Luo_Anita_Experimental Collaborative Work with GenAI (PDF)
Anita Luo
Dr. Jana Fedtke
PoH: AI-Assisted Writing, Agency, and Authenticity
October 16, 2024
Experimental Collaborative Work with GenAI
Project Proposal
In this project, I aim to use generative artificial intelligence (GenAI) tools to produce promotional material for tourism in South Africa. The graphic design process will involve using AI-powered tools built into Canva such as Magic Media and Magic Design which are Canva’s GenAI models for images, videos, and graphic designing. Canva makes designing more accessible to non-designers with its customizable templates and user-friendly interface. With this new set of AI-powered design tools, Canva is increasing its accessibility and potential use in the mass population. To evaluate its efficacy, I intend to create a poster of South Africa that is an accurate representation of the country and its features while evaluating how much producer oversight is required, in terms of text prompt lengths or rearrangement of design elements while editing, to achieve this result. By actively collaborating with these GenAI tools, this project is an opportunity mainly to analyze the advantages and shortcomings of GenAI in creativity, problem-solving, and decision-making; and to identify ethical implications in terms of intellectual property, bias, and the potential influence of AI-generated content in the future.
The Creative Piece
Throughout the creative process, I documented my interaction with the GenAI tool by saving drafts, screenshots, and logs of my prompts and the responses I received.
Magic Design
Firstly, I used Magic Design and typed a generic prompt “south africa tourism poster” to investigate how the GenAI Tool would interpret it. The outputs all consisted of natural landscapes of South Africa or the South African flag. Although nature reserves are one of the main tourist attractions in South Africa, I believe that the prevalence of nature visuals further exacerbates the existing stereotype that the country of South Africa mainly encompasses a safari scene.
For my next step, I edited my previous prompt and included the requirement to have various tourist attractions to explore the extent of the GenAI tool’s discriminatory output. I found that Magic Design had the limitation of only having one image in its outputs. Furthermore, I only had the option to put a one-time prompt so I cannot provide more prompts to further develop the initial generated outputs.
I specified my prompt further by adding “nature and city attraction” and “more than 2 images.” The purpose of doing this is to attempt to solve the challenges in the previous step. This time I received more variation in images. However, the limitation of having one image was still evident. Furthermore, I noticed that the text produced in the image was generic and could be applicable to any country advertised. Thus, in the next step, I aimed to resolve this issue.
In the next prompt, I wanted to explore Magic Design’s capacity to design and understand the context. Thus, I refined my prompt to include a specific design style that is contemporary and specific South African phrases. As a result, I found that there were improvements in the design’s complexity and arrangement of elements. Furthermore, I noticed that some outputs also contained the term “Rainbow Nation” which is a popular term used in South Africa.
My final choice for the design template is due to the relevance of the term “Rainbow Nation.” I dismissed the image as I intended to use Magic Media to add more visual elements.
The interface for Magic Media is easy to interpret. However, similar to Magic Design, the GenAI tool only allows one prompt at a time.
My first prompt was “South African locals showing tourists around South African tourist attractions with a smile on their faces.” Through this prompt, I intend to see a variety of races and tourist attractions. However, the outputs were limited to only the black race, and the location of the background was imperceptible.
As a result, I had to be decisive and particularize the output I wanted by stating that I wanted to generate content with South Africans of different ages and races in a particular place in the forest ziplining. Unfortunately, the outputs were still dominated by people of the black race.
To tackle the limitation of representation, I wrote what races I wanted which required prior knowledge of South Africa’s diverse population.
One particular shortcoming I noticed was the inability of Magic Media to put different groups of people of different demographics in one output.
On the other hand, I removed any words or phrases relevant to South Africa to test whether the bias towards the black race in the outputs was a result of the prompt having an association with South Africa. Thus, I found that the outputs displayed more diverse races including the white race. The results confirm that the AI model associates South Africa with the black race.
Unfortunately, Magic Design did not have a function to rearrange visual elements so I had to organize the media independently.
In order to test whether the AI model has the capacity to put people of different races into one image, I specified my prompts further. The outputs showed that it is possible to put people of different demographics together. However, the GenAI tool requires the user’s preference to do so and automatically groups people of the same background together. In a larger context, the AI model’s output is a reflection of a lack of diversity and inclusion culture between different groups of people.
As an experiment, I also test Dall-E. The tool displayed similar limitations to Magic Media concerning the strong association of South Africa to the black race, and the lack of inclusivity of different types of people.
Reflection
Generative AI (GenAI) tools massively reduce the time investment needed to create content. Thus, in a fast-paced world, GenAI tools such as Magic Media and Magic Design are transforming the ideation process for both designers and non-designers in many different disciplines in which any person can conceptualize an idea and create a proof of concept swiftly. However, with the assistance of AI, these GenAI tools allow us to explore how the generated outcomes redefine the perceived creativity of a piece of work, how it differs from human-only collaborative processes, and how it further complicates the existing authorship attribution problems in the world. As the role of the “author,” I aim to use Magic Media and Magic Design on Canva to create a poster that will promote features of South Africa, such as landscape, activity, and people, accurately. As a South African, I will be evaluating how much guidance, refining, and editing is needed from me to create a satisfactory promotional poster of my country that avoids potential bias, stereotypes, or misrepresentation.
For the process of this project, I first inputted prompts in Magic Design. My approach to inputting prompts was to first type a simple prompt that was open to many interpretations to investigate whether the GenAI tool had enough context to generate an accurate outcome. After finding drawbacks to the generated outcome and identifying areas in need of improvement, I would refine my prompts through additions and alterations to the previous prompts. This procedure was repeated until I found an output that was free from potential issues. Satisfactory outcomes were found through this iteration process. The same iteration process is used with Magic Media. In the final stage, I organized the media on the poster in a way that I deemed presentable and usable in a real scenario of tourism advertisement. Throughout the process, I worked closely with the GenAI tool as it required heavy producer oversight. I felt like a supervisor who adopted autocratic leadership. Firstly, all the ideas were developed by myself and there was little input from Canva in giving suggestions or feedback to me about alternatives or possible improvements to my initial concepts and prompts. There was no presence of a dialectic relationship between the GenAI tool and me—the interaction was purely on a utilitarian basis. In this case, I have the sole role of determining the direction of the project. In my opinion, the lack of dialogue with the GenAI tool is one of the biggest differences between AI assistance and human assistance. Nevertheless, there are many other GenAI tools with the capacity for dialogue. Secondly, Canva mostly produces less than satisfactory results. As a result, I am required to provide guidance, support, and identify development needs to Canva in terms of specifying what I envisioned to a large extent. Therefore, I have control over the generated outcomes.
On the one hand, Canvas algorithms can augment and emulate human work which saves money and time. For example, I do not have to take or hire other people to take photographs in South Africa. Furthermore, I do not have to go through any procedures and spend time to ensure I am using copyright-free material—or to ensure I have my own intellectual property right to the advertisement. According to Canva’s AI Product Terms, users who utilize Canva’s Magic Studio features for commercial purposes may not have exclusive rights to the generated outputs. Additionally, there are instances in which the generated images have challenged my ideas by producing unforeseen visual elements that can enhance the poster’s purpose. For example, for one of the prompts I only included “Cape Town.” However, it generated Table Mountain which I did not envision initially. I accepted the image due to the fact that Table Mountain is one of the classical landmarks in South Africa. Thus, this was a moment where the AI’s contribution felt truly “creative.”
On the other hand, there are many drawbacks to achieving my creative goal of producing an accurate and reliable advertisement of South Africa. I have mentioned before that there is a high demand for producer oversight through increasing specificity in prompts. This high producer oversight is also due to a limited prompt input option—you cannot refine images with follow-up prompts. I also encountered, as Joy Buolamwini calls it, the “coded gaze” which refers to algorithmic bias. I discovered when my prompts included “South Africa” the generated images only consisted of people of the black race. When I specified for people of different races and ages, the GenAI tool continued to generate similar results as before. Thus, this experience informs us about how AI is shaped by and shapes society. By having preconceived and oversimplified images of a “mid-tier” or “secondary” country such as South Africa, which may have less focus in the datasets that train the AI models, I am concerned about the potential misrepresentation and misinformation that these tools may provide to people outside the country—exacerbating the misunderstanding between different groups of people. Furthermore, there was also the tendency for the GenAI tool to group people of similar demographics together—a reflection of a lack of diversity and inclusion culture in society. Relating to ethical implications, there has been an increasing concern that algorithms used by GenAI tools produce discriminative outputs as they are trained on data that is embedded with societal biases. Additionally, Canva discloses that it does not provide any guarantee that the content generated is cleared for use, particularly if the output reproduces text and images from existing works. There are potential challenges in ensuring that all outputs generated are free of legal risks too.
Lastly, the application of Canva’s GenAI tools for commercial use problematizes authorship attributions. Canva requires users to disclose that AI has been used to generate content according to their sharing & publication policy. However, producers are found to receive more credit for work when they are assisted by algorithms, compared to assistance from humans (Jago and Carroll). Through my experience of using Magic Design and Magic Media, I strongly agree that algorithmic assistance demands more producer oversight than human assistance—at least for simple and accessible GenAI tools such as those of Canvas. Despite my contributions, from a legal perspective, I am not assigned credit. In the greater context of distribution, compensation, and accountability, authorship is crucial. Therefore, tension is found between the social perceptions of authorship attributions and the actual legal authorship attributions in the area of AI. Thus, in light of the possible legal risks of claiming intellectual property as discussed before, I believe that the creative industry could embrace this new phenomenon of a free, “non-proprietary” domain of content creation separate from the domain characterized by patents, copyrights, and trademarks.
In this project, I have outlined the advantages and disadvantages of using GenAI tools such as Magic Design and Magic Media on Canvas and their impact on the traditional process of content creation. Despite the GenAI tools’ ability to reduce work time and money, they exhibit and contribute to existing societal biases in their outputs; they particularly showcase racist underpinning and lack of diversity. Artificial neural networks may be both affected by and affecting the masses—“a cycle of bias propagation between society, AI, and users” (Vlasceanu and Amodio). In order to mitigate the discriminatory outputs, a high level of producer oversight is required to control the generated content within the objective of the project. Thus, I believe that these AI tools should be seen as “extensions of human organs” (Agüera y Arcas 5). The project contributes to the discourse of AI authorship attributions in which there is a limitation of having a copyright over the generated content in spite of the level of contributions of humans and AI. However, as a result, a new form of freedom is introduced in the creative sector.
Works Cited
Agüera y Arcas, Blaise. “Art in the age of Machine Intelligence.” Arts, vol. 6, no. 4, 29 Sept. 2017, p. 18, https://doi.org/10.3390/arts6040018.
Buolamwini, Joy. “How I’m Fighting Bias in Algorithms.” YouTube, TED, 30 Mar. 2017, www.youtube.com/watch?v=UG_X_7g63rY&t=156s.
Jago, Arthur S., and Glenn R. Carroll. “Who made this? algorithms and authorship credit.” Personality and Social Psychology Bulletin, vol. 50, no. 5, 3 Feb. 2023, pp. 793–806, https://doi.org/10.1177/01461672221149815.
Pavlick, Ellie. “From the MIT GenAI Summit: A Crash Course in Generative AI.” YouTube, MIT AI ML Club, 14 Mar. 2023, www.youtube.com/watch?v=f5Cm68GzEDE&ab_channel=MITAIMLClub.
Vlasceanu, Madalina, and David M. Amodio. “Propagation of societal gender inequality by internet search algorithms.” Proceedings of the National Academy of Sciences, vol. 119, no. 29, 12 July 2022, https://doi.org/10.1073/pnas.2204529119.