Seven Perspectives on 'the Augmented Imagination Project'

5. Computer-Generated Poetry vs Computer-Assisted Poetry

The distinction between computer-generated and computer-assisted poetry is usually not clearly drawn. It is, after all, reasonable to think that all poetry that is generated by computer algorithms - 'computer-generated poetry' - is at the same time co-created by the human author of those algorithms, and therefore best thought of as 'computer-assisted poetry'. 

To take a well-known example, the 'stochastic texts' created in 1959 by Theo Lutz (often thought of as the first computer poet) involved algorithms for picking words at random from a database derived from Kafka's The Castle, and for arranging them into grammatical sentences using logical constants ('Not', 'and', etc.) resulting in lines like the following:

Not every look is near. No village is late.
A Castle is free and every farmer is distant.

Is such poetry 'computer-generated' or 'computer-assisted'? It is computer-generated in that the output is wholly determined by computer algorithms. But it could equally be said to be a 'computer-assisted' poem, since the algorithms were written and the source text selected by Lutz himself: Lutz, we may say, produced a poem with the assistance of a programmable computer. The same could be said for a wide range of what Christopher Funkhouser and others call 'computer poetry', e.g. Jackson Mac Low's 'diastic' poems or Kenner and O'Rourke's TRAVESTY texts.

However, the computer-generated/computer-assisted distinction gains renewed value when we consider other kinds of text. A subtradition within computer poetry, sometimes traced back to Charles Hartman, uses computer programs to create something like first drafts of a poem (see e.g. Ch 7 of Hartman's book Virtual Muse) which can then be edited into shape. Hartman used his program Prose for this purpose, and other more recent programs such as Chris Westbury's Janusnode and Antonio Roque's various Gnoetry programs have user interfaces that encourage such a practice.

With poems produced by such post-algorithmic editing, it becomes highly doubtful that the term 'computer-generated' is still appropriate: the end product is no longer wholly (or even significantly) determined by computer algorithms. On the other hand, in such cases the term 'computer-assisted' continues to seem entirely apposite. This suggests we draw the distinction between computer-assisted and computer-generated poetry according to whether or not there is further editing of the usual compositional sort after the algorithms have run their course. We may schematize this idea as follows:

In discussing computer-assisted poetry, it will also be useful to distinguish further (even if this is ultimately a matter of degree) between those programs that are intended by their creators to be used primarily for computer-assisted poetry in the above sense, and those that are not. We can trace a subtradition (within the subtradition of computer-assisted poetry!) of creating and using programs of the former sort, programs whose stated aims are to do no more than assist with more traditional modes of composition. Perhaps the most prominent example of this sort of program is technology entrepeneur and futurologist Ray Kurzweil's Cybernetic Poet software, whose intended function is to assist poets by suggesting words to fit into or complete existing lines of poetry. The Augmented Imagination Project can be located within this subtradition.

A distinguishing feature of the Augmented Imagination Project, seen in the context of the history of computer-assisted poetry sketched here, is its simplicity. Its word-generating algorithms use neither 'Markov chains', nor 'n-grams', but simply pick words at random alphabetical constraints. Arguably, this simplicitly is its strength. Markov models in which probabilities of word selection are weighted according to properties of a source text are arguably more suitable for computer-generated poetry, where the aim is a finished product resembling existing poetry, than for computer-assisted poetry, where the aim is ultimately to provide an effective interface for human-computer collaboration in poetic composition. Hence the importance of the visual interface in the Project, which is a metaphor for the way words emerge from the blank, non-verbal spaces of the mind during the imaginative process.

The complex algorithms of Cybernetic poet, or of Janusnode, bring both advantages and disadvantages. On the one hand they allow additional specification of the kind of word to be selected (e.g. the kind used by an existing poet in Cybernetic poet, or drawn from a particular thematic area in Janusnode). On the other hand, in doing so, they implicitly narrow down the type of poetry for which the program can usefully assist the poet. By renouncing such complexities, the Augmented Imagination Project leaves wide open the the range of poetries than can be created with its assistance.

It is true that the Project does introduce a certain level of complexity by means of alphabetical criteria for the selection of words. However, something like this form of complexity is arguably essential for any computerized poetry assistant, since alliteration and assonance are arguably essential to good poetry of all kinds. It is true that the algorithms of the Project are particularly suited to assisting with the production of poetry in the 'Oulipian' vein, in which alphabetic constraints (such as the omission of a particular letter throughout the text) are intentionally used. One should not assume, however, that correct use of the interface implies that every word used in the final poem is a word generated by the program. This is by no means mandatory: it is of course quite possible to include connectives or other words which do not meet the alphabetic constraints being used by the generator, during the 'editorial' process of producing the finished poem.