Notes & Major Points

_8-3_Doyle_Generative_AI_and_Illustration.pdf

Home >English homework help >Notes & Major Points

Generative AI and Illustration: Questions From the Field

Doyle, Susan

Journal of Illustration, Volume 11.1, Special Issue: “Illuminating the Non-Representable”, 2024

Doyle, Susan. “Generative AI and Illustration: Questions From the Field.” Journal of Illustration, vol. 11.1, 2024, pp. 165–89.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

jill 11 (1) pp. 165–189 intellect limited 2024

journal of illustration Volume 11 Number 1

www.intellectbooks.com 165

SUSAN DOYLE Rhode island School of Design

Generative AI and illustration:

Questions from the field

Abstract

The multi-year international project ‘Illuminating the Non-Representable’ (IN-R) sought to consider the breadth of possibilities in contemporary illustration practice. A question was how, in an ever-more global context, illustration might sensitively communicate concepts of ‘the other’ – meaning persons who do not share heritage or characteristics of the perceived audience or illustrators themselves. With growing AI image generation in 2022, new questions arose regarding types of AI images users were prompting and whether resulting images perpetuate bias inherited through machine learning that is trained on databases already proven to encode bias. This article shares an analysis of a sampling of AI-generated images in response to those questions and includes expert opinions on the benefits and limits of AI creativity, ethical issues related to plagiarism and the unauthorized scraping of copyrighted works into training databases, as well as more generally on the erosion of professional practice that generative AI portends.

Keywords

artificial intelligence Midjourney AI image scraping plagiarism illustrators copyright bias

Received 30 March 2024; Accepted 1 June 2024; Published Online November 2024

https://doi.org/10.1386/jill_00089_1

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

166 journal of illustration

Under a project entitled ‘Illuminating the Non-Representable’ (IN-R) led by Hilde Kramer (University of Bergen), a group of educators, researchers and art practitioners set out to consider whether cross- disciplinary art forms, not limited to traditional forms of image-making, can convey complex mean- ing that might be regarded as ‘illustration’ (Kramer). One key question that arose was how illustration manifests notions of ‘the other’, meaning persons who are from outside the dominant culture of a particular region. The representation of marginalized groups, especially by someone not of that social group themselves, can be fraught because heritage, traditions or even certain physical attributes may not be sensitively perceived or shared. This article, which was originally given in June 2023 at the Falstad Centre, Norway, at the last of three IN-R symposia, was principally conceived to consider questions relating to current trends (2022 to mid-2023) in AI image generation as they intersect with this question. Analysis of a large sampling of images made by online image generators revealed that representation of marginalized social groups was indeed problematic. This article shares that original analysis and practical concerns related more generally to AI image generation as it has become more widely embraced by illustration practitioners. It also includes new interview data and updates as of spring 2024 to reflect changes in AI image generation technology and discussions by researchers, politicians and professional illustrators.

This article shares analysis of those images from the perspective of an illustration educator and is certainly not a technical treatise on AI technologies. Emphatically, my research is not approached with any expertise in computer programming, but rather through the eyes of one who analyses how pictures create meaning and influence taste. Because AI generation platforms are generally accessible and make picture generation easy and inexpensive to both amateurs and professionals, AI images can and do displace commissioned human art, and they exercise influence over visual culture. This is of key importance to illustration education.

Technology and illustration

Unlike all prior forms of illustration, generative artificial intelligence (Gen AI) does not rely on the handiwork of an artist. Rather, Gen AI programs use algorithms to compose new images by synthe- sizing digital image information in response to textual prompts made by a human user. Freely availa- ble, images generated by AI have been compelling enough to win art and photography competitions (Zhang 2023b) (Figure 1) and even to falsify an image of a disaster that was so believable that its appearance on the social media site Twitter (renamed X) briefly sent the New York Stock Exchange reeling (Bond 2023) (Figure 2). If the trend towards massive adoption of AI image generation contin- ues apace, questions arise as to how such images might impact not only the news but also the prac- tice of illustration – forcing educators, publishers and artists themselves to grapple with the ethics and potential legal ramifications of using AI-generated images in a rapidly evolving visual culture.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 167

Figure 1: Pseudomnesia: The Electrician was submitted by Boris Eldagsen and won the Creative category, Sony World Photography Awards, 2023. Artwork created using DALL-E 2, in Boris Eldagsen, n.d., https://www.eldagsen. com/pseudomnesia/. Accessed 15 June 2023.

Figure 2: Artwork attributed to AI, in UKR Report, Twitter, 22 May 2023, https://x.com/ukr_report?lang=en. Accessed 15 June 2023.

~ UKRREP0RT O • @UKR_Report

3:10 PM • May 22, 2023 • 1,366 Views

https://www.eldagsen.com/pseudomnesia/

https://x.com/ukr_report?lang=en

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

168 journal of illustration

We have become accustomed to search engines retrieving and qualifying seemingly infinite amounts of information and images without much concern about how that information was deter- mined. Yet newly available AI programs that manifest ideas in words and pictures without explicit artistic control by a human – unthinkable even a few years ago – have shaken our collective consciousness.

This disruption can in some ways be compared to the paradigm shift in visual culture when photography became viable for large swaths of the population in the nineteenth century. The one- to-one correspondence of light’s action on photo-sensitive substrates reshaped the very notion of what was ‘actual’ in pictures. The apparent objectivity of photos infiltrated the mindset of everyday folks and called into question the inherently subjective processes of handmade images, no matter how realistically drawn. A century later, Adobe Photoshop opened up means of digital painting, compositing and retouching that gave digital artists the power to create and edit illus- trations and falsify photographic images in previously inconceivable ways. Perfection was possible and the means undetectable – and the status of photos as verifiable ‘truth’ was diminished.

Both photography and digital painting, however, rely on human skills gained through long hours of practice, and a level of consciousness is required to make compelling images. This is not so with AI image generators that output pictures in seconds. Industry watchers estimate that between the first public access of DALL-E 2 in April 2022 through August 2023, 15.470 billion were generated by just four sites – DALL-E 2, Stable Diffusion, Midjourney and Adobe Firefly – roughly the same number of photos that were taken in the 150 years from 1826 to 1975 (Valyaeva 2023). Through AI image generation, the creation of hyper-realized imagery by non-artists is possible, raising altogether new questions about authorship, truth and art education that will certainly upend the status quo of illustration practice.

How does AI image generation work?

AI compositions are not derivative in the way a human might copy or digitally collage. Rather, AI generators algorithmically synthesize source material. Through a process referred to as machine learning, the system correlates attributes and similarities in keyword-tagged pictures to encode and prioritize in a predictive way desirable visual characteristics to make a new image based on the words of a verbal request: a.k.a. the ‘prompt’. The programs make use of references to visual quali- ties, stylization and other contextual information associated with the pictures. AI images are not wholly original, yet they are not easily attributable to any single source whose images were used in the training process.

AI generators have reached their current capabilities in part because of access to massive amounts of information ‘scraped’ from the internet and other databases used for AI training. Debates exist

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 169

over whether the billions of images made available under the aegis of non-profit research should continue to be collected and shared through mega non-profits like Large-scale Artificial Intelligence Open Network (LAION) and Common Crawl, since those databases are accessible to both non- profit and commercial AI generation entities.

Further questions persist around what is being output by computers utilizing those hundreds of billions of images since, after all, computers do not possess any human sense of ‘right and wrong’ in synthesizing available data, and to date there is no consistent industry oversight of AI.

Research + process

Motivated by the consternation surrounding AI image generation, I studied a sampling of AI-generated images in the spring and summer of 2023 in order to evaluate images that were output via publicly accessible image generators. The goal was to develop objective data around an otherwise largely subjective debate.

Like many professional illustrators and academics, I tested some of the freely available AI image generation sites like DALL-E 2 and Open AI that let users view and iterate their own images through successive prompts. Midjourney AI, however, hosts a ‘Community Feed’ that openly shows images prompted by all users within a given time frame and, thus, offered greater access to what others were prompting.

Midjourney is accessed through the Discord user interface, a free online text and video chat site that hosts public and private community groups (Minor 2024). After establishing an account, I was able to view all the images showcased on the Midjourney Community Feed at a given moment in time, in a vertically scrolling grid (Figure 3). In order to understand the user’s intention and, thus, gauge if the illustration expressed it, I also needed to access what had been requested – the text ‘prompt’. By purchasing a Midjourney subscription, I gained access to enlarged images with their full prompts, available in sequence without the need to backtrack to the Community Feed homepage between each picture.

Each work session commenced as follows: I captured the Community Feed in a series of screen- shots. I then accessed a close-up of each picture in a pop-up window and made a screenshot of it. I transcribed prompt information and cross-checked to make sure that each image in the session was accounted for in the data sheet. All images in this article were downloaded from the Midjourney Community Feed by the author unless stated otherwise.

I evaluated the attributes of each AI-generated image using limited criteria such as whether the image seemed to mimic handmade or digital illustration or if it appeared to be a photograph, as well as names of artists or historical art styles that were in the prompt. I noted whenever race, gender and cultures were specified, and when Black, Indigenous and people of colour (BIPOC) were depicted, Figure 3: Screenshot

by Susan Doyle of the Midjourney Community Feed.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

170 journal of illustration

and whether those images manifested racial or gender bias. Fantasy was prevalent in the Midjourney Community Feed, so I tracked the percentage of images that were ‘unreal’ – including mythological subjects and science fiction.

Because I was looking for evidence of the benefits and challenges Gen AI might pose to indi- vidual artists and to illustration practice as a whole, it seemed relevant to determine whether the AI-generated images were more-or-less what was articulated in the text prompt – and, therefore, perhaps viable in a professional workflow, meaning that an art director or an editor could use AI rather than commission a human. Images that did not include or allude to the requested subject matter or that were stylistically unrelated to the prompt were called out as uncorrelated/poorly correlated.

Over 2.5 months, I analysed 1795 images, recording the wording of prompts and attributes of the images. No restrictions were applied in collecting images because the intent was to faithfully reflect user outcomes.

Findings

The Midjourney site generated a wide range of illustrations. 28.9 per cent of prompts called for simu- lated photography and the images were generally very believable as such. Some prompts were quite exacting and included names of camera brands, particular lenses, F-stops and mention of the styles of certain photographers or particular filmic moments, as if capturing a film still. For instance, in addition to the subject matter and dramatic lighting, the prompt for Figure 4 specified simulation of a photo taken by a Nikon 27 camera with an 105-mm focal length.

71.1 per cent of images were ‘illustrations’, meaning they did NOT prompt for or look like photographs. The largest subcategory of ‘illustrations’ I labelled ‘rendered realism’ because they depicted volumetric forms with convincing light and shadow and attention to surface and texture. In Figure 5, for instance, the image on the left was based on the prompt ‘A black cloud rises over a residential street with park cars, in the style of dutch and flemish [sic], vivid energy explosion, monumental vistas, dark, white, and light grey’, while the image on the right was based on a prompt for ‘a cute kitten inside of foliage, white and green, depic- tions of inclement weather’ with mention of artists Artgerm (Stanley Lau), Shilin Hwang and Thomas Kinkade.

Certain prompts straddled semantic lines with prompts that included multiple or conflicting terms: photo, photorealism (photorealism is historically a genre of painting), painting and photo- realistic. In such prompts, I made an assessment as to whether the outcome was indistinguish- able from a photograph or, even if tightly rendered, looked more like a digital illustration than a photograph.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 171

Figure 4: Image created using Midjourney by Landerholm writing the prompt seen below the image. AI credibly simulates a photograph of cinnamon rolls taken by a Nikon 27 camera with an 105-mm focal length.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

172 journal of illustration

Figure 5: Images created using Midjourney by users writing the following prompts: (left): ‘A black cloud rises over a residential street with parked cars, in the style of dutch and flemish (sic), vivid energy explosions, monumental vistas, dark, white and light gray, dutch seascapes, multilayered, wimmelbuilder’; (right): ‘Cute kitten sitting inside of foliage, in the style of art germ, shilin huang, thomas kinkade, water drops, lush scenery, white and green, depictions of inclement weather.’ Examples of art categorized as ‘rendered realism’. Fair use.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 173

A relatively small 23.4 per cent of prompts asked for some level of abstraction or artistic styli- zation. This subcategory included images with visual attributes that simulated analogue drawing (pencil, charcoal, ink, etc.) and gestural or abstract painting, vector or comic art or faux craftworks like quilling (curling and folding paper into pictures) and felting that appeared to be dimensional handmade art. Images of this type were categorized as ‘not rendered/not realism’ and further tagged as ‘prompt specifies art style’. Abstracted or painterly output tended to be less successful than the photo-realistic work. For instance, Van Gogh came up many times in prompts, and AI imitated his work mainly in terms of the thickness of paint daubs, which looked more-or-less like frosting rather than capturing an expressive synergy of mark, colour and iconography associated with the post- impressionist artist (Figure 6).

31.5 per cent of all image output fell into the category of ‘Fantasy’, which includes futur- istic, otherworldly or mythological characters and scenes; not surprising since Midjourney and Discord are popular among gamers and game designers (Jordan 2024) with more than 60 per cent of Discord users being under the age of 34 (Abreu 2023). Fantasy images imitated photog- raphy as well as stylized or illustrative approaches. At times, the prompt was clearly about a character or narrative from a known text, mythology or entertainment genre; but in some cases, the prompt did not specify anything inherently ‘fantastic’ yet the AI generated a visual treatment that implied the unreal or supernatural nonetheless. Most of the illustrations in this genre were derivative of existing video game motifs or well-trodden themes from fantasy films (Figure 7).

AI is capable of generating eerily perfect human ‘specimens’ and sometimes depicted persons that bore facial traits similar to known persons, while not necessarily being a portrait per se. ‘Fashion’ images and celebrity portraits (Figure 8) were uncannily convincing – perhaps in part because cultur- ally we are accustomed to highly retouched images in pop culture, advertising and celebrity photos. While persons operating in the public eye would likely be accustomed to their images appearing in tabloids or other news feeds, falsified portraits raise ethical and legal issues, as in the case of actress Scarlett Johansson who sued Lisa AI: 90s Yearbook & Avatar after her voice and likeness were simu- lated in an AI application (Hobbs 2023).

Six per cent of all sampled AI-generated images wholly or largely did not correlate with the prompt. Such aberrations may be akin to the ‘hallucinations’ that have plagued Chat GPT users, or they may be a by-product of prompts laden with contradictory elements or inclusion of lesser- known artists. For instance, the prompt for Figure 9 specified both Van Gogh and Art Deco – two very dissimilar aesthetics. The resultant image has neither the impasto nor chromatic qualities of Van Gogh, nor the dedication to the streamlined symmetry of Art Deco. The prompt also requested a ‘Jazz Age nightclub’, not a street scene, so it would have been further categorized as ‘Prompt uncor- related/poorly correlated’.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

174 journal of illustration

Figure 6: Images created using Midjourney by users writing the following prompts: (left): ‘A mesmerizing image of an extraordinary paper, quilling creation, depicting a 3-D flower sculpture with incredible attention to detail and realism. The artist employs a variety of quilling techniques and subtle color variations to re-create the flowers leaves patterns, delicate antennae, and graceful form. White background. The composition highlights, the innovated fusion of natural beauty and Intricate artistry, inviting….’ (right): Vincent van Gogh inspired oil painting of sunflowers with a brilliant blue butterfly sitting peacefully on the love…’ While the image on the left does a credible job of mimicking the craft of quilling, the image on the right, generated from a prompt that asked for Van Gogh, was not very convincing. Fair use.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 175

Figure 7: Images created using Midjourney by cryptophrankv and Beer Monkey writing the prompts visible beneath each. Fantasy images made up almost 1/3 of the total images requested on Midjourney.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

176 journal of illustration

Figure 8: Images created using Midjourney by Hanzo and daohuozhe428... writing the prompts visible beneath each. AI fashion ‘photography’ and celebrity portraits are often convincing.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 177

Racial and gender bias in Gen AI

As data-driven output aggregated from multiple sources of information, Gen AI might seem to oper- ate outside of human sentiment, thus minimizing the likelihood of advancing subjective viewpoints. On the contrary, this current research finds that databases inherit human-designed metrics and are, thus, biased, like search engines that categorize and prioritize results in ways that perpetuate gender

Figure 9: Image created using Midjourney by user writing the following prompt: ‘Art Deco angel: a woman in sunglasses stands in a glittering jazz age nightclub with a Van-Gogh style background of…’ Fair use.

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

178 journal of illustration

and racial discrimination (Birhane and Cummins 2019; Noble 2018; Quach 2023). A 2021 study of the LAION-400M dataset (filtered through Open AI’s dataset training [CLIP] system) found that the dataset contained ‘troublesome and explicit images and text pairs of rape, pornography, malign stereotypes, racist and ethnic slurs, and other extremely problematic content’ (Birhane et al. 2021: 1).

Further, stock image providers act as defacto gatekeepers that select work considered viable in the marketplace and, therefore, influence what photographers and artists make. In a 2020 study of Shutterstock and Getty Images, researcher Fernanda Carrera concluded stock images uphold racial, gender prejudices and ‘use of stereotypes, unequal and discriminatory narratives, as well as hyper-ritualizations of culturally determined behaviors’; reifying stereotypes of attributes like kind- ness, aggressiveness, beauty and ugliness (Carrera 2020: 225). Since stock images were scraped for AI image training, Gen AI is likely to propagate biased narratives from its sources unless they are expressly managed to avoid defaulting to stereotypes. This is what Google attempted with its Gemini AI image generator, which was abruptly paused on 22 February 2024 after it drew criticism from users because it inaccurately depicted historical subjects by including diverse races among the American ‘founding fathers’ that were certainly not present in the eighteenth century (Robertson 2024). According Ars Technica, Google’s very public glitch was likely caused by the same kind of code that Open AI has discreetly used since July 2022 ‘whereby its system would insert terms reflecting diversity (like “Black”, “female” or “Asian”) into image-generation prompts in a way that was hidden from the user’ (Edwards 2024).

A key part of my research project was to assess whether such race and gender biases are evident in the process of AI image generation. To test this, I collected data on how often people of colour were prompted and how they were represented. Of the 1795 images, only 905 contained people and only 186 (20.5 per cent) of them included persons of colour (Figure 10). People of Asian descent were the most prevalent BIPOC group represented (Figure 16).

These results are statistically significant because Midjourney is a US-based company and accord- ing to the 2020 United States Census, 40.7 per cent (USCensus.gov 2020) of the population in the United States self-reports as non-white. It should be noted that Midjourney can be accessed from outside the United States, and users are not identified by heritage. However, the relatively low percentage of images that include persons of colour suggests racial bias since the total global popu- lation and hypothetical users outside the United States are majority non-white.

Further, when positive modifiers (beautiful, good, cute) were specified, and there was no mention of race or ethnicity in the prompt, over 70 per cent of images defaulted to white subjects. Conversely, when persons of colour or of a particular ethnic group were shown, the prompt had specified that cultural or ethnic group in 70 per cent of cases.

Certain prompts lumped together disparate ethnic groups. The prompt for Figure 12 (left) called for a cybernetic ‘Aztec, Egyptian, Moroccan Empress’ as if to say all those cultures are

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 179

indistinguishable. The resultant hodgepodge of costume included a Renaissance-style bodice with a fantasy feathered headdress that was not culturally specific. When the AI generator was instructed to illustrate persons of colour, 38 per cent of those prompts also included language that exoticized the character or included modifiers that specified physical exaggeration or other stereotypes. For exam- ple, Figure 12 (centre) prompted for an ‘African American female, very supple chest with lustrous skin and narrow waist’. The artists specified were Alphonse Mucha (1860–1939), an Art Nouveau era poster artist known for his highly decorative portrayals of women, and Milo Manara (b. 1945), a contemporary Italian comic book artist known for his erotic images of women. This suggests the

Figure 10: Representation of people of colour specified in prompts. Graph by Susan Doyle.

P.O.C. in Al Generated Illustration Sample

Total Images 1795

Images with People 905

Images showing POC 186

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

180 journal of illustration

Figure 11: Representation by ethnicity specified in prompts. Table by Susan Doyle.

Figure 12: Images created using Midjourney by BrandonMc, printsmith, and ZacharyTiger writing the prompts visible below each. All three images exemplify problematic bias including the blurring of cultural distinction and/or overt sexualization of females.

P.O.C. Representation by Ethnicity/Race in Al Generated Sample Middle N.American

Asian African Eastern Indian Egyptian Latino-a Indigenous Multi Gypsy "Pygmy" TOTAL

113 51 3 1 1 3 1 11 1 1 186

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 181

prompter’s intention was to explore a hybridized stylization of seductive female beauty, but incon- gruously the AI generator surrounded the ‘American’ female with lions rather than American iconog- raphy – which suggests a keyword/database connection that problematically links ‘African’ with lions – even in the context of ‘African American’.

Gender bias was clear in many images: 9.3 per cent included suggestively posed or physically distorted females. They were in fact four times more likely than males to be sexualized whether or not it was explicitly requested in the prompt. For instance, the prompt for Figure 12 (right) asked for a ‘photo’ of a ‘flirty instagram [sic] influencer, coy expression and full body, Sakimichan, WLOP’. Sakimichan is a popular digital artist who typically shows females as comically buxom with tiny waists, and WLOP is an ArtStation artist known for scantily clad, fantasy females – thus, the popular art references predicted the sexualized image results by expressly requesting it.

The ethics of artistic derivation

Controversy over the ethics of AI training on the works of artists is ongoing, not just because artists’ works were scraped from the web without their consent but also because colossal databases allow prompters to request an image ‘in the style of -----’ (insert artist name here). My analysis tracked when individual artists were named in the prompt and when an artistic style or genre was requested. As noted above, prompts that combined several artists were not uncommon and some images did not correlate with the style prompts, which may be attributed to a paucity of training samples or inaccurate keyword indexing.

Of art historical styles, Art Deco and Art Nouveau were the most frequently prompted, while Pixar and Disney were the most requested contemporary art studio ‘styles’ to be named. Anime – now a global phenomenon with Japanese roots – came in third. Pop-culture trends like ‘Steampunk’ appeared often as well. An interesting prompt technique is to add ‘core’ to any noun to signal an aesthetic ‘genre’ – presumably based on commonly accepted stereotypes (‘grandmacore’ ‘princess- core’, etc.). Of living artists, Artgerm and James Jean (b. 1979) were the most prompted in my sample. Of those deceased, Alphonse Mucha (1860–1939) and Jean Henri Gaston Giraud (aka Moebius) (1938–2012) were the most frequently invoked.

Plagiarism vs. creativity?

AI-generated images do not copy or composite exact elements from specific images the way a human might collage an image together. Rather, AI facilitates synthesis of many instances of a subject, artists’ styles or artistic genres. Therefore, one might question if these visual renditions are techni- cally plagiarism in the traditional sense. We can certainly sympathize with artists whose works were scraped without authorization and perhaps argue that they are due to some recompense because

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

182 journal of illustration

their art is in some way deployed within the algorithms. But, how do we reconcile this new paradigm with existing laws that were not designed to deal with this innovation in making? Is this output truly a new kind of creativity?

Notwithstanding that thinkers have grappled for centuries with the notion of what ‘creativity’ is, AI researchers and developers promote the idea that creativity involves making ‘something new and useful’ (Haase and Hanel 2023). But successful creative acts also involve problem definition (articu- lating what one wishes to make, show and discover) as well as evaluative, critical judgement of the output in order to consider whether it communicates the essential idea in an aesthetically appropri- ate way. At this point in time, AI does not initiate requests for art nor consider or interpret source material in the way a human does. Rather, through learned keyword + image pairs, AI’s digital neural networks predict what is being asked for in the prompt (Hardesty 2017). Only through iterative trial and error initiated by humans does the AI refine or improve versions of the output.

Since ‘problem definition’ is analogous to prompting, and the application of evaluative criteria remains a human prerogative, in current Gen AI use, two of three fundamental aspects of a ‘creative act’ are still performed by a human, even though the output is generated by an AI. This more or less aligns with the findings of Keith Kirkpatrick at the Association for Computing Machines (ACM.org), who qualifies AI as a catalyst to creativity, stating that AI is capable of ‘producing unique combina- tions of familiar ideas, creating new works based on the attributes of previous works, and by offer- ing new ideas based on combinations of attributes and ideas that humans may not have thought of during the creation of a new work’ (Kirkpatrick 2023: 21–23).

I interviewed several practitioners for their views. According to American illustrator Lars Grant- West, who creates works for a high-profile game company,

Having a trained eye is still important, and some who call themselves ‘creators’ of these things, can’t pick the good ones or the bad ones. So, a lot of times you’ll see dozens and dozens of pictures of the same thing. It’s just people being agog over the level of finish and not being trained enough to understand what, or where the flaws are.

Kirsten Zirngibl, a digital artist and game designer who is proficient at AI, expands on its uses:

Generative AI can be integrated with every phase of the creative process. Ideation (probing the boundaries of the possibility space), iteration (design variations), rendering (3D render or line drawing diffused to a tighter full-colour presentation) and asset creation for compositing (photobashing). In the realm of 3D, it can be used to generate crude models and textures for game level prototyping.

(Zirngibl 2024)

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 183

Like Grant-West, Zirngibl points out that human artists using AI benefit from strong abilities to paint, draw and solve design problems because software programs do not have the ability to critique the efficacy of output or know how to correct problematic aspects of images. Humans with artistic expertise are still key in finalizing more nuanced aspects of the artistic outcomes.

Practical and legal impacts of AI

According to Artisana, an organization that tracks developments in Gen AI, deployment of new AI tools has been a major factor in a 70 per cent reduction in illustration jobs in the Chinese game industry (Zhang 2023a). This contraction was in part caused by government video game licensing controls aimed at curtailing computer gaming. Some artists who have not lost their jobs outright have been relegated to correcting AI-generated works at a lower compensation rate.

This trend increases fear among artists, as well as educators globally whose students hope to enter the gaming, animation and movie fields. Entry-level employment in these areas often includes time- intensive work like asset development or iteration of characters or scenes in visual development, and those jobs, predominantly digital in nature, could perhaps be more economically performed by Gen AI, even after factoring in the cost of humans to edit thousands of AI-generated sketches.

One thing slowing a seismic shift towards replacing artists with Gen AI in the United States is copyright regulation that requires ‘human authorship’ for granting any application for US copy- right. That law was tested in the case of Zarya of the Dawn, a comic written and designed by Kris Kashtanova. When it was disclosed that the comic’s images were made by Midjourney AI, the copy- right office rescinded the original image protections citing that the illustrations were ‘not the product of human authorship’; but they affirmed Kashtanova’s copyright for the written text and the book design (Edwards 2023).

This ruling presents an obstacle for animation, gaming or movie studios seeking to use Gen AI to output ‘finished art’ for larger projects. Under the current laws, such work will be difficult to defend against copying. However, one can foresee a future in which highly capitalized studios win protections for AI-generated content created through proprietary databases or that use patented AI software trained on their own ‘house style’. Until then, creative entities that migrate artistic output to Gen AI are likely to see it appropriated by competitors, leading to cross-bred content in enter- tainment genres that favour the bottom line over human inspiration and the more time-consuming alchemy of artistic exploration.

As illustrators test the legal boundaries for Gen AI copyright, legal actions pertaining to the unauthorized scraping of online visual content into AI training databases are ongoing. A class action lawsuit was brought in 2023 by artists whose art was appropriated by AI companies including Stability AI, Midjourney and DeviantArt. The suit asserts that over 4700 artists’ works have been used

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

184 journal of illustration

to train AI generators (Richardson 2024). In another case, the royalty-free stock image archive Getty Images claims that millions of its images were illegally scraped by Stable Diffusion/Stability AI under the aegis of ‘fair use’ (Vincent 2023).

In addition to concerns over AI database origins, plagiarism in the sense of falsified authorship using AI is also a concern, and one that can have swift economic implications in the rapid online marketplace. A recent case involved illustrated pocket calendars purportedly created by French illus- trator Jean-Baptiste Monge that appeared on Amazon.com. The products were not copies of his existing images; rather, they were AI-generated versions of his style. Monge posted a screenshot on Facebook in protest and the items were eventually removed from Amazon (the author reached out to JB Monge who declined to comment for this essay).

AI and the future of illustration

In March 2023, a petition was signed by more than three hundred fifty technology and AI inven- tors warning of the dangers of unregulated AI from political, ethical, commercial and global safety perspectives (Anon. 2024b). Ironically, some of the same tech/media giants who warned of the dangers of AI continue to lead and benefit from ongoing AI research, including Sam Altman at Open AI and Dario Amodei, CEO of Anthropic. A year later, in March 2024, the European Union adopted legal parameters around AI development (Pressroom, European Parliament; Anon. 2024a). To date the United States has passed no federal regulation of AI development or practices, although US President Biden signed an ‘Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence’ (Briefing Room, Presidential Actions, Whitehouse; Anon. 2023) in October 2023 that set a 270-day deadline for the development of guidelines and safety practices. Until legislation is actually enacted, and likely thereafter, the illustration field will be left to grapple with its own use of AI image generation whether for its innovative potential, editorial convenience and cost savings, or for artistic inspiration.

As AI image generation becomes more prevalent in visual culture, we will no doubt witness and in some ways be collectively responsible for which illustration practices remain viable. All of the artists I interviewed, whether they are AI users or not, believe that Gen AI will impact certain areas of illustration more than others. Kirsten Zirngibl predicts that:

Gen AI will still likely knock out a significant chunk of low-specificity entry-level projects, gigs done for small publishers/producers. Examples include small press tabletop game inte- rior illustrations, book covers, spot illustrations, album covers, and marketing content […] any place where the client’s budget is tight and they don’t need something hyper-specific.

(Zirngibl 2024)

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 185

Boston illustrator John S. Dykes (who does not use Gen AI) is philosophical over whether it is a bad thing if AI takes over jobs like spot illustrations that are currently ‘good enough’ and mostly serve ‘a decora- tive function’. Dykes predicts AI will be more threatening to artists whose work involves photo-based manipulation or that have a graphic style – but perhaps less so for those with an organic ‘hand-feel’. He concludes: ‘On the bright side, maybe in the longer term after the initial shock, from an illustration standpoint, AI might raise the quality. What is the adage: a rising tide raises all ships?’ (Dykes 2024).

Grant-West, who teaches digital art and character design at Rhode Island School of Design, has experimented with AI quite a bit and is less sanguine:

The problem is that an AI can make images in a fraction of the time it would take a human to draw or paint an iteration of an idea. So, I have not quite resolved how to compete with it, or if it’s even possible to compete on its own level; but it’s the first time you’ve ever had all three: the good, quick and cheap.

(Grant-West 2024)

Along with my colleagues, as an art educator I have concerns not only for the future of young artists entering the field who may be displaced by AI image generation tools but also for the impact that Gen AI will have on teaching and learning for art students overall. Art is a way of knowing. Anyone who has studied art or art history knows mastery over materials is essential to the development of one’s artistic voice. A critical eye develops only through looking at a lot of art over time, whether one is a painter, draftsman, ceramist or furniture maker. Unique works of art involve trial and error, deep engagement with practice, and patience. Students are already inundated by art possibilities on the internet and they want results fast, so tools like Gen AI have the potential to undermine time- intensive but important aspects of traditional training unless students understand the value of learn- ing the fundamentals through personal practice. It is up to instructors to help them set boundaries and monitor AI use by students.

AI generation presents a particular challenge in teaching Illustration because the 2D digital envi- ronment is often where students learn to draw and paint using programs like Adobe Photoshop, which now embeds its ‘Firefly’ image generator into the Photoshop program start-up menu, thus encouraging students to initiate their artwork with AI exploration. Bundled in Creative Cloud, Firefly generated 1 billion images in the first three months after its launch in May 2023 (Valyaeva 2023).

Artists, educators and researchers who care deeply about the impact of images on thought may not agree on the future of AI image generation technology, but the centrality of humans in shaping visual expression is undisputed. While those in the vanguard like Kirsten Zirngibl speculate that Gen AI may bring ‘an artistic renaissance fuelled by this tech in the hands of those with vision, where art will expand in new ways/depths that are hard for anyone to imagine now’, veteran illustrator Anita

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

186 journal of illustration

Kunz expresses an important proviso: students must remember that ‘[t]hey’re not a programme, they are a sum of their whole life experience’ (Kunz 2024).

Acknowledgements

My thanks to Anita Kunz, Lars Grant-West, John S. Dykes, Kevin Mutch and Kirsten Zirngibl, who generously offered their time and expertise in discussing AI from their unique artistic perspectives.

References

Abreu, Yerain (2023), ‘How to create AI art with Midjouny (sic) (easy mode)’, Maker Hub, 2 March, https://blogs.baruch.cuny.edu/makerhub/how-to-create-ai-art-with-midjounry/. Accessed 23 June 2024.

Anon. (2023), ‘Executive order on the safe, secure, and trustworthy development and use of Artificial Intelligence’, The White House, 30 October, https://www.whitehouse.gov/briefing-room/presi- dential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development- and-use-of-artificial-intelligence/. Accessed 19 April 2024.

Anon. (2024a), ‘Artificial Intelligence Act: MEPs adopt landmark law’, Official Pressroom European Parliament, 3 March, https://www.europarl.europa.eu/news/en/press-room/20240308IPR19015/ artificial-intelligence-act-meps-adopt-landmark-law. Accessed 15 March 2024.

Anon. (2024b), ‘Statement on AI Risk’, Center for AI Safety, 30 May, https://www.safe.ai/work/state- ment-on-ai-risk. Accessed 4 June 2024.

Birhane, Abeba and Cummins, Fred (2019), ‘Algorithmic injustices: Towards a relational ethics’, arxiv, 16 December, https://arxiv.org/abs/1912.07376. Accessed 17 June 2023.

Birhane, Abeba, Uday Prabhu, Vinay and Kahembwe, Emmanuel (2021), ‘Multimodal datasets: Misogyny, pornography, and malignant stereotypes’, arxiv, 5 October, https://arxiv.org/abs/2110.01963. Accessed 6 April 2024.

Bond, Shannon (2023), ‘Fake viral images of an explosion at the pentagon were probably created by AI’, NPR, 22 May, https://www.npr.org/2023/05/22/1177590231/fake-viral-images-of-an-explo- sion-at-the-pentagon-were-probably-created-by-ai. Accessed 20 March 2024.

Carrera, Fernanda (2020), ‘Race and gender of aesthetics and affections: Algorithmization of racism and sexism in contemporary digital image banks’, Matrizes, 14:2, pp. 217–40, https://www.revis- tas.usp.br/matrizes/article/view/167187.

Dykes, John S. (2024), interview with S. Doyle, 28 March. Edwards, Benj (2023), ‘AI-generated comic artwork loses US Copyright protection’, Ars Technica, 23

February, https://arstechnica.com/information-technology/2023/02/us-copyright-office-with- draws-copyright-for-ai-generated-comic-artwork/. Accessed 5 June 2023.

https://blogs.baruch.cuny.edu/makerhub/how-to-create-ai-art-with-midjounry/

https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/

https://www.europarl.europa.eu/news/en/press-room/20240308IPR19015/artificial-intelligence-act-meps-adopt-landmark-law

https://www.safe.ai/work/statement-on-ai-risk

https://arxiv.org/abs/1912.07376

https://arxiv.org/abs/2110.01963

https://www.npr.org/2023/05/22/1177590231/fake-viral-images-of-an-explosion-at-the-pentagon-were-probably-created-by-ai

https://www.revistas.usp.br/matrizes/article/view/167187

https://arstechnica.com/information-technology/2023/02/us-copyright-office-withdraws-copyright-for-ai-generated-comic-artwork/

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Generative AI and illustration

www.intellectbooks.com 187

Edwards, Benj (2024), ‘Google’s hidden AI diversity prompts lead to outcry over historically inaccu- rate images’, Ars Technica, 22 February, https://arstechnica.com/information-technology/2024/02/ googles-hidden-ai-diversity-prompts-lead-to-outcry-over-historically-inaccurate-images/. Accessed 9 April 2024.

Grant-West, Lars (2024), interview with S. Doyle, 29 March. Haase, Jennifer and Hanel, P. H. (2023), ‘Artificial muses: Generative artificial intelligence chat-

bots have risen to human-level creativity’, arXiv, https://arXiv:2303.12003. Accessed 23 April 2023.

Hardesty, Larry (2017), ‘Explained: Neural networks, Ballyhooed artificial-intelligence technique known as “deep learning” revives 70-year-old idea’, MIT News Office, 14 April, https://news.mit. edu/2017/explained-neural-networks-deep-learning-0414. Accessed 19 March 2024.

Hobbs, Jack (2023), ‘Scarlett Johansson takes legal action against AI app ripping off her likeness’, New York Post, 2 November, https://nypost.com/2023/11/02/entertainment/scarlett- johansson-takes-legal-action-against-ai-app-ripping-off-her-likeness/. Accessed 24 June 2024.

Kirkpatrick, Keith (2023), ‘Can AI demonstrate creativity?’, Communications of the ACM, 66:2, pp. 21–23, https://cacm.acm.org/magazines/2023/2/268946-can-ai-demonstrate-creativity/fulltext. Accessed 19 June 2023.

Kunz, Anita (2024), zoom interview with S. Doyle, 27 March. Minor, Jordan (2024), ‘What is discord and how do you use it?’, PC Magazine, 13 March,

https://www.pcmag.com/explainers/what-is-discord-and-how-do-you-use-it. Accessed 23 June 2024.

Noble, Safiya Umoja (2018), Algorithms of Oppression: How Search Engines Reinforce Racism, New York: New York University Press.

Quach, Katyanna (2023), ‘4chan and other web sewers scraped up into Google’s mega-library for training ML’, The Register, 20 April, https://www.theregister.com/2023/04/20/google_c4_data_ nasty_sources/. Accessed 19 June 2023.

Richardson, Aran (2024), ‘The 4,700-artist AI controversy: Artists accuse Midjourney and other AI firms of unauthorized use in lawsuit’, Yahoo Finance, 29 February, https://finance. yahoo.com/news/4-700-artist-ai-controversy-182916995.html?guce_referrer=aHR0cH M6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAALf7gY47qAca89bLYhMbh 9h9ashGr7-BM7fLgA24y5VywJDkqddu__hEmKlaD8b4AUv2aR-1-LGLccY0qGrrVYtO_ LA19ZcCT0NpfGHEERNZCnkvJl0gJWfK_ODbepRSjTKNPibtyAowtn1KEQ4svEPjvNdRsA7Kth WOYOKJePW9&guccounter=2. Accessed 8 April 2024.

Robertson, Adi (2024), ‘Google apologizes for “missing the mark” after Gemini generated racia- lly diverse Nazi/ Generative AI has a history of amplifying racial and gender stereotypes – but

https://arstechnica.com/information-technology/2024/02/googles-hidden-ai-diversity-prompts-lead-to-outcry-over-historically-inaccurate-images/

https://arXiv:2303.12003

https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414

https://nypost.com/2023/11/02/entertainment/scarlett-johansson-takes-legal-action-against-ai-app-ripping-off-her-likeness/

https://cacm.acm.org/magazines/2023/2/268946-can-ai-demonstrate-creativity/fulltext

https://www.pcmag.com/explainers/what-is-discord-and-how-do-you-use-it

https://www.theregister.com/2023/04/20/google_c4_data_nasty_sources/

https://finance.yahoo.com/news/4-700-artist-ai-controversy-182916995.html?guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAALf7gY47qAca89bLYhMbh9h9ashGr7-BM7fLgA24y5VywJDkqddu__hEmKlaD8b4AUv2aR-1-LGLccY0qGrrVYtO_LA19ZcCT0NpfGHEERNZCnkvJl0gJWfK_ODbepRSjTKNPibtyAowtn1KEQ4svEPjvNdRsA7KthWOYOKJePW9&guccounter=2

Delivered by Intellect to:

Washington University in St. Louis (wustl)

IP: 128.252.67.66

On: Mon, 24 Mar 2025 14:26:12

Susan Doyle

188 journal of illustration

Google’s apparent attempts to subvert that are causing problems, too’, The Verge, 21 February, https://www.theverge.com/2024/2/21/24079371/google-ai-gemini-generative-inaccurate-historical. Accessed 26 February 2024.

United States Census Bureau Official Website (2020), https://www.census.gov/quickfacts/fact/table/ US/RHI225221. Accessed 2 June 2023.

Valyaeva, Alina (2023), ‘AI has already created as many images as photographers have taken in 150 years. Statistics for 2023’, Every Pixel Journal, 15 August, pp. 1–7, https://journal.everypixel.com/ ai-image-statistics#:~:text=More%20than%2015%20billion%20images,reach%20the%2015%20 billion%20mark. Accessed 15 March 2024.

Vincent, James (2023), ‘Getty images is suing the creators of AI art tool stable diffusion for scra- ping its content’, The Verge, 17 January, https://www.theverge.com/2023/1/17/23558516/ai-art- copyright-stable-diffusion-getty-images-lawsuit. Accessed 16 March 2024.

Zhang, Michael (2023a), ‘China’s video game AI art crisis: 40x productivity spike,70% job loss’, Artisana, 12 April, https://www.artisana.ai/articles/chinas-video-game-ai-art-crisis-40x-produc- tivity-spike-70-j. Accessed 8 June 2023.

Zhang, Michael (2023b), ‘Artist refuses prize after his AI image wins at top photo contest’, Petapixel, 14 April, https://petapixel.com/2023/04/14/artist-refuses-prize-after-his-ai-image-wins-at-top- photo-contest. Accessed 17 April 2023.

Zirngibl, Kirsten (2024), e-mail correspondence with S. Doyle, 12 April.