Selected Papers of #AoIR2024: The 25th Annual Conference of the Association of Internet Researchers Sheffield, UK / 30 Oct - 2 Nov 2024 Suggested Citation (APA): Miltner, K.M. (2024, October). “A.I. is holding a mirror to our society”: Lensa and the Discourse of Visual Generative AI. Paper presented at AoIR2024: The 25th Annual Conference of the Association of Internet Researchers. Sheffield, UK: AoIR. Retrieved from http://spir.aoir.org. “A.I. IS HOLDING A MIRROR TO OUR SOCIETY”: LENSA AND THE DISCOURSE OF VISUAL GENERATIVE AI Kate M. Miltner University of Sheffield Introduction Since the release of the open-source AI model Stable Diffusion in August 2022, a series of apps that create AI-generated portraits, or avatars, have exploded in popularity. One of the most notable examples is AI-powered photo editing app Lensa, owned by Prisma Labs. Lensa went viral in late 2022 due to the launch of its Magic Avatar feature, which uses the Stable Diffusion model to create fantastical (and occasionally bizarre) portraits of users in a series of themes, such as “Fairy Princess”, “Superhero”, and “Astronaut”. After the Magic Avatar launch, Lensa was installed over 13.5 million times worldwide and users spent over USD$29 million on the app in 12 days (McCluskey, 2022). In response to Lensa’s international popularity, a global discourse developed that both hyped and critiqued the Magic Avatar phenomenon. This paper analyzes the global English-language press coverage of Lensa and finds that it echoes existing technological discourses, focusing on the app’s predatory data practices, the biased content it produced, and the user behaviors associated with it. I argue that this coverage provides evidence of discursive closure (Deetz, 1992; Leonardi & Jackson, 2003; Markham, 2021) around both the risks and the potential of visual generative AI in a manner that supports the maintenance of the status quo. I also suggest that the press coverage of Lensa, which both articulates key AI-related harms and frames those harms as intractable and insolvable, creates a discourse of inevitability (Leonardi & Jackson, 2003; Markham, 2021) that has implications for how these issues are understood by the public, and for the approaches that are taken to address them. Method The study corpus consisted of 135 unique articles from the global English-language syndicated and tech press. Articles were sourced from Nexis using the term “Lensa AI” in English-language coverage across all regions from November 2022 to July 2023. Topically irrelevant results were excluded from the corpus, and any repeated articles were deduplicated. I then took a “snowball” approach (see Braithwaite, 2016) to the content linked in the Nexis results, adding any new articles. I then read these articles and followed those links until reaching saturation/failing to come across any new material. I analyzed the corpus using reflexive thematic analysis (Braun & Clarke, 2019; Byrne, 2020). Following Braithwaite (2016), my goal was to offer an “illustrative rather than exhaustive account” of the Lensa discourse and demonstrate its most salient features (p. 3). This study has obvious limitations in that the corpus is limited to English- language coverage; the discourse about Lensa likely has a wider global audience than the countries reflected in this analysis. Nonetheless, the discourse surrounding Lensa had distinct contours that were quite consistent across the countries represented in the corpus, offering a clear illustration of the key issues associated with visual generative AI at this point in time. Findings Out of the 12 countries represented in the corpus, there were four countries that comprised the majority of English-language Lensa coverage: the United States, Mexico, the United Kingdom, and India. The United States and Mexico dominated the coverage, with 34% and 31% of the corpus, respectively. News items from the United Kingdom comprised 14% of the sample, and Indian stories were 13%. Collectively, these four countries made up 92% of the corpus. The remaining 7 countries (Australia, Canada, Egypt, Pakistan, Singapore, Switzerland, and the UAE) collectively composed 8% of the corpus. There were two distinct types, or genres, of content in the Lensa discourse: scandal coverage and explainer coverage. Scandal coverage, which comprised 42% of the corpus, focused on perceived wrongdoing or norm violations related to the app and its use. The other type of coverage was explainer coverage, which constituted 58% of the corpus and focused on describing what Lensa is and how it worked. In addition to the two categories of coverage, there were three overarching themes across the sample. The first theme focused on predatory data practices. This theme was the most prevalent in the corpus, reflected in 56% of the articles and within six countries (USA, Mexico, UK, India, Australia, and Canada). It largely focused on the unethical or non-consensual use of data in AI technologies, discussed through two distinct, but related topics: the use of art as training data in the Stable Diffusion model and Lensa’s privacy policy, which stated at the time that users’ photos could be used to train the app. The second theme was related to biased content. Articles relating to this theme referred to the gender and racial biases reflected in some of the images generated by Lensa, including the sexualization of women, the lightening of darker skin tones, and the generation of explicit content. This was the second most prevalent theme in the corpus, reflected in 47% of articles, and was present in coverage in the US, UK, Canada, Australia, India, Mexico, and Switzerland. The third theme focused on user behavior. The content in this theme described how people used the app, including celebrities and political figures. It was reflected in 37% of articles within five countries (USA, Mexico, UK, India, and Australia). Discussion The key themes of “sex, art theft and privacy” (Biron, 2023) in the Lensa discourse connect to existing framings and understandings of sociotechnical concerns. The discussions of Lensa’s predatory data practices closely align with popular discourses about users as exploited data subjects that took off exponentially in the wake of the Cambridge Analytica scandal (e.g., Isaac and Hanna, 2018). Critiques of Lensa’s sexualized and racialized results map directly onto discourses about “algorithms of oppression” (Noble, 2017) and other forms of gendered and racialized algorithmic biases, and analyses of Lensa user behaviors were plucked directly from the selfie moral panic playbook, with their insistence that digital technologies amplify our most deleterious narcissistic tendencies (e.g., Tiidenberg, 2018). Repeating existing framings of these sociotechnical issues perpetuates the status quo. While the discourse about Magic Avatars identifies key AI-related problems, it also treats those problems as intractable obstacles that are too big to solve; it is taken for granted that these systems or models will inevitably cause harm. In this way, we can see discursive closure at work: by acquiescing to a discourse of inevitability (Leonardi and Jackson, 2003; Markham, 2021) about generative AI, the Lensa coverage offers “a particular view of reality that is maintained at the expense of equally plausible ones” (Deetz, 1992, p. 188). A range of ideas have been proposed by computer scientists, artists, and other stakeholders (e.g., Crabapple, 2022; Hundt et al., 2023) to address AI- related harms, but within the Lensa coverage, no such solutions are visible; instead, we are told that when it comes to generative AI, “the horse has already left the barn” (Hatmaker, 2022). Markham (2021) explains that the power of anticipatory logics that flow through everyday discourse around emerging technologies end up building and reinforcing “a hegemonic ideology of external power and control”, strengthening the dominant frames of inevitability and powerlessness (p. 384). As a result, more creative approaches to the challenges and potential harms of technologies like generative AI end up being foreclosed or dismissed as unlikely, constraining the boundaries of possibility and limiting a more expansive vision of what they could– or should– be. References Biron, B. (2023, January 16). Sex, art theft, and privacy: Lensa exploded overnight — and now the avatar app is dealing with public backlash. Here’s what you should know. Business Insider. https://www.businessinsider.com/lensa-ai-raises-serious- concerns-sexualization-art-theft-data-2023-1 Braithwaite, A. (2016). It’s About Ethics in Games Journalism? Gamergaters and Geek Masculinity. Social Media + Society, 2(4), 2056305116672484. https://doi.org/10.1177/2056305116672484 Braun, V., & Clarke, V. (2019). Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health, 11(4), 589–597. https://doi.org/10.1080/2159676X.2019.1628806 https://www.businessinsider.com/lensa-ai-raises-serious-concerns-sexualization-art-theft-data-2023-1 https://www.businessinsider.com/lensa-ai-raises-serious-concerns-sexualization-art-theft-data-2023-1 https://doi.org/10.1177/2056305116672484 https://doi.org/10.1177/2056305116672484 https://doi.org/10.1080/2159676X.2019.1628806 https://doi.org/10.1080/2159676X.2019.1628806 Byrne, D. (2022). A worked example of Braun and Clarke’s approach to reflexive thematic analysis. Quality & Quantity, 56(3), 1391–1412. https://doi.org/10.1007/s11135-021-01182-y Crabapple, M. (2022, December 21). Op-Ed: Beware a world where artists are replaced by robots. It’s starting now. Los Angeles Times. https://www.latimes.com/opinion/story/2022-12-21/artificial-intelligence-artists- stability-ai-digital-images Deetz, S. (1992). Democracy in an age of corporate colonization: Developments in communication and the politics of everyday life. SUNY press. Hatmaker, T. (2022, December 5). Lensa AI, the app making ‘magic avatars,’ raises red flags for artists. TechCrunch. https://techcrunch.com/2022/12/05/lensa-ai-app-store- magic-avatars-artists/ Leonardi, P. M., & Jackson, M. H. (2004). Technological determinism and discursive closure in organizational mergers. Journal of Organizational Change Management, 17(6), 615–631. https://doi.org/10.1108/09534810410564587 Markham, A. (2021). The limits of the imaginary: Challenges to intervening in future speculations of memory, data, and algorithms. New Media & Society, 23(2), 382– 405. https://doi.org/10.1177/1461444820929322 McCluskey, M. (2022, December 14). Why It’s So Hard to Resist Turning Your Selfies Into Lensa AI Art. Time. https://time.com/6240648/lensa-ai-psychology-behind/ Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press. Tiidenberg, K. (2018). Selfies: Why We Love (and Hate) Them. Emerald Publishing. https://doi.org/10.1007/s11135-021-01182-y https://doi.org/10.1007/s11135-021-01182-y https://www.latimes.com/opinion/story/2022-12-21/artificial-intelligence-artists-stability-ai-digital-images https://www.latimes.com/opinion/story/2022-12-21/artificial-intelligence-artists-stability-ai-digital-images https://www.latimes.com/opinion/story/2022-12-21/artificial-intelligence-artists-stability-ai-digital-images https://techcrunch.com/2022/12/05/lensa-ai-app-store-magic-avatars-artists/ https://techcrunch.com/2022/12/05/lensa-ai-app-store-magic-avatars-artists/ https://doi.org/10.1108/09534810410564587 https://doi.org/10.1177/1461444820929322 https://time.com/6240648/lensa-ai-psychology-behind/