Tool,companion or a catalyst force? Exploring sociotechnical imaginaries Within AI livestreams’ communities of practice

Researching AI imaginaries

Public discourses on technological developments display ‘a rhetoric of prospective potentials that innovation sets free. This rhetoric not only enduringly frames the perception of businesses and customers for a technology but also creates an element of performativity’ (Bareis and Katzenbach, 2022: 860). Current AI discussions illustrate this point and several studies show that media discourses have become ‘sensationalized, industry-driven, and politicized’ (Brennen et al., 2018; Goode, 2018), presenting an over-hyped vision of AI by focusing on its potential and capabilities (Elish and Boyd, 2018).

Previous research observes that news media and governments engage in AI discourses that promote narratives framing the technology primarily in respect to economic benefits, technological progress, and geopolitical competition (Bareis and Katzenbach, 2022; Nguyen and Hekman, 2022a; Nguyen and Hekman, 2024; Zeng et al., 2022). Nguyen and Hekman (2022a) and Nguyen and Hekman (2024) identify four major meta-frames focusing on economic, cultural, social, and political impacts of AI. Similarly, Scott Hansen (2022) finds several prevalent themes in AI news: AI as an autonomous entity that will develop past humans, machines and humans working complementary, AI taking over humanity, humans being passive and excluded from AI development, and AI making positive contributions to humans and serving human development (2002: 65‒69). Zeng et al. (2022) highlight the transmission of dominant frames in the context of social networks and mainstream media in the Chinese AI discourse, showing that narratives and imaginaries disseminated by the Chinese government remain dominant without relevant ‘counterpublics’ challenging these ideas. Concerning national policy, Bareis and Katzenbach (2022) explore imaginaries in national AI strategies of the USA, China, and Germany. They observe that imaginaries dominating governments’ views promote the idea that AI is essential for national competitiveness and national security (868). They portray AI as inevitable (869) and that it could serve as a technological solution for social problems and challenges, ultimately enhancing the nation's overall well-being (869).

Previous studies reveal the considerable influence of elite narratives dominated by governments, media, and tech businesses, alongside a lack of critical and alternative viewpoints, in shaping public discourses on AI. Discourses do not only convey how stakeholders view a technology but highlight the salient aspects that interest them. Imaginaries, Vicente and Dias-Trindade (2021) claim, ‘reside in the reservoir of norms and discourses, metaphors and cultural meanings’ (710). Importantly, previous research suggests that AI is undergoing a process of mediatisation. Mediatisation, as a ‘metaprocess’ (Krotz, 2009; as cited in Hepp, 2016), describes how societies evolve when ‘everyday practices increasingly rely upon media and become “moulded” by them’ (Hepp, 2016: 919). Different stakeholders mobilize their visions to shape future developments, regulations, and uses, operating in what Jasanoff and Kim (2009) describe as the ‘understudied regions between imagination and action, between discourse and decision, and between inchoate public opinion and instrumental state policy’ (123).

However, a limitation of previous studies is the focus on mainstream discourses. Expanding the analytical angle is crucial to understand how the sociotechnical imaginary landscape is configured and to recognize the role of alternative voices. This does not only mean that more research is needed on discourse cultures outside of the ’West’ by, for example, shifting focus to East-Asian or African AI discussions. There are manifold discourse formations defying geopolitical categorisation in online communities where people with varied backgrounds engage in the formation of sociotechnical imaginaries through daily practice. Analyzing alternative discourses can reveal the complex interplay between mainstream imaginaries in the broader public sphere and community-driven imaginaries among non-elite participants who share an interest in AI.

Alternative imaginaries among CoPs

In contesting elite-centric analyses of media culture, Kellner (2020) points to the relevance of smaller communities and individuals:

[e]veryone can contribute to discussions, critique, and media creation through social media and new technologies and platforms (…) even create their own narratives, images, analyses and media artifacts as the tools of media production become part of the digital devices of everyday life. (2‒3)

This challenges critical research to engage with highly dynamic and diverse discursive formations in the digital sphere. Similarly, Bruns (2023) criticizes the idea of a unified or dominant ‘digital public sphere,’ suggesting that it is more accurate to think of a ‘fractured digital sphere.’ This fragmented landscape reflects the diverse nature of multiple social groups and online cultures existing in parallel. Accordingly, Vicente and Dias-Trindade (2021) argue that sociotechnical imaginaries have come to ‘incorporate greater intellectual plasticity through contributions that highlight a contested nature of sociotechnical imaginaries and the crucial importance of studying the (in)visibility of the more localized origins and circulation of alternative imaginaries’ (711).

CoPs (Wenger, 1999) that gather on online platforms around shared interests are one important example for these alternative discursive formations. What makes them insightful cases for studying AI imaginaries outside of mainstream narratives is their direct engagement with technology for specific goals. CoPs are often ‘pioneer communities’ (Hepp, 2016), exhibiting self-selecting and territorial traits. Examples for CoPs include collaborative groups that come together based on a shared interest in the interplay between society and the Internet of Things (IoT), self-measuring technologies and datafication, and open-knowledge movements (Hepp, 2016: 921‒923). These communities are media-related, that is, ‘they are constituted by technical means of communication’ (Hepp, 2016), utilizing the networking and creative potentials that digital media afford them. Often, CoPs play a ‘forerunner role’ in the adoption of and experimentation with emerging technologies, including AI (924). Communities around AI-generated livestreams are an illustrative example for CoPs exploring the capabilities of GenAI for specific creative purposes. These communities collaborate not only in creating content in line with their sub-cultural interests; they also socialize around shared norms, develop common vocabularies, negotiate and determine values, and curate AI imaginaries (Hepp, 2024). AI livestream communities can be considered ‘cultural intermediaries’ (924), bringing together members with diverse backgrounds, such as developers, content creators, journalists, and viewers. These communities are likely dominated by a similar demographic ‒ well-educated white men ‒ common in other online emerging technology spaces, which is important to consider when evaluating the diversity of viewpoints.

Importantly, CoPS discourses are considered ‘alternative’ vis-a-vis mainstream discourses because they take place among users that do not necessarily have a broad public audience in mind and may not apply strategic agenda-setting and framing in their communication; the assumption is that they are sharing genuine viewpoints among more or less equal discourse participants. Their views on AI can align with mainstream narratives but may also diverge in crucial respects. What makes them alternative is not that they per se represent marginalised positions in technology discourses but the context in which their discourse is performed.

Based on the discussion above, the present study investigates AI discourses within the Discord community of Nothing, Forever. It is the first AI livestream ever created to implement an interactive and dynamic system that gives users partial control over its narrative, which contributed to its popularity. Aside from the streaming platform (Figure 2), most community interactions take place on the external communication platform Discord, where viewers participate in discussions on a wide range of topics related to the stream, technology, and media culture. Users discuss technical aspects of the generated content, including the evolution, current application, and future of AI.

OPEN IN VIEWER

Figure 2.

‘Nothing, forever’ being transmitted through the streaming platform Twitch.

The analysis is guided by the research question (RQ) How do communities of practice construct, construe, and shape the sociotechnical imaginary about AI, particularly in the Discord community surrounding the AI-generated livestream ‘Nothing, Forever’? The empirical part is further guided by two specific sub-questions (SQs):

What is the dominant AI imaginary present and circulated in the Discord community?

The present study aims to understand the influence and dynamics of CoPs in the development and dissemination of AI sociotechnical imaginaries as well as how alternative discursive formations relate to mainstream AI discourses.

Data retrieval and pre-processing

Via the open-source tool ‘Discord Chat Exporter’ (Tyrrrz, 2023), all chat messages and meta information were retrieved from ‘AI-discussion,’ resulting in a dataset of 7,934 individual texts posted by 847 unique users. The dataset spans from February 2023 to December 2023, which covers all available data at the time of extraction. Different Natural Language Processing (NLP) methods in Python 3 were used for the computational-quantitative analyses. First, the texts underwent several pre-processing steps, including removal of HTML tags, URLs, lowercasing, expanding contracted verbs or abbreviations, and removal of emojis, digits, punctuation, and special characters, as well as lemmatization and tokenization using SpaCY (Honnibal et al., 2023). Next, part-of-speech tagging was applied to distinguish nouns from adjectives and verbs. This was necessary for more efficient text analysis where only one word type was considered relevant (e.g., topics based on nouns).

Automated text clustering and network analysis

To identify frames, texts were clustered with term frequency-inverse document frequency (TF-IDF) vectorization. Methodologically, frames are considered clusters of texts that share similar words. For example, a cluster of documents frequently sharing words such as ‘risk,’ ‘ethics,’ ‘privacy,’ and ‘bias’ could be considered as the ‘AI Risks’ frame (Nguyen, 2023). The TF-IDF algorithm calculates the importance of a word in a document by measuring its frequency in that document and normalizing it based on how frequently the term appears across all documents. Those sharing distinct words are then grouped together via clustering such as k-means. A key challenge is to determine optimal k, that is, the number of text clusters and thus frames expected to emerge from the data. There are different quantitative measures to approach a range of values for k that ‘best’ represent the diversity of clusters in a given dataset. To determine the optimal number of clusters, Silhouette Scores (Rousseeuw, 1987) were calculated, before applying the k-means++ clustering algorithm (Ostrovsky et al., 2012). Generally, the silhouette scores were low as tested for a range of k from 5 to 20 (0.071‒0.084), indicating considerable overlap and homogeneity in the discourse. Eighteen clusters emerged as the statistical optimum (0.084) and was retained since it revealed niche discussions that would have been otherwise collapsed in a lower cluster solution. The 18 clusters representing frames were manually reduced to four broader socio-technical imaginaries based on thematic similarities.

Next, a similar TF-IDF adjacency matrix was created to visualize word co-occurrences in Gephi. This matrix contained all nouns in a bi-gram disposition (e.g., ‘AI creativity’). The matrix was loaded into the Gephi visualization software and then spatialized via the Fruchterman Reingold layout algorithm. Modularity was employed to identify and color-code communities within the network by assessing the concentration of edges relative to a random distribution of links across nodes. Communities exhibiting a high degree of similarity were manually merged. The resulting clusters were then compared with the TF-IDF clusters, facilitating validation of the clustering and enabling a visual exploration of the relationships between frames.

Close readings

Close reading is crucial to enrich computational analysis (Franzini et al, 2015). To this end, at least 30 texts or more -depending on the cluster size- were manually analysed. This exploration of 600 full texts supplemented the labeling of frames next to the bag-of-words representation from the computational analysis. Simultaneously, it helped with further analysis of framing practices and sentiment towards AI. It is important to mention that the algorithmic text clustering does not account for the context in which words are frequently mentioned; the results merely provide orientation for the researchers to discern what issues and topics dominated the discussion and can be interpreted as frames. With the help of close readings, it becomes possible to characterize and reflect on each of the frames identified with the computational distant reading.

It is important to note that, while very insightful for an exploration of AI discourses and imaginaries within a specific CoP, the case of Nothing, Forever comes with several empirical limitations. First off, it concerns a special interest community that is likely to consist of like-minded individuals with a particular affinity for technology. While the CoP is global in scope, it is by no means representative of a more diverse broader public but largely consists of individuals with arguably advanced knowledge especially about technical aspects of AI compared to other lay audiences. Secondly, it is only one example of a vast diversity of CoPs that engage with AI for various creative tasks (e.g., image art, audio, etc.).

Distant Reading: A tech-enthusiastic imaginary of AI as a tool for creative processes

A single socio-technical imaginary presenting AI from an instrumentalist and mostly tech- enthusiastic viewpoint dominates the community discord. This becomes apparent when closer investigating the different themes that emerged from clustering the individual frames. The broader AI imaginary is nuanced within a tool-centric interpretative horizon but connects only to a limited extent to cultural and social questions.

The text clustering and manual inspection yielded 18 frames (Appendix 1), most of which centre on content creation and scripting. 77.4% of all user messages evoke the frame Interactive Narratives/Scripting about content-related decisions, followed with considerable distance by Collaborative Art (3.2%), Script Feedback/Meta-Talk (3.0%), and GenAI Training (2.8%). Fewer discussions focus on reflections about Scriptwriting, AI-Generated Art, AI Livestream, and Technical Prompting (each accounting from 1 to 2.6% of all messages), with limited messages discussing AI for Media Production more broadly and Sensory and Expression (<1%).

The different frames were further grouped based on thematic similarities (Appendix 2), underscoring that most of the community discourse construes and expands a sociotechnical imaginary that mostly portrays AI as a Tool for Creative Practices (92.4% of all messages). AI Discussions among users are primarily centred on the purpose at hand, that is, deploying the technology to create content and steer the narrative of Nothing, Forever in a direction that resonates with the communities’ expectations. This entails critical discussions about how to achieve these objectives. Relatedly but to a much lesser extent, users focus on AI's Technical Foundations (4.3%), the implications of Human-AI Interaction on a more abstract level (1.7%) and societal issues under Ethics, Labor and Social Impact (1.6%).

The strong focus on AI as a tool with creative potentials and an overall more technical perspective among users is further supported by semantic network visualizations of frequently co-occurring words and bigrams. This approach allowed for exploring framings that are commonly associated with one another, as well as the identification of issues that were not captured by the text clustering method.

The general network graph (Figure 3) indicates a strong presence of four key areas of debate. Two of these areas focus on technical discussions related to AI, such as machine learning, large language models (LLMs), and training data, as well as the use of AI in livestream production, including voice cloning, the video game engine used in the stream, and scriptwriting (Figure 4(a)). The other two domains are centered around discussions on content generation, streaming, scriptwriting, episode commentaries, and debates on art, AI implementation, and work (Figure 4(b)).

OPEN IN VIEWER

Figure 3.

Ai discussion network visualized in Gephi. Every node in the network represents word associations (bigrams and trigrams). Each cluster of words represents relevant discussions for the channel.

OPEN IN VIEWER

Figure 4.

Relevant domains of discourse. (a) (Left) Represents AI technical and AI-related discussions about software, characteristics and domain-specific knowledge. (b) (Right) Portrays various discussions about art, content generation, streaming, and scripts.

Altogether, these four domains of discussion support the findings from the text clustering and thematic grouping of frames. First, Figure 4(a) reveals the presence of the AI as a Tool for Creative Practices and Technical Foundations themes, where domain-specific discussions focus on understanding AI's inner workings, as well as its implementation in content generation. Second, Figure 4(b) points to Human-AI Interaction, featuring conversations about the content production process, content-related topics (e.g., improving AI's scriptwriting, making it more humorous), and discussions centered on streaming, art authenticity, and work. The fourth theme, Ethics, Labor and Social Impact, although much less represented in the discourse, emerges from the interrelation between these domains. Here, AI implementation intersects with scriptwriting, image creation sparks art debates, and discussions about streaming and daily routines are intertwined with efforts to understand this technology.

Zooming-in on specific nodes allows for further exploration on the impact and implementation of AI in everyday practices, supporting the idea of AI mediatisation. For example, the node ‘chatgpt write’ (Figure 5(a)) seems mostly connected to technical aspects of AI. However, words indicative of ethical issues and challenges such as privacy, biases, authorship, copyright, or creativity are noticeably absent. The partial network around the node ‘game engine’ (Figure 5(b)) implies that users perceive AI in the context of content production and the role of different trending technological developments (e.g., stable diffusion).

OPEN IN VIEWER

Figure 5.

Word associations selected in Gephi. The left figure (a) shows every node connected to the bigram ‘Chatgpt write’. The right figure (b) shows relevant nodes connected to the bigram ‘Game Engine’.

Another prominent node based on frequency of occurrence and centrality is ‘ai generate’ (Figure (6)). It appears primarily associated with terms about human-ai relationships, such as ‘people ai,’ ‘procedurally generate,’ ‘uncanny valley,’ ‘create ai,’ ‘generate model,’ ‘imagine ai,’ ‘love ai,’ or ‘art ai.’ These discussions demonstrate an understanding of the processes involved in AI-generated content but also potential misuse of AI in video, such as in ‘deep fakes.’ This is one of the few instances where a potential risk of AI is addressed (albeit not necessarily from a critical angle).

OPEN IN VIEWER

Figure 6.

Word association selected in Gephi. Every node connected to the bigram ‘Chatgpt write’.

Accordingly, inspecting the network graphs prompts exploration of what issues and topics might be missing. For instance, there is a noticeable scarcity of word associations related to governance (e.g., regulation, governance, bill, law, policy), which suggests that the community may show limited interest in these areas.

In what follows, we further explore the four themes shaping the sociotechnical imaginary qualitatively in two subsections, based on their shared emphases: primarily on technological aspects, on the one hand, and on user-centric as well as social implications, on the other.

Close Reading of themes

Most discussions centre on AI's capabilities for audiovisual content creation and frequently link to technical questions, both concerning current possibilities and limitations but also with an eye on future potentials. In the latter respect, the community discourse reflects techno-optimist notions prevalent in mainstream discourses. Given the context of the Discord CoP, this is little surprising as it arguably attracts users with an interest in technology, including individuals with a hands-on professional background in tech-related sectors. Indeed, the CoP is apparently considered a pool for recruiting AI expertise (as illustrated in one post: ‘Hi all - any devs with openai experience that have bandwidth and are looking for work? reply here and ill msg ya - cheers!’). Generally, The boundaries between the four larger themes are often fuzzy, as discussions about steering the narrative can quickly pivot to questions about the underlying technology and vice-versa.

AI as a tool for creative practices and technical foundations

The largest theme AI as a Tool for Creative Practices consists of frames with a focus on hands-on content creation with the help of GenAI and steering the narrative of the show. There is a constant shift back and forth between these dimensions, where the context of interaction with the AI and among users as well as the affordances of the technology pre-determine the points of attention in user exchanges. For example, some users post prompts, while other discuss plot devices, the context of the show, and the implementation of creative ideas:

Runs in peanut butter (note: prompt for the AI)

Creative goals and technical factors lead here to the emergence of a techno-artistic discourse where AI capabilities and limitations take centre stage. This includes meta-discussions about prompting, where users reflect on and debate strategies for writing effective prompts/prompting practices and link it to current technological capabilities:

Without reading further, this sounds like someone “discovered” prompting. You can get ChatGPT to do the same thing if you tell it to respond as a given celebrity. The increase is nice though

There is generally a strong focus on various technical issues, both conceptually about AI more generally and specifically for concrete application purposes:

We've basically modeled neural networks after the most basic algorithmic function of neurons, to be trained to perform a task with a fuzzy idea of “correct”.

While this is a dominant characteristic of the general discourse, the technical outlook is particularly pronounced in the theme Technical Foundations:

[…]you will need to do modeling through Pixiv's V-roid, export to general 3D file then import to blender/unreal/unity to be able to do high resolution cross platform video output and sophisticated music videos[…]

Such comments indicate that certain sub-discussions require an advanced level of knowledge about AI and its underlying principles, which is characteristic for tech-centric pioneer communities. This includes critical voices that seem to bemoan hype-centric jargon in AI discussions. For example, some users ask for ‘stop using big words to say that ai is a tool’ and argue that ‘ai is a tool and that is all.’ This extends to emphasizing the importance of human creativity but also technical expertise and know-how in using AI:

I think one of the key things that people new to this tech need to understand is that it needs hella human supervision and design. chiefly, you need to understand procedural generation. And if you want to make something that really makes an impact, an understanding of how people […] interpret narratives.

Importantly, AI is positioned towards achieving very specific and simple goals. While useful and innovative, perceived limitations of the technology are pointed out as well:

AI is best used as a toy. I am going to write funny stories or make funny audio or generate funny images like basically just for entertainment or shitpost purposes […] It's not very good at anything else yet.

Despite this generally more sober outlook and a focus on concrete application, there is a line of reasoning that places emphasis on optimisation and eventual perfection of AI via technological progress, more data, less regulation, and sufficient funding:

After all, a model can only be as good as the data it's trained on (I think).

Relatedly, others echo narratives about the inevitability of super-human AI and motifs of the so-called singularity as promoted by influential futurists in popular business-oriented tech discourses:

Once AI supersedes the intelligence of all human beings combined, access to the AI superintelligence will become the most valuable resource.

Taken together, the exploration of the two closely intertwined dominant themes surfaces an AI imaginary that has a strong utilitarian outlook, appears overall enthusiastic about the technology, and positions it to some extent as an ‘insider topic,’ demanding sufficient expertise to understand it.

Human–AI interaction and ethics, labor and societal impact

The theme Human–AI Interaction connects creative, technical, and futuristic notions to the relationship between users and technology, partially evoking quasi-anthropomorphic visions of AI with emphasis on dialogue, conversation, and assistance. It further underscores a mostly tech-enthusiastic view. AI is portrayed not only as a technical tool enhancing the creative process but as a collaborator to engage with through communicative interaction. This implies a synergetic connection between users and AI technology, where the latter is assigned some agency in scriptwriting, directing shows, prompting, scripting, designing, generating scenarios, etc. Some users reflect here on AI as an anthropomorphized companion and potential future developments of technology beyond the context of the CoP:

Us KNOWING that the ai is the main character is what makes it entertaining because we can cheer it on, laugh at its problems, celebrate it when it gives us output that makes sense to us, etc.

More explicitly, there is an attribution of human characteristics to AI in phrases such as ‘a computer's mind’—drawing from a common metaphor of AI as a brain- and ‘let it be itself and letting it do what it wants,’ as if it had a mind of its own. However, other users seem to point to a dependency of AI on human direction and argue that it cannot understand in a sentient sense:

To be honest, the AI isn't sentient enough to get “Canceled”. It's just running lines of code that make it speak. It probably made an oopsie because it read a random line of code off and without the usual moderation in place it spit it out. The AI doesn't know what offensive means.

Overall, users tend to recognize AI as a valuable resource for generating creative output. However, there are some who seem to have an ambiguous view on AI's capabilities and potential future development, pointing to (incremental progress) but also broader implications on its perceived value:

If the AI got too realistic or started making too much sense i feel like it'd lose a lot of the charm it has currently

A smaller number of comments link community interests and AI practices to broader societal and ethical issues, albeit limitedly. These form the theme Ethics, Labor and Societal Impact, capturing more critical views of AI's cultural and social implications. AI is seen as a transformative agent, with a shift toward mediatization, where technology integrates into daily tasks and shapes practices across domains. Discussions reflecting this tend to downplay ethical concerns and ignore critical perspectives, while emphasizing AI's potential, sometimes exaggerating its capabilities. Recurring narratives include AI and art, job replacement, AI dominance, and the future of employment, often framed with techno-determinist notions:

We're not even sure what human intelligence and consciousness is in discrete terms or whether it's strictly superior to AI.

Yet, at the same time, some users appear cautious about these developments and point to technological hurdles that need to be overcome first to unlock AI's full potential, especially regarding memory as both a computational resources and an ‘ability’:

The biggest thing is that we need AI to have longer and more robust memories. And to understand how to sort what's important to remember or not itself so it can optimize that process without a human guiding it.

Others locate challenges on a social dimension by pointing to the ways tech businesses work and how that affects the creative sectors but also see concrete empowering potential for wider user base:

The problem is the economic system itself, not whether the artist is assisted by AI systems.

Despite a shared enthusiasm for the technology, not all users are fond of the corporatization of these technologies. This resonates with ideals in the broader open source movement, where one main principle is ‘peer production,’ meaning that users support the idea of making source codes, blueprints, and documentation freely accessible to anyone:

I mean, a while back, Google wrote a big long thing about not wanting to grant open access to their AI art bot because of ethical concerns. Now they're out in the wild and we're all like “oops, jobs gone lol”.

Importantly, several frames associated with this imaginary focus on technical aspects of AI, such as data training, and model development. They rarely address issues regarding developer ethics and critical perspectives on AI adoption and misuse. Similarly, issues such as copyrights, environmental impact, monetization, and platform economies, are discussed to a lesser extent. Although discussions about AI regulation are rare, some community members do address the issue, indicating that AI will be regulated in a preventive rather than reactive manner:

All the laws and regulations will be preventative, addressing terrible things that have already happened well after they did.

Generally, AI is seen as ubiquitous, underlining its growing role and its expanding applications across various domains (e.g., content creation, production, automation, entertainment, and scriptwriting). It is noticeable, however, that the use of GenAI triggers a critical discussion about art and what its use implicates for the future of the production of creative content:

Expect a lot of messy legal battles over AI generated content in the future. the ruling that AI-generated art is not subject to copyright will be challenged multiple times. I hope it holds up, but we'll see how things go.

There is a reason why nobody is hiring people doing AI art. They all look the same.