Edinburgh Festival, Generated · 7–31 Aug 2020


About: The Team | How the AI Works | Rules of Curation | Brand Identity | Illustrations | Acknowledgements and Disclaimers | Press and Coverage

The illustrations that accompany the AI-generated show listings throughout the ImprovBot project are, with a few rare exceptions, not custom-made for a given post: they are randomly assigned. To this extent, they are the functional equivalent of stock images. The full set of several hundred illustrations is also available for reuse under a non-restrictive licence.

Originally posted to this site and the project’s Twitter feed, the illustrations can be viewed in a re-sort by category: Identity: ImprovBot | Identity: Improverts | The Bot, incl. Multiples Edinburgh and Multiples Shakespeare | Moresque | Capture | Dataviz | Base64 | Network | Hardmod | Mojibake | Noise | Extras

Method and Software

Despite the AI-generated texts they complement, the ImprovBot illustrations have only a modest claim to being ‘generative‘. Created on a Linux system with unremarkable processing and memory specifications, the illustrations are built from elements that were manually shuttled back and forth between the vector editing application Inkscape and the bitmap editing application GIMP. Liberal use was made of the deformation filters available in the G’MIC-qt image processing package. The occasional special effect is owed to Photomosh.

The creative process of these illustrations is underpinned by very little theory. The illustrations are governed by a collagist’s need to juxtapose divergent materials and a cartoonist’s desire to make incongruent juxtapositions somehow startling, or even witty when it goes well on a good day.

The illustrations loosely adhere to the rule of tincture, an ancient colour protocol designed to heighten the contrast between the figure and ground elements in flags and coats of arms. Then there’s the countervailing rule of diapering, of texture that diminishes or even disturbs such contrast. Throughout the illustrations, these rules are often honoured in their breach, as it isn’t always clear which element is the figure and which is the texture.

Consider the colour combination of blue and white, which for some reason – maybe in homage to IBM’s corporate colours – dominates the conventional stock image representation of AI. On the whole, the ImprovBot illustrations perpetuate this dominance, but they also counter it with a riot of other colours, often drawn from the heraldic palette, often also drawn from accidental computer output.

Iconographic Elements

The Improvbot illustrations are composed from a limited set of elements that are either sourced from the iconographic repertoire of conventional AI stock imagery – or they are not, in which case they extend that repertoire and might be said to mock or deride it. The cartoon character of The Bot is a figure often displayed against a background textured with elements such alphanumeric code strings or the nodes and edges of a network graph, but there are a few other elements as well. Let’s review the main ones!

The Bot

The Bot Awakens

The humanoid robot is not just a perennial standby of the science fiction genre in the arts, it is also the most inescapable cliché of the stock illustrations that often accompany news articles on artificial intelligence. The ImprovBot project embraces this cliché as a central element of its brand identity, but also extends it to its illustrations. The illustrations mostly retain the hard edge, the reduced detail, and the flat colours of the graphic design idiom that typically expresses a brand identity; in so doing, they also replace the high-tech, high-gloss glamour of the stereotypical AI robot illustration, substituting it with with the matter-of-fact blandness of a social media avatar placeholder.

The Bot’s blandness is played upon in a number of adaptations, iterations and permutations which include the cherub, the hard-hat builder, the Bard, the Jolly Roger, HAL 9000, the songbird, the sunface, the moonface, the virus, and the only two illustrations custom-made for a particular show listing, the H.R. Giger and the Botspotting piece. The latter is a parody on the Trainspotting movie poster featuring a number of eminent Elizabethans (left to right: Edward de Vere, Elizabeth Tudor, William Shakespeare, Francis Bacon, and Robert Fludd).

Illustration set: The Bot, includes Multiples Shakespeare and Multiples Edinburgh

Base64 and Network Graph

Base 64 and Network Graph

In addition to the robot, AI’s stock image iconography often uses other elements such as the noughts and ones of binary code, or the nodes and edges of a network graph. The ImprovBot illustrations duly embrace these two clichés as well, although the chosen network graph is cheerfully wonky and the binary code is transformed to an arbitrary block of Base64 code, set in the OCR-A typeface. This combination, assuming it hasn’t been done elsewhere, is herewith offered as a novel contribution to the computerly visual vernacular.

Illustration sets: Base64, Network

Glitch Elements

Going beyond conventional AI iconography, the ImprovBot illustrations use glitch elements. Glitch art is one of the more self-consciously ‘computational’ modes of visual work, as the glitch – a malformed computing output that can either be encountered unintentionally in the course of day-to-day work, or be deliberately induced – foregrounds the computer’s agency or instrumentality in accidental outcomes.


Pixel Capture

The term ‘capture’ is a clunker, but it stuck around from the artist’s habit of taking screenshots, formally known as ‘screen captures’, especially of screen renderings that are mildly off kilter or have gone haywire altogether. The earliest such captures that were worked into the ImprovBot illustrations are the accidental result of a hardware failure that happened more than a decade ago and caused a browser to render web pages as an intriguingly patterned but otherwise useless array of pixels. Some of the gaudily multi-coloured strands of enlarged single pixels are of more recent vintage and were deliberately induced by maltreating an innocent bitmap image by means of one desperate method or another.

Illustration set: Capture



Another terminological dud, ‘hardmod’ is short for ‘hard modification’, a name for the angular, out-of-whack result of an attempt to vectorise some enlarged pixel arrays that had been sourced from a glitch capture.

The capture and hardmod elements were both used in the ‘proof of concept‘ demo that got the illustrations accepted as a part of the ImprovBot project.

Illustration set: Hardmod



Mojibake is the garbled screen rendering that results from text being decoded with an unsuitable character encoding. The aliased, pixellated mojibake used throughout the ImprovBot illustrations was drawn from the artist’s digital archives, and was gathered in Japan, a decade and a half ago, most likely of a Japanese web page decoded with a Western character encoding.

Being quite chunky, especially compared with the Base64 code, the mojibake element often assumes an intermediary position between figure and texture throughout these illustrations.

Illustration set: Mojibake



Among the glitch elements, finally, ‘dataviz’ refers to images derived from the result of one particular computational mishap that occurred to a complex data visualisation presented in an ACM research paper (Ogawa & Kwan-Liu, 2010). The research project’s flagship SVG version of this visualisation wouldn’t open in one application and opened with serious deformities in another, thereby providing an opportunity for experimentation that perversely relishes both the modernist-constructivist aesthetic of data visualisation and the wilful negation of this aesthetic in its post-punk, post-modernist deconstruction. We reckon that such deconstruction is fair game under the visualisation’s CC-BY-NC licence, and that the discussion in this essay satisfies the licence’s attribution requirement.

Illustration set: Dataviz

Moresque Ornamentation


Moresque is a tessellating floral pattern of Ottoman provenance that was widely used for ornamenting material objects throughout Europe from the late 15th to the early 18th century. Primarily chosen as a heritage counterpoint to the high-gloss futurism of the typical AI stock illustration, its long and tangled stems may also invoke neural pathways, thereby alluding to the neural networks at play in the text part of the project. Moresque patterns also respond very well, and often in delightfully surprising ways, to algorithmic deformation through image processing.

As an ornamental pattern the moresque might be expected to add texture to a background, but it’s quite apt to stand out as a figure against a ground in its own right, or hover somewhere in between. It shares this ambivalence with the mojibake element, which might be why the two often go together quite well, a combination that is particularly striking due to the thematic discrepancy between the historic ornament and the computational artifact.

Illustration set: Moresque


Prof Branestawm and Lehrer Lämpel

Extras include characters such as Prof Branestawm, borrowed from the book cover of a recent work by Melissa Terras, and the Teal Deer, which served as the mascot of an earlier project initiated by Melissa. Closely related to the Teal Deer, the Unicorn is an extra created especially for this project in honour of Scotland’s national beast.

Illustration sets: Extras, Noise

Reuse and Legal

All text and images on ImprovBot.ai and its associated Twitter feed are distributed under the CC-BY-NC licence: feel free to take, remix, and play!

A subset of the ImprovBot illustrations is also available under a less restrictive licence that allows for commercial use and does not require attribution.

A zip archive of the full illustration set is available for downloading (as is a text file of the generated show listings).