens if that person wins, say, a Guggenheim Fellowship? Would the entire endeavor of poetry be devalued?

We are rapidly moving toward an era in which scenarios such as this will become more than just thought experiments. One of our few saving graces is that today’s algorithms have not yet progressed to the point of being able to fool everyone. Yet they are still at least moderately sophisticated, and competitions such as Dartmouth’s will ensure that they only grow more so.³⁰

What then are the limitations of poem-generating algorithms? Could a machine ever write an epic poem such as Paradise Lost, or a narrative poem such as B.H. Fairchild’s “Body and Soul”? A coherent sestina or heroic crown of sonnets? Could it even write a simple extended metaphor such as e e cummings’s “she being Brand,” which describes a speaker’s attempts at learning to drive a car with a manual transmission, but which is obviously about—without ever saying so—the speaker’s first copulatory encounter?

Answers are, unfortunately, unclear. Scholars such as Margaret Boden, a professor of cognitive science at the University of Sussex, have argued that computers have already modeled the three main types of creativity—combinatorial, exploratory, and transformational—providing them with the ability to mimic creative thought. However, the combinatorial creativity necessary to create fresh poetic imagery is unlikely to be found in artificial intelligence:

Most random word combinations… would be senseless. A practiced poet might be able to use them in a way that showed their relevance—to each other and/or to some other ideas that we find interesting. But the computer itself could not. Unless the programmer had provided clear criteria for judging relevance, the random word combinations couldn’t be pruned to keep only the valuable ones. There are no such criteria, at present.³¹

On the other hand, Hiroshi Yamakawa, Director of Dwango AI Laboratory and one of Japan’s leading artificial intelligence researchers, is working on creating “a human-like artificial general intelligence (AGI) by learning from the architecture of the entire brain”—a process which uses “the integration of artificial neural networks and machine-learning modules while using the brain’s hard wiring and a reference,” resulting in a problem-solving ability that “will be more similar to human intelligence.”³² Such a creation would almost certainly push the boundaries of machine-generated poetry. According to Yamakawa, the average estimated delivery date among researchers for this AGI is 2030.

The future being murky, it only makes sense to focus on the present, and to limit as much as possible, for as long as possible, the probabilities of machine-generated poems being mistaken for human. This task falls exclusively on the poet, since technology will continue its inexorable march, and must begin with an examination of which techniques, styles, and characteristics unite, or at least bring closer together, the machine- and human-generated poem. Here at least some introductory answers are possible, for an examination of Erica T. Carter’s poems, as well as those most often identified on botpoet.com as having been created by a computer (regardless of their origin), reveals a set of underlying assumptions on the part of the reader: that those poems that are less fluid, more grammatically problematic, less immediately intelligible, and more fractured, are more likely to be seen as the product of an algorithm.

Put another way, those poems that are most often seen as computer-generated are those that align most closely with Tony Hoagland’s definitions of a “skittery” poem. In his 2006 essay “Fear of Narrative and the Skittery Poem of Our Moment,” Hoagland notes that there is a specific type of modern poetry that is characterized by “obliquity, fracture, and discontinuity,” a “mistrust of narrative forms,” and “a pervasive sense of the inadequacy or exhaustion of all modes other than the associative.”³³ These poems, Hoagland argues, are quite fashionable at the moment, but fall into the traps of much of fashion: they “lack…perspective” and don’t “recognize the deep structure of whatever manners [they are] adopting.”³⁴

Hoagland furthermore says that “Speed, wit, and absurdity”³⁵ are frequent qualities of such poems, and he provides several examples, including this, from G.C. Waldrep’s “Watercooler Tarmac”:

My harvest has engineered a sanctioned nectary.
The transmission of each apple squeals when I apply the
 compress.
All my obsequities have finished their summer reading,
they are diligent students,
they understand the difference between precision and
 Kansas.
This was before I had pried up the floorboard to see what
 was ticking underneath.
I keep busy, every plane that flies through my sky
requires help, sign language for the commercial vector.
My octave’s intact so this may be working.

Hoagland argues that such poems “showcase personality in the persona of their chatty, free-associating, nutty-smart narrators. It is a self that does not stand still, that implies a kind of spectral, anxious insubstantiality. The voice is plenty sharp in tone, and sometimes observant in its detail, but it is skittery. Elusiveness is the speaker’s central characteristic.”³⁶ Cynically, it could be summarized of such poems that inasmuch as they can be assumed to have meaning, they ready to take one’s doctoral examinations when one can take a sentence such as “the child fell down” and rephrase it as “the infant exhibited a strong transitory propensity towards ground-wards tropisms.”

It is perhaps no coincidence that the skittery poem most closely aligns with the perceived-as-computer-generated poem for, as Avery Slater argues, the existence of poem-generating algorithms is actually “a false-positive, a ‘success’ that is only possible once the poem has been rendered so already fragmented/trans-rational/suggestively disconnected in the modernist era.”³⁷ But it does mean that any poem that relies too heavily on the fractured, discontinuous, and oblique, or otherwise fulfills many of Hoagland’s definitions of “skittery”; that presents a sequence of disconnected words, phrases, lines, or sentences that seem to rely on illusory pattern perception to supply the connections; that conflates surface-level complexity with intellectual depth; that lacks Shuan Sim’s “coherence of ideas”; that does not have an identifiable subject matter; or that, in short, does not appear to have an intended meaning, moves ever closer toward the line where the products of machine and human become indistinguishable. Therein lies the danger.

This does not mean the modern skittery poem needs to be abandoned. Nor does it mean that poems should not be difficult. Indeed, difficulty may be what saves the poem from the machine: Reginald Shepherd has eloquently argued that “the intellectual is an essential element in poetry,” and that “mystery in poetry can be a lure…[o]ne wants to solve the mystery, or at least to better understand its source.”³⁸ The problem for Shepherd is the poem that is obscure: that is difficult “in an attempt to cover up [its] vacuity.”³⁹ Here the algorithm can take over, for a cleverly programmed sequence of random words could create the illusion of obscurity. But, at least for the moment, algorithms cannot simulate the intellectual, as seen in their inability to move much beyond skitteriness, and so that is where the poet must focus her attention. To do otherwise is to be complicit in the continuing devaluation of her own work.

---

Jon D. Lee is the author of four books, including IN/DESIDERATO and An Epidemic of Rumors: How Stories Shape Our Perceptions of Disease. His poems and essays have appeared or are forthcoming in several literary journals. He has an MFA in Poetry from Lesley University, and a PhD in Folklore. Lee teaches at Suffolk University, where he also serves as a Senior Poetry Reader for Salamander.

---

Notes