Augmenting Human Creativity with AI

Artificial Intelligence is often perceived as neutral and objective, masking biases, hidden labour, and infrastructural power. Transparency has become a pressing concern in generative AI, as these systems increasingly influence how knowledge, culture, and decisions are produced. This provocation presents The Models, an interactive installation by the artist duo dmstfctn, that stages large language models as improvised performers within the tradition of Italian Commedia dell’Arte. By embodying tendencies such as hallucination, antagonism, and servility, the work invites audiences to both believe in and question the machine. Building on this case, the talk introduces the notion of a smark stance: an oscillation between enchantment and critique enabled by a temporary suspension of disbelief. I argue that this ambivalent engagement allows generative AI to be experienced as neither infallible nor merely instrumental, opening a space where errors become legible and meaningful. Artistic practices like The Models suggest that fairness in AI can be cultivated through experiences that expose, rather than conceal, the systems’ limits.

AI has made significant inroads into the creative and cultural sectors, both as a tool or medium and as a socio-political object. In this presentation, my aim is to highlight the friction that exists between technological innovation in AI and artistic practice. To do this, I review current knowledge on AI practices (pre-genAI and genAI) in artistic practices. I show that the latest generations of AI have diminished artists’ agency over these technologies, both from a material and a cultural and political point of view. I then propose a framework for understanding the power relations between innovation and creative practices in three dimensions: cultural, structural and ethical. This framework is a way of encouraging more research into these socio-technical issues, beyond the opportunities and challenges that AI poses for the arts sectors.

(Artistic Provocation)

This artistic research demo presented an ongoing project that explores sound generation through a small, carefully curated dataset composed of socio-political sonic material. Countering the tendency to optimize for scale and efficiency that dominate contemporary machine learning practices, the project employs a small-scale sound corpus to train a generative sound model capable of producing unexpected and speculative sonic outputs. The results thus serve as an inquiry into how machine learning systems mediate complex situated lived realities. As such, the project foregrounds the ethics and aesthetics of working with data through embodied interaction with relevant material to construct a dataset, situating model training as a relational and iterative process unlike result-driven paradigms. In contrast to large-scale models reliant on massive amounts of compute and extractivist data logics, this work investigates how slow and small datasets and attentive training practices can open alternative modes of human-machine collaboration. By treating sound generation as a site for speculative experimentation beyond accurate representation, the project contributes to ongoing debates in artistic research and critical AI about the politics of data as knowledge, the materiality of training processes, and the possibilities of alternative machine learning practices.

The purpose of my presentation was to focus on the marginalising aspects of technology adaptation as well as ways of using GenAI to suggest different kinds of histories, realities, and desires. I presented a recent essay that I co-authored [2] suggesting that applying a 10-year perspective makes it possible for us to see aspects of a technology, for instance what it marginalises, that are in many ways difficult to see on a day-to-day basis. In the essay, we use the example of how IoT technologies enables harm-doing in domestic abuse situations to ask how we might as researchers be better at understanding the potential harms that any technology (here: generative AI) will be able to lead to. I then proceeded to presenting two artists who use GenAI in their practice. First was the Swedish artist Arvida Byström, who in the piece “In the Clouds” (2024) worked with apps that can visually strip naked any human on a photograph. With this, she interestingly explores how the algorithms in the app “dream-up” impossible bodies, both expanding the idea of what a body might look like and explosing how the algorithm has a very narrow idea of the female body. I recently co-authored an article (in Danish) on this project [1].

Second artist is Danish artist Kristoffer Ørum, who uses GenAI to imagine what the cities and countries he works with could look like, had our social ideologies been different. Specifically, he asks what the city of Aarhus (Denmark) would look like if the positive energies and ideologies of egality and creativity that we also find in hip-hop culture had been the dominant political organising principle. In total, the aim of my talk was to use my recent work/research to spark conversations on how we might think about which worlds are being built with GenAI.

Group 1 examined how AI systems can meaningfully engage with the situated, idiosyncratic, and often non-linear nature of creative practices. Participants identified a central limitation of contemporary generative AI tools: their tendency to promote design fixation, homogenization, and “echo chambers” through simplified input–output interaction models. Such paradigms obscure the rich, iterative, and exploratory processes that characterise creative work. In response, the group explored how AI systems might be designed to surface and support the in-between, i.e., the intermediate states, transformations, and uncertainties that occur between prompt and output. Emphasis was placed on interaction paradigms that invite tinkering, transparency, and serendipitous discovery, enabling creators to navigate, manipulate, and reflect on generative processes rather than merely consume results. The group proposed a multi-stage research agenda. First, ethnographic studies with creative practitioners would document existing workflows, values, and frictions in AI-assisted creation. These insights would inform a shared knowledge base mapping current AI tools, their affordances, and, critically, what is missing. Building on this foundation, the group envisioned speculative and experimental design explorations that challenge linear interaction models, for example, through visualizing generative processes, enabling non-linear intervention, or introducing productive friction. The overarching goal is to reconceptualise AI systems not as automated generators, but as creative partners that preserve authorship, support exploration, and amplify the uniqueness of individual creative voices.

Group 2 explored how AI systems might better support “humaning” by expanding interaction beyond text-based interfaces toward multimodal, embodied, and relational forms of engagement. Rather than treating multimodality as a purely technical extension of existing interfaces, the group focused on what the generative and fuzzy qualities of contemporary AI introduce as fundamentally new interaction possibilities. Central to this exploration was a shift from efficiency-driven conversational models toward slower, more reflexive, and contemplative modes of interaction, including tactile, contextual, and sensual interfaces that engage the whole body.

The group also investigated “black box” strategies not as mechanisms for transparency, but as sites for intimacy, viscosity, and negotiated understanding, asking what it might mean to relate to AI without fully opening it or optimizing it. They discussed hegemonic narratives embedded in conversational AI, proposing alternatives such as conversational consent, bidirectional negotiation, and embodied manifestations of AI from products to robotics. They also foregrounded the importance of limits, articulating the concept of “un-AI”: negative spaces, black holes, and contexts where AI should deliberately not exist.

Group 3 addressed the fundamental challenge of evaluating creativity in human–AI collaboration, arguing that prevailing evaluation approaches are ill-suited to creative practice. Current methods are often derived from machine learning benchmarks or productivity-oriented software metrics, and hence tend to reduce creativity to binary distinctions (e.g., good vs. bad, human vs. machine) or artifact-centric measures that neglect process, experience, and meaning. The group proposed a shift toward holistic, process-oriented evaluation frameworks that reflect creativity as situated, contested, and value-laden. Rather than privileging output quality alone, evaluation should account for dimensions such as expressivity, authorship, control, ownership, novelty, authenticity, trust, and the productive role of randomness and imperfection. Participants emphasised that creativity often emerges through struggle, iteration, and discomfort, which are qualities that are often obscured by efficiency-driven metrics. To operationalise this perspective, the group outlined several directions for future work. These include developing functional applications that embed alternative evaluation logics, exploring participatory and narrative-based assessment methods, and supporting community-driven governance frameworks for evaluating creative AI systems. At a deeper level, the group called for a craft-based epistemology of AI, one that recognises tacit knowledge, embodied practice, and experiential judgment as legitimate forms of evaluation. Ultimately, the group aims to establish new standards for assessing creativity that resist reductive quantification and instead honour the complexity of creative work in AI-mediated contexts.

Group 4 explored the question of attribution as an important concept in contemporary creative practice delineating reception and authorship of creative products in the context of the use of Generative AI. New technologies often challenge traditional approaches to notions of authorship in creative production. As Benjamin argued, new technologies can transform what constitutes a creative product, challenging traditional notions of authenticity, uniqueness and value of creative work [1].The emergence of generative AI systems has introduced another such challenge to our notions of creativity and authorship. In 2022, James M. Allen created a visual art piece entitled Théâtre D’opéra Spatial using Midjourney, which won a fine arts competition with a $300 cash prize [4]. When the artist submitted a request to the United States Copyright Office for registration of the image, the request was refused because “the Work contains more than a minimis amount of AI-generated content” which can not be separated from human contribution and thus could not be said to be the work of “human authorship” [6]. As generative AI becomes ever more integrated into everyday creative practices, from essay writing to image creation, the question of attribution of authorship becomes ever more important across domains.

What does authorship look like when agentic AI systems are involved in co-production of content? Group and collective authorship of creative content has a been a point of debate and concern for decades. Authorship and attribution are a type of contested perfomativity, a form of cultural production and form the basis for valuations of creative output [3]. What happens when we begin to think in terms of LLMs though – who is then speaking through the LLMs as these systems ingest human-generated content and ideas and then reproduce, remix, reformulate them in response to prompts? Who is included and who is excluded a culturally contested space of creativity? The practice of naming and attribution of the creative product can privilege the loudest voices and has repeatedly motivated explorations of alternatives to traditional naming and authorship expressions [2]. With the use of generative AI, how do we mark and signal authorship and what cultural changes might come about as “with the use of AI” becomes a common signal?

Past work at IBM on attribution for working with AI systems has explored attribution as a form of ownership or acknowledgement of contributions of different types made to a creative work. Acknowledging that different communities have different attribution standards (the credit taxonomy), He, Houde and Wiesz [5] developed the IBM AI attribution toolkit (https://aiattribution.github.io/), providing a language to describe how AI was used to create a work. Clearly the new vocabulary is needed to signal and debate emerging notions of (co)authorship, ownership, credit, provenance when creative work is produced using generative AI systems. The questions become ever more complex however, as the proliferation of these systems across creative digital tools can result in creative practice that may be unaware of the use of Generative AI.

The group discussed many aspects of attribution, coming up with a loose taxonomy of what might constitute appropriate tracking and provenance marking:

• $\mathrm{■}$

  provenance, authorship, ownership, attribution

• $\mathrm{■}$

  origin or location – effort that goes into creation of output

• $\mathrm{■}$

  tracking from output to original – tracing back to original sources

The resulting discussion noted that in the end we arrive at impossible tasks of provenance tracking, such as: what would it take to trace the origin for every text? Creative content production is ubiquitous, but where are the limits to attribution? Does everything need to be attributed or only things generated with particular types of generative AI systems? Is the question of attribution rooted in the use of prompts or should we consider all generative AI tool use? Current debates propose technical marking of AI generated content through the C2PA provenance marking standards (https://c2pa.org/) or watermarking and other computationally more intensive approaches. Yet the question remains on what ought to be marked, when and why. What effects visibly or invisibly marking AI-generated or AI-assisted content will have on cultural production, creative pursuits and what comes to constitute authenticity for the work of art in the age of Generative AI. The group identified productive questions for research that is crucial for both tracking the normative evolution of how people will interpret AI-produced or AI-augmented content and, potentially, developing mechanisms for affecting this normative evolution by providing new mechanisms for marking and reflection.

Creativity and improvisation go hand in hand. How better to think about creativity than through improvisational practice? This performance brought together four members of the Dagstuhl Seminar with their own individual improvising practices in an ad-hoc multimedia exploration of inter-agencies between humans and technologies. Dave Murray-Rust uses the laptop as an instrument, delegating micro-decisions to algorithmic processes while the musical shaping stays with fingertip controls. Philippe Pasquier marshals collections of musical agents to work with and against musical impulses, inhabiting a space between conducting and performing, and layering drones using the Kenaxis live software. Courtney Reed starts with the expressive range and power of operatic voice, then dives into a sea of electronic alterations and additions. Joe Lindley works visually, using realtime input to shape a stream of computational imagery through increasingly abstracted yet coherent dreamlike terrains. Performing for the first time, without rehearsal, they created the piece here as an example of lived practice within the discursive frame of the seminar.