ARTificial Intelligence: opportunities for art librarians to engage students’ critical thinking with AI image generators

Abstract

Art librarians work with images. It’s one of the things that separates us from many of our fellow subject librarians. As the academy continues to grapple with the benefits, drawbacks, and effects of AI, art librarians are uniquely positioned to teach students how to critically engage with AI image generators. Considerations concerning copyright, bias in datasets, formal analysis, and AI image generators’ potential as an art medium are some examples of topics that art librarians have at their disposal.

Introduction

AI image generators, like DALL-E and Midjourney, currently exist in an ambiguous artistic space. On one hand, you have the negatives; images created solely through these applications have caused backlash when they’ve won juried awards, artists’ work was used without consent in AI image generators’ datasets, and who owns the copyright to images generated by AI is murky at best. The other hand is more positive; artists are using AI as an artist tool to great effect for both initial brainstorming and finished artworks. How can art librarians prepare their students for a future where AI is a certainty when AI’s current relationship to art is so uncertain?

Critical discussions with students about AI image generators’ relationship to copyright, the biases inherent in AI image generators’ datasets, asking students to determine which image is AI generated and why, and questions that consider AI image generators’ potential place in the canon of Western art history are some examples of ways art librarians can engage their students in critical thinking about AI image generators. The first three are broadly applicable to not only art majors but the general student population. The other two are more specific to art disciplines. Regardless of the audience, these conversations can be held as full, one-shot class sessions or shorter discussions.

Ethics and Copyright

The first two instructional opportunities lay in how AI image generators deal with copyright and how they attempt to mitigate bias in the dataset. These topics extend beyond art majors to the broader campus community as they speak directly to topics all students need to know to understand how AI image generators work and their potential implications for society.

Art librarians can raise these ethical considerations by asking students to consider AI image generators’ copyright policies and publicly available critiques of those policies, including recent court cases and rulings. In the US, only human beings can own copyright. This means that people that create images with AI don’t necessarily own the copyright to that image. For example, the US Copyright Office has ruled that Jason Allen, who submitted a work generated from a prompt he authored in Midjourney to the 2023 Colorado State Fair and won in the amateur digital arts category, does not actually own the copyright to that image.¹ Anyone can print the image on any object they’d like and sell it for a profit. Knowing this might deter someone from using the software should they need, or want, to retain copyright.

While US copyright law is clear that only humans can own copyright, there does seem to be some murkiness in the area of what constitutes the intervention needed to make an AI image ‘made by a human.’ Kris Kashtanova used Midjourney to illustrate her 2023 graphic novel Zarya of the Dawn. The US Copyright Office ruled that while she owns the copyright of the written text and sequence of images in the novel, she does not own the copyright to the individual images made by Midjourney.² This suggests that there is some amount of artistic choice that qualifies as enough to merit ‘humanness.’ Kashtanova’s individual images can’t be copyrighted but the sequence chosen by the artist can. Additionally, artists’ work was used to train AI image generators without the artists’ consent - Getty Images is currently suing Stability AI³ and a group of artists are suing both Stability AI and Midjourney.⁴ Students that upload their work to AI image generators for manipulation might very well find their work added to the dataset without their consent, unless they read the terms and conditions.

Originally this idea was conceived as a stand-alone, one-shot lesson where student groups research individual AI image generators and present their findings to the class in a debate format.⁵ The following questions would guide research and discussion: ‘How do the various AI image generators compare and contrast in how they handle copyright?’, ‘Which AI image generator has the ‘best’ copyright for users? Why?’, ‘Which AI image generator has the ‘worst’ copyright for users? Why?’, ‘What does it mean if you didn’t find any information on copyright and your AI image generator? Why might this be problematic?’, and ‘Would you hesitate to use any of these tools based on how they handle copyright?’.

However, should the librarian not be able to dedicate an entire class period, the lesson can be shortened. For example, prior to class, the librarian could research various AI image generators’ copyright policies, supply them to students, then poll students asking whether they would use the AI image generator and ask them to explain why. Another exercise would involve asking students whether they knew that they don’t necessarily own the copyright to images they generate with AI, provide current examples from the US Copyright Office, and ask whether students would still use the generator.

Bias in the Dataset

AI image generators mirror and reinforce the biases held by the society that created them. DALL-E 2, for example, has implemented mitigations so that the prompt ‘C.E.O.’ is more racially and sexually diverse in its generations.⁶ These mitigation attempts can sometimes have unintended consequences. When DALL-E 2 took out sexualized content it removed many images of women, resulting in fewer images of women in the dataset in comparison to men.⁷ AI image generators also tend to privilege photography as a style, both in the results it returns and what it is best at.⁸ After all, AI image generators are trained on images on the internet, not reality. All oil paintings on the internet are photographs of oil paintings.

This learning opportunity was, like copyright, originally conceived as a full, one-shot lesson where students researched bias in AI image generators’ datasets in groups and then presented them to the class. Questions could include: ‘Does the AI image generator acknowledge bias in its dataset? If so, what biases are present? If not, can you find other sources that discuss bias in the AI image generator’s dataset? What are the biases?’, ‘Why might an AI image generator not acknowledge bias in its dataset and why might that be problematic?’, ‘What are the consequences of bias in the dataset?’, ‘Does the AI image generator attempt to thwart bias in its dataset? If so, how?’, ‘Are there other rules and/or regulations and/or restrictions in your AI image generator’s dataset?’ and ‘What else did you learn about your AI image generator’s dataset?’ The discussion could also be shortened. For example, the librarian could provide students with some ways datasets are biased and then ask students possible consequences.

Formal Analysis/Spot the AI

Formal analysis, a mainstay of art history, also provides an opportunity to engage students with AI image generators by asking them to consider formal elements and principles of design. As such, it might appear this opportunity is less applicable to non-art majors than discussions around copyright and bias in datasets. This is not the case.

The core of this lesson lies in another art history mainstay, the comparing and contrasting of images. Before class, the librarian uses an AI image generator to generate a two-dimensional artwork, like a painting or drawing, in the style of a famous artist. The consideration of which artist to choose is important and finding the right one might take some time. Bias in datasets takes many forms and, since datasets are created with images scraped from the internet, AI image generators are better at mimicking artists who appear frequently online. Then, compare the AI generated image to an artwork created by that artist, but not one that would immediately be known. Ask students to determine which image was generated by AI and which one was created by the artist and have them elucidate their reasons why.

While this discussion focuses on an example that uses art history, it could be applied to other disciplines. For example, one could substitute a photograph of a famous event in a history class. The discussion is also flexible enough that it can be broadened to a full class period or condensed into a short discussion. For an art student audience, you might ask them to pay particular attention to the formal elements and principles of design. However, as ‘formal elements’ and ‘principles of design’ aren’t terms often heard outside art and design classrooms, using them explicitly might confuse non-art students.

The identification of the ‘correct’ answer is not the overarching goal of this lesson and students should be made aware of that fact. While, at the time of this writing, it is still possible for those with a keen eye to determine at least some AI generated images, AI image generators are advancing so rapidly that whether this remains possible is questionable. Thus, it seems that demonstrating to students the challenges inherent in determining whether an image was generated by AI or created by a human is a more worthwhile pursuit.

AI and Photography

The final two opportunities, discussions that center around AI’s potential as an art medium in the Western canon, were created with art students in mind and developed around extending common discussions students already have in many art history classes. While they are more specific in nature than the previous three opportunities, they can still be applied to both upper division art courses, where students are likely art majors, and lower division art courses that are typically composed of both art and non-art majors, at least at a traditional US university.

Today we consider photography to be an art medium, however, this was not always the case. When photography was invented in 1822, the artworld didn’t immediately celebrate the invention of a new art medium. The question of whether photography was an art or a technology was hotly debated for years by both artists and critics. In fact, we see this debate echoed in our own Library of Congress Classification System, where photography is classed in the technology range, T, instead of the fine arts range, N. On one side, you had artists who bemoaned the death of painting now that photography could more easily fulfill paintings’ historical role of reproducing reality.⁹ On the other, you have critics and artists arguing that since photography removes the hand of the artist and depends on a mechanical process, photography cannot be considered art. AI image generators’ potential acceptance, or not, as an art medium and the line in between art and technology is akin to photography’s history. Like photography, the debate surrounding AI image generators, in part, hinges on whether the invention is a tool capable of art making.

Librarians can approach this opportunity, as those opportunities previously discussed, in one of two ways. They can either dedicate an entire class session, outline the history of photography, and discuss how AI could, potentially, become accepted as an art medium based on the following questions: ‘Should (or will) AI art ever be considered ‘real’ art?’, ‘How are AI image generators different from photography? How are they the same?’, ‘Should we consider AI image generators an artistic tool as well as a technological tool? Where would the line be between the two?’, ‘Do you think AI will cause digital photography to decline as a medium in favor of a return to film photography and darkroom processes?’, and ‘Can the artist ever be replaced by a machine or technology?’

Condensing this opportunity down to a shorter discussion is slightly limiting in the sense that it can only be done in classes where you know students have already discussed the invention of photography. Of course, this is not hard to determine with a short conversation with the professor. If the librarian knows their students are familiar with the arguments surrounding photography’s acceptance as an art medium, a short conversation can be initiated by briefly reminding students about the topic and asking some, or all, of the questions above.

AI and Conceptual Art (or What Would Duchamp Do?)

Another topic many art courses cover is Dada artist Marcel Duchamp, his invention of the readymade, and the idea that the concept behind the artwork is the most important element, superseding both aesthetics and craftsmanship. In 1917, Duchamp submitted his readymade Fountain, a urinal turned on its side and signed R. Mutt, to an exhibition at the Society of Independent Artists. Despite the exhibition not having a jury process, the work was not accepted. A defense of Fountain was penned in The Blind Man,¹⁰ a Dada journal Duchamp edited. The essay called into question the definition of art held until that time, arguing that art was not made by the artist’s skill or talent with material but rather the artist’s imagination. How does this concept relate to AI image generators where an image is generated based on the words, or idea, of the person requesting the image?

To break this down further for students, the librarian can ask the following questions: ‘Is the idea behind the art the most important thing? Why or why not?’, ‘How does this relate to AI? Does AI ‘count’ as Conceptual Art?’, ‘Do you think AI will make the hand of the artist more important?’, ‘Do you think AI will make craftsmanship more important?’, ‘What would Duchamp think of AI?’. As with the previous opportunities, librarians have two ways they can approach this discussion with students - either as a full, one-shot discussion or a shorter activity. The choice to shorten the activity is likely only advisable in courses where students have already discussed Dada and/or Conceptual Art.

Conclusion

AI image generators will undoubtedly change art, and the world, in many ways. While impossible to predict without the benefits of hindsight, art librarians have the opportunity to engage students’ critical thinking with these topics in the current moment. Some of these opportunities include discussions about ethics and copyright, bias in the dataset and formal analysis, topics that can be approached with students in any major. Other opportunities, likely more appropriate for art students, include considerations of how AI image generators might potentially affect the history of art by comparing and contrasting them to the acceptance of photography and conceptual art as legitimate art mediums. Given the unknown potentials of AI, students need to thoughtfully analyse its use.