Many readers of this blog would be familiar with the poem generator: Gnoetry. It is, basically, a computer program with which a human collaborates to create new poems out of a pool of texts. It is a form of constrained writing and an experiment in human-computer collaboration. It has been described and labeled in many ways, such as unconscious thinking (Stern, 2004) and Nintendo for poets (Carnig, 2005), reading the poems has been likened to actually getting to listen to HAL from 2001 sing Daisy (Montfort, 2004), while the co-creator of the program has been called a used-car salesman (cited in this post). The endeavor of computational poetics is also intriguing to some, for poetry generation in particular can simultaneously push[es] on the architectural ramifications of meaning, style and subjective state (Mateas, 2004b). I have been intentionally provocative here, because Gnoetry, computational poetics and Aritificial Intelligence in general is. There is a long history to this area of experimentation, a wonderfully dramatic trail of controversay can be followed through the Racter scandal and debates around Gnoetry. This post is meant to complement texts that cover these issues, so please peruse the further reading at the end of this post for a contextual guide. Instead, this post is an interview with the proud parents of Gnoetry: Eric Elshtain and Jon Trowbridge.

Lets begin with a bit of background to Gnoetry and then well progress through the process of creating and reading a gnoem. Terminology: A gnoem is the resulting output from a collaboration between Gnoetry, the selected texts and the regenerations undertaken by the human user (gnoet). [Note: All quotes in this post are from (Elshtain, 2006a) unless otherwise stated.]

The Birth of Gnoetry 

WRT: Where does the word ???Gnoetry??? come from?

Eric Elshtain: Hard G Gnoetry is a recursive acronym (i.e. the acronym itself is an expansion of one initial, like GNU=Gnu Not Unix) of the sort enjoyed by the open-source community Jon is part of.  Neither of us, alas, remembers what the letters stand for now, though y does stand for yokels.

WRT: Tell me how Gnoetry started. 

EE: It started out of a conversation Jon Trowbridge and I had at a local greasy spoon.  We noted that we have some data points in common:  we come from two fields which people feel comfortable with publicly denouncing, in fact, denouncing right to the face of their practitioners:  math and poetry.  How many times had we both heard Math?  I hate math, or I hated having to read poetry in school &c.  So we decided to mix the two to double the denouncements!  Actually, we talked about the intersections between math and poetry (syllabic and metrical forms), Oulipos algorithmic literature, and early attempts to make poetry with computers (Manchester Mark I; early haiku machines; the Travesty generator written by Hugh Kenner and Joseph P. ORourke.  Jon said I can do that and off we were.  He quickly was able to get the computer to spit out random strings of language, and we noted that there were some real gems.  Each step of the process we noticed something really interesting happening, which encouraged us to continue.

WRT: Jon is now a programmer at Google, but was a student at Chicago with you? Is that how you met? 

Jon and I met through a mutual friend.  He was an undergrad at the UofC but he had proudly dropped out of grad school by the time I joined the UofC gradaute school rank and file. 

WRT: Eric, are you still doing a thesis on poetry and science? What are you investigating?

EE: I am investigating the use of poetic language in scientific writings and the scientific writings of poets in the 18th and 19th centuries, mostly in Britain and Germany. 

I presume Gnoetry is part of the research method?

EE: No, Gnoetrys not part of the thesis.  I think of Gnoetry as adjacent to the thesis, but not part of it. 

About Gnoetry
The first step of the process in creating a gnoem is to select a text from the pool in Gnoetry. 

WRT: I have counted 71 individual texts and four collections of various authors (see list). They are collected mainly from Project Gutenberg and the others are from the web. Who chooses the texts to input and why? 

Eric Elshtain: Its mainly been Jon who has chosen the texts according to his whim and to the strict dictates of copyright law. [Although,] McKenzie Wark sent us a proof of his Hacker Manifesto after he saw our work. Jon sent him a few haiku generated with his own text.

Once the texts are selected, Gnoetry then analyses them, looking for patterns: which words appear next to each other and how often. Jon describes the process:

Once the computer has assembled all of the necessary information about how words appear relative to one another in the source text, it can use that information to randomly generate new sequences of words subject to the rules and patterns that it observed.  The frequency with which a word appears in the original text is an important part of the model.  When generating these new sequences of words, the computer will use common words more often than rare ones, and likewise will use common combinations of words more often than rare combinations.  If the language that is produced in this way is grammatical, it is only a happy accident resulting from the computers meticulous aping of the original author???s patterns of word choices.  For example:  the computer will never put three verbs in a row.  Not because I have explicitly programmed it to avoid doing that, but because three verbs in a row is usually not a word-combination that will appear in the original text.  And if this construction never appears in the original text, there will be no potential for that sort of a construction in the statistical model we build up ??? it will assigned a probability of zero, and thus will never appear when we are generating our random stream of language. (Trowbridge, 2006)

WRT: A gnoet can select any nubmer of texts to input. Eric, you???ve experimented with selecting one text and mixing it with another repeatedly, then one text only. What have you found to be the differences in the experience of constructing and the subsequent output? Do you have a preferred ingredients volume? Do others?

Eric Elshtain: At the Field Tests, John Tipton really got into using one short source text in order to generate a gnoem that insists on repeating certain words and phrases over and over, hence freeing him to compose something that repeats the phrase in this unnerved seven times in a rowsomething no human poet would do on her own, I dont think.  Matthias Regan likes to use a huge number of texts and generate long, long stretches of blank verse a kind of machine-writ Prelude as he calls it that he uses as a rough draft for poems of more of his own devising.  Im fairly conservative in my approach, insomuch as I like to see the authors in the gnoems, to utilize what Ive taken to calling gestalt markers that the reader can hold onto, be them a characters name or a recognizably, say, Conradian or Joycean, Whartonian, &c. phrase or piece of style.  Given a greater number of source texts, the language becomes a bit more natural in some waysfewer repetitions, more room for grammtical and syntactical continiuty (for the most part, though look at the haiku above perfectly grammatical utterances!)  It all really depends on the taste of the gnoet…     

It is then back to the gnoet, who then decides what output the gnoem will be.

WRT: The various outputs seem to be haikus, blank verse, 16-line quatrains, sonnets and fragment. Are there others? Why these structures? Are you considering or planning to output a novel? 

Eric Elshtain: The structures are haiku, renga, tanka, blank verse (with the choice of various line and stanza lengths) and the nonce syllabic forms of the end-users own devise.  These structures are there to make the output recognizably poetry, for one, but also to up the ante on the idea of constraints within constraints within constraints…  And no, we have no plans for story or novel output those structures are well covered in the world of computer-generated language we think. 

WRT: Have you considered a speech output of the gnoems? 

Eric Elshtain: Maybe when we can get Bob Dylan to be the outputter. 

WRT: The next step in the process is the ???regenerate??? function, which allows the user to make changes to the output at the word, phrase, sentence and stanza levels. These changes pull on more words from the text(s) selected. Could you explain the ???regenerate??? function more and provide an example of the development of a verse using this function?

EE: Say Im composing some haiku using the statistical analysis of Poes Fall of the House of Usher and we get:

I know not how, to
the base.  A valet, of the
invalid, were, ha! 

So, I like that first phrase I know not how, but want to change the rest.  I highlight it and hit renegerate and get:

I know not how, to
the words of one of my friend
I was aware, fear.

I decide I like it to one and make another regeneration:

I know not how, to
the words of one of the wind;
and thus I pondered.

Viola!  Not bad.  Thats all language from the source text.  However, the first haiku I generated to show how regeneration works was:

His ordinary
manner had vanished.  There was
a species of awe.

Now, that needed no regeneration at all. 

For those that consent, a chapbook is then created of the outputs and is published by Beard of Bees. These chapbooks usually have a description by Eric and includes details on the texts used in the gnoem.

WRT: Why the date & time identifier on some outputs in the chapbooks? 

EE: Initially, Jon put that there in order to be able to reproduce the chronological sequence of composition when putting together the artifacts from the Field Tests.  I got intrigued with it as a way to communicate how rapidly finished gnoems can be produced, and I have also begun to think about the time stamps as a way to historicize the gnoetic moment of composition, as a new way to consider Wordsworths notion of a spot of time, in which an image in a poem which is tied to the psychosocial development of the author.  A Gnoetic time stamp marks the beginning of the social, semantic and semiotic development of the gnoem itself.  

WRT: In the chapbook with your niece as gnoet , you describe how she shared her gnoems:

Over the 2005 Thanksgiving holiday, I invited my 11-year-old niece, JoAnn Welch, to use Gnoetry0.2. After a single demonstration, she took the reins and composed the ten haiku that make up this short chapbook she decided to title, simply, Gnoetry. She???d sit for a while at my desk, then run into the room with another wonderful haiku that she???d recite for the entire family. (Eric Elshtain & JoAnn Welch, 2005)

It seems from this and other reports you give on Field Tests that a joy of Gnoetry is not in the construction or creation of a text, but in the sharing of the constructed text. Why do you think this is?

EE: I think the joy comes from both events:  the joy of a successful collusion with the software, and then the joy of sharing that result with others.  Its really in the sharing, though, that one can fully appreciate that ones become a true gnoet.  We have a friend, whom well call Bob (because thats his name) who really does not have a poetic bone in his body but who announced to the room during Field Test Two Im a Poet!! and read to us his gnoem.  It was a great moment Gnoetry can help anyone enter into the fraternity of poets, without the pain.

WRT: In ???Her Social Frame,??? you describe the experience of Gnoetry as follows:

Gnoetry belies the myth of individual creation and, hence, eliminates psychodrama from the act of reading. The resulting ???a-ha!??? when reading a gnoem is not based upon an epiphany gestalt, but is borne out of the elegence with which Gnoetry finds poetic solutions to a poetic problem: how would Edith Wharton do things with renga? (Eric Elshtain and Gnoetry, 2005)

Could you expand on this? How does Gnoetry eliminate psychodrama from the act of reading? I find your point about the breakthrough moment being stimulated by Gnoetry???s ???poetic solutions to a poetic problem??? interesting. The joy is not in finding a solution or having your own solutions confirmed, it is about discovering how Gnoetry has solved the poetic problem. I wonder if the same joy could be garnered from a human???s attempt to do this? I don???t think so. Therefore, rather than the joy being that the poetic problem is solved, the joy is from the fact that Gnoetry seems to pull it off. Gnoetry succeeding makes it more alive, an equal player perhaps? So, it is not what Gnoetry does, but that Gnoetry does something. Do you agree? What are your thoughts? 

EE: Youre right.  That Gnoetry does something at all blows people away though I would say that what it does is fairly awesome to folks as well.  That the machine and the human are equal collaborators has been our point all along most other programs either emphasize the machines output or the humans in-put.  Jon has struck an extraordinary balance between these two effects.  The human attempts you speak of are called procedural verse and, as John Tipton points out in the essay he reads at our demos The Oracle at Chicago, the procedures are more often than not more interesting than the poetic output.  So the joy there is in the feeling of the procedure, and in the appreciation that someone when through all of that to make a poem.  I think I should revise what I said in the preface, and say that Gnoetry finds statistical solutions to a poetic problem.  Given that statistical basis, and the fact that the language is not being culled from the inner workings of an individuals mind (which never happens in real life anyway Gnoetry helps belie the myth I speak of, in a long line of arguments against the notion of genius and such) hence relieving the language from the typical understandings of authorial intent and freeing the reader to supply what meaning she wants to supply to the language of a gnoem.  What I mean is this one can never ask of a gnoem What did the author mean?  Thats a meaningless question given the way a gnoem is composed.  Theres no one authorial psychology behind the gnoem Gnoetry doesnt kill the author, it multiplies the author and so no self to be stripteased out of the poem.  The reader in fact becomes the ultimate authority on what the gnoem means semanitcally and especially semiotically not that a gnoem can mean anything, but a gnoem thrives in its ambivalence and ambiguity. 

WRT: You posit, in a discussion about The Dublin of Doctor Moreau, the concept of gestault markers, that you describe as allusions to the source texts or source material (Elshtain, 2006b, 6). Allusion implies intentionality, something a computer program does not have in the human sense. The allusion, then, is imposed on the experience by the reader. Do you think the reader experiences these uncanny references/repeats at the word and phrase levels as allusions or something else? What could that else be if intention is not there?

You also go on to say that the poem composed with Gnoetry relies in many respects, though not entirely, upon ones knowledge of the source texts (ibid.). Obviously readers will have a range of knowledge of the source texts and some with little or none. They can all, potentially, still garner some experience from the reading. Could you venture a few possible approaches to or modes in which a gnoem could be read?

EE: I was just considering the use of the term allusion since it does, indeed, imply intent.  Its difficult, I realise, to discuss this stuff without terms that imply agency.  Ill still hold that these gestalt markers *act* like allusions in the readers mind even though they are just the results of statistical probabilities.  They are part of the solution to the statistical problem posed by the code but the reader is not reading statistics, she is reading language that comes from somewhere specific, and those spots of re-cognition will behave like allusions, intent or no.

That said, if the reader has no knowledge of the source texts no gestalt markers to rely upon her reading will hinge on the syntax, on the shadows of style left over from the source texts authors.  Upon this syntax can hinge not the individuality of the author(s) but the individuality of the reader.  Since a gnoem is an open field of signification, the reader can experiment with being a reader she can read for narrative; new sentences (ala Ron Silliman); ode; panegyric; she can test the linguistic waters and see which reading brings the language together bestor she can read for the disjunctions; for the uncanny; for the hallucinatory possibilities of language derived from a prosthetic poetic mind.

WRT: Gnoetry analyses the relationships between words in the original texts. It could mean that all subsequent output in that session is somewhat derivative of the writing style of the original author. Have you found this to be the case? How much text is needed to give a fairly representative emulation of style? 

EE: That was one of the first things I noted when I was experimenting with Gnoetry 0.0 and 0.1.  I would also go so far as to say that the gnoems are not derivative of but just ARE the writing style of the original author.  0.1 had the ability to make sonnets (end-rhyming became a problem once enjambment was introduced) and I made many sonnets using about a dozen works by Joseph Conrad.  I realized that the sonnets were written, in fact, by Joseph Conrad.  All this tells us something about the power of syntax and grammar which, it turns out are the fossils of voice, diction, style, whatever you call it.  One needs but one text bu an author for the gnoem to represent that authors voice, but ofttimes the form one chooses mitigates this representation.  For example, I tried Whartons Custom of the Country in blank verse and it was not Wharton for some reason, that text in renga form is Wharton 100%.  Go figure. 

WRT: In 100 Tankas From the Verblen (Elshtain, 2004), you observe that the gnomes of Veblen???s text have five features: ???economy of language; strict predetermined form; serial and syllogistic qualities; and stereotypical relationship to a particular moment of insight into and within nature???. It seems that you can extrude the core style of an author through the patterns that Gnoetry recognizes. You don???t seem to have pursued this mode of analysis though? Why?

EE: My line of inquiry has taken me away from the idea of Gnoetry as a kind of interpreter of textual style and voice, as a kind of reader of texts itself.  Im hoping that some other folks will take up this and other ideas. Wed rather that Gnoetry not just be a pet project, but something that others will examine in print. 

WRT: Gnoetry 0.2 was introduced in Field Test Two, April 2004. I note a difference with Gnoetry 0.1, is that each line does not need be accepted or rejected in its entirety anymore. What other differences are there between the programs? 

Eric Elshtain: Enjambment was the Big Thing Jon went after in 0.2 at the suggestion of Charles O. Hartman (author of Virtual Muse). The stochastic algorithm changed as well, in order to give the end-user the ability to make lower-level changes to the text through regeneration.  Also, through the Field Tests, Jon was able to gather information from the poet end-users and he made changes to the user interface according to many suggestions, like the ability to statistically weigh each text before and during composition; he added a nonce syllabic form function; a bi-focal view of the gnoem in progress (a magnetic poetry view on the left, and a as printed view on the right); and the ability to undo regenerations.   

WRT: Eric, Jon Trowbridge and John Tipton have given talks about Gnoetry at many venues. How is it received?  

Eric Elshtain: Matthias Regan a colleague from the UofC has also been on the road with our travelling side show.  And theyve gone very well, especially given the variety of audiences weve presented to thus far, from community college crowds to poetry crowds to more academic ones.  We have yet to be angrily denounced (though I did present a paper at the Uof C where people were very uncomfortable with my arguments about the diminishment of human intent and agency in the composition of gnoems, and I was once accosted in Iowa City no surprise, I guess and called, mysteriously, a used car salesman) but one of these days…  People are both frightened and intrigued by what the machine does (though weve grown to realize that many people leave our demos with some odd ideas about what the machine actually does.  Use the word statistics in a room filled with humanities folks and they glaze over, alas).  When we do demo and go through the process, people exhale with astonishment at what the machine spits out.

Personally, I like the idea that Gnoetry can spit out some amazing things; I also like the idea that Gnoetry, a computer program, is less inhibited by rules than humans:

Gnoetry shows what happens when one abides by local syntactic and metrical rules and forgo conventional ???meaning-full??? structures. While those after noumenal or phenomenal truth might find gnoetry ???soulless,??? for others gnoetry is semantically unrestrained and therefore a ???liberated??? poetry. No human would be able to ???get away with??? constructing an iambic pentameter line with the single phrase ???the door,??? the machine liberates the poet from this self-consciousness, lending insight into the source texts (since gnoetry plainly evokes them, revealing a kind of semantic texture) and the poetic/ gnoetic process. The machine can do what we can not; it can be much more cavalier with signification than the human poet. Perhaps Gnoetry is the 21st Century Aeolian Harp.??? (Eric Elshtain in Gnoetry and others, 2004)

Do you have any questions for Eric or Jon and any resources or views youd like to add? Go ahead.


Elshtain, Eric (2006a) Re: Gnoetry0.2′, personal email to Christy Dena, 14th March. 

Elshtain, Eric (2006b) Gnoetry0.2 and the Continuing Effort to Make Poetry Mean, unpublished essay.

Elshtain, Eric and Gnoetry (2005) Her Social Frame, Beard of Bees, Chicago, 26, June [Online]

Elshtain, Eric and Gnoetry (2004) 100 Tankas from the Verblen, Beard of Bees, Chicago, 11, March [Online] Available at:

Elshtain, Eric and JoAnn Welch (2005) Gnoetry, Beard of Bees, 28, Dec [Online] Available at:

Gnoetry and others (2004) Field Test Two, Beard of Bees, Chicago, 13, April [Online] Available at:

Trowbridge, Jon (2006) How It Works, unpublished essay.

Further Reading

Carnig, Jennifer (2005) Gnoetry Creates Powerful Poetics Language of Human and Machine, The University of Chicago Chronicle, 28 April [Online] Available at:

Mateas, Michael and various (2004a) The Dublin of Dr Moreau, Grand Text Auto, 18 Feb [Online] Available at:

Mateas, Michael and various (2004b) The Poet Laurete and the Machine, Grand Text Auto, 23 April [Online] Available at:

Montfort, Nick (2004) Coding and Execution of the Author, [Online] Available at: Originally published in: The Cybertext Yearbook 2002-2003, eds. Markku Eskelinen & Raine Kosimaa, pp. 201-217, Research Centre for Contemporary Culture, University of Jyväskylä, Finland, 2003. [This text may not correspond exactly with the printed text; the latter should be considered definitive.]

Montfort, Nick and various (2004) Coding and Execution of the Author, Grand Text Auto, 11 March [Online] Available at: 

Stern, Andrew and various (2004) Unconscious Thinking, Grand Text Auto, 4 May [Online] Available at:


7 Responses to “Gnoetry: interview with Eric Elshtain”

  1. 1 Jeremy Douglass

    I enjoyed this interview, and was particularly fascinated by Eric’s mentions of both free and open source software movements and public domain / copyleft commons movements. The connections seem to run deep - not just in the name Gnoetry but in the site footer that declares “All Beard of Bees publications are produced on GNU/Linux systems using only Free Software.”

    Am I correct in assuming that more recent versions of Gnoetry are still available under the GPL? If so, where is it available? I’d also be interested in hearing if the code-authors (Jon? Eric and Jon? more?) have any thoughts on that other dimension of collaborative authorship: donations to or modifications / forks of the Gnoetry source code.

  2. 2 Jon Trowbridge


    I never found the time to package up the Gnoetry 0.2 source code for release, but the subversion repository that it lives in is publicly available from svn://
    Getting it running will probably take a little work.

    Gnoetry 0.3, which is still under development, is not available. In theory I would like to eventually release it under the GPL, but that probably won’t be possible since parts are going to be built on top of proprietary technologies from my current employer. Maybe I’ll be able to release a subset of the code, but we’ll have to see how things work out.

    And if there are any poets who know how to program out there, I would welcome external contributions to the Gnoetry source code. Anyone interested in tinkering with Gnoetry should contact me.

  1. 1 WRT: Writer Response Theory » Blog Archive » WRT interview with Chris Crawford
  2. 2 WRT: Writer Response Theory » Blog Archive » WRT interview with Chris Crawford
  3. 3 We Revise Together: Blogging on Writer Response Theory at WRT: Writer Response Theory
  4. 4 Red Rover Experiment #9, Gnoetry at Impossible to work
  5. 5 Gnoetry interview with Eric Elshtain at WRT Writer Response Theory | Wood TV Stand

Leave a Reply

thesis writing service