Assisted Poetic Creation


At a time when Artificial Intelligence is part of many aspects of everyday life, how can it help us in today's textual creative processes? Poetry is often perceived as the source of pure inspiration before a blank page. This, however,  it is forgetting all too quickly the degree to which it comes with traditions well known to poets – dictionaries of synonyms or rhymes, learning how to write. Assisted creation, in fact, is no issue in photography, as seen in the digital support available on smartphones to edit a photo’s brightness, sharpness, or other features. In the same way, Artificial Intelligence or parametric fonts (developed in recent years) and self-generated layout can support the creative process.

But is Artificial Intelligence actually becoming creative? Instead of taking the often illusory, sometimes frightening, option of replacing man with still woefully inadequate text generators, we chose to intervene in the different layers of poetic creative writing. Man and machine collaborate in a constant dynamic: while the machine suggests, Man always chooses. Machine never replaces Man, but rather, helps him make a choice. This workshop invites you to explore the path of poetry with new tools, as if in an assembly line leading to a vast scroll that collects and assembles the exhibition’s poetic energies.

The stages of
assisted creation

In order to help the Artificial Intelligence generate a first draft, three elements are left to the visitor's choice: the poem’s form (quatrains, haiku, free verse), theme (love, death, war ...), and emotions. A text is generated. To eliminate any syntactic errors or spelling weaknesses from the text generators (which work by statistical reconciliation), the visitor can correct the text and improve it after this first step. Using the cursors, the visitor gives the text emotional colour: more joyful or melancholic?

To finish the poem, the calligraphic tradition must of course be considered. This is done through finding suitable font styles. The font and layout used can strengthen the text’s meaning and give a visual sense of ​​the content the words convey; this last stage of the creative process and final rendition focus on these specific elements. The creator can modify the machine-generated font to a certain extent to accentuate its severe appearance or, on the contrary, to facilitate its legibility. From there, the poem continues along its path. It is either printed during the exhibition on a huge sheet, can be individually searched online, or can photographed on a smartphone and made shareable at will.


The system relies on artificial intelligence to learn the probabilities of sequences of words, using an artificial neural network with multiple layers and temporal recursion.  The weights of the connections between artificial neurons are optimized during a training phase, by observing actual sequences of characters from about 6000 poems.  After training, the system is able to estimate the probability of any sequence of characters and words, but also to create new sequences with high probabilities, which are displayed as lines and stanzas of poetry.  The form parameters are constraints on the lengths of the sequences; themes and emotions result from the combination of different AI models, trained separately on poems annotated as representative of specific themes or emotions; and rhymes are adjusted by searching among the most probable words those with matching endings.

Nicolas Baldran, CfFP
David Héritier, CfFP 
Andrei Popescu-Belis, HEIG-VD 
Antonio Rodriguez, UNIL 
Aris Xanthos, UNIL

Forms, themes, and emotions:
Andrei Popescu-Belis, HEIG-VD 
Aris Xanthos, UNIL 

Typography and printing: 
Nicolas Baldran, CfFP 
David Héritier, CfFP 

Programme development: 
Gabriel Luthier, HEIG-VD
Valentin Minder, HEIG-VD
Alejandro Ramirez Atrio, HEIG-VD

David Deppierraz 

Denis Correvon ​​​​