Experiential AI: A Transdisciplinary Approach

• Experiential AI is a transdisciplinary paradigm that transforms opaque algorithmic processes into tangible, co-creative experiences using art, design, and technology.
• It integrates methodologies like immersive installations, creative residencies, and iterative workflows to reveal hidden data flows and enable human-machine interaction.
• The paradigm prioritizes transparency, accountability, and public engagement by involving users in the active co-design and experiential examination of AI systems.

Experiential AI denotes a transdisciplinary research and design paradigm in which the internal mechanisms, agency, and consequences of artificial intelligence systems are rendered materially, cognitively, and affectively apprehensible through situated, often co-creative, experiences. Originating in the intersection of art, science, design, and engineering, Experiential AI aims to bridge the epistemic gulf between opaque algorithmic processes and human understanding, recasting explanation from a matter of textual or post hoc commentary to one of multisensory interaction, narrative embodiment, and democratized agency. Core to this paradigm is the conviction that explanation, legibility, and accountability in AI systems can only be meaningfully achieved when the systems become objects for lived engagement, co-design, and public critique, rather than passive recipients of technical documentation or static outputs (Hemment et al., 2019, Hemment et al., 2023, Hemment et al., 2023).

1. Foundational Principles and Formal Definitions

Experiential AI is formally defined as “a new research agenda in which artists and scientists come together to dispel the mystery of algorithms and make their mechanisms vividly apparent,” treating algorithms “not simply as hidden mechanisms to be documented, but as phenomena to be experienced, questioned, and co-shaped” (Hemment et al., 2019). The scope of Experiential AI moves beyond technical explainable AI (XAI) to encompass:

• The creation of tangible artifacts—installations, performances, interactive interfaces, physical computing or sensory experiences—that reveal, manipulate, or dramatize algorithmic operations.
• The design of situations wherein human subjective agency and machine agency are entangled, allowing users to intervene in, or directly sense, the causal chain of AI-derived outcomes.
• The embedding of AI within narrative or role-play contexts that surface its ethical, social, and economic implications in a shared, participatory frame.
• A normative commitment to transparency, accountability, and expanded public engagement in AI conception, deployment, and audit (Hemment et al., 2019, Hemment et al., 2023).

This paradigm differentiates itself from classical XAI (which typically provides feature-level or decision-level explanations) by reframing explanation as a dynamic, artistic, and socially embedded process. Mathematics and formalism are deployed mainly to structure the mappings between technical processes and experiences, e.g., through cyclic mappings: $E = g(f(\mathcal{D}; \theta)), \quad K = h(E, H)$ where $f$ is the AI algorithm, $g$ an artistic/design transformation, $E$ the experience, $H$ a human participant, and $K$ the resulting knowledge or changed awareness (Hemment et al., 2023).

2. Methodologies and Process Models

Experiential AI operationalizes its goals through transdisciplinary collaboration models in which artists, scientists, designers, and engineers co-develop AI artifacts and public interventions (Hemment et al., 2019, Hemment et al., 2023):

• Residency structures pair artists with machine learning labs and technical teams for periods of creative prototyping, with each party introducing methods, tools, and framing from their disciplinary expertise.
• Process models such as the “4As” (Aspect, Algorithm, Affect, Apprehension) form a double-diamond pattern: diverging on social and conceptual issues ($\rightarrow$ Aspect), converging on technical implementations ($\rightarrow$ Algorithm), diverging into creative artifacts ($\rightarrow$ Affect), and converging in public engagement ($\rightarrow$ Apprehension) (Hemment et al., 2023).
• Creative workflows are iterative and recursive:
  1. Identification of opaque algorithmic mechanisms or data flows in need of exposition.
  2. Artistic framing and material selection (e.g., visual, auditory, tactile, narrative media).
  3. Model perturbation or generative transformation (e.g., re-training, bias exacerbation, data remixing) to magnify salient algorithmic dynamics.
  4. Embedding in social context (gallery, workshop, participatory performance).
  5. Observation, documentation, and iteration based on audience response and reflection.

Table: Example Creative Pipeline (Image-Latent Space Shaping) (Hemment et al., 2023)

Stage	Operation/Formula	Role
Data Ingestion	$f$0, $f$1 (two sets of images)	Input Definition
Encoding	$f$2, $f$3	Representation
Axis Defining	$f$4	Semantic Dimension
Slider Exploration	$f$5, $f$6	Interactive Probing
Generation	$f$7	Output Rendering

This structure puts direct manipulation of model internals in artists’ (and in principle, users’) hands, surfacing the boundaries and idiosyncrasies of AI latent spaces.

3. Artistic Practice as Epistemic Mediation

Artistic practice is a cornerstone of Experiential AI, serving not as an ancillary method but as an epistemic strategy to mediate between opaque computational phenomena and interpretative human faculties (Hemment et al., 2019, Hemment et al., 2023). Artworks act as boundary objects: at once encoding technical logic (through data feeds, neural activation patterns, or classification rules) and translating this logic into sensory or narrative forms aligned with cognitive and affective human registers.

Notable modes include:

• Glitch-based visualizations (e.g., recursive feedback in neural layer outputs) expose the nonlinearity and instability of model representations (see Mario Klingemann’s “Neural Glitch”).

• Generative adversarial portraits make biases and construction choices in training corpora both visible and debatable, sparking discussions on authorship and originality (cf. Robbie Barrat’s GAN portraits).
• Wearable adversarial interventions (e.g., CV-Dazzle) dramatize surveillance and resistance by materially staging face recognition system failures on the human body.
• Participatory performances (role-play workshops, interactive installations) that elicit user reflection on data ethics, curation, and control.

Crucially, the role of artists extends to co-creation of new socio-technical configurations, facilitating non-passive, dialogic experiences with algorithmic systems.

4. Agency, Legibility, and Evaluation

Central goals of Experiential AI are to increase both the legibility (the ease with which non-specialists can apprehend AI operations) and the agency (the capacity for users and creators to manipulate, contest, or reconfigure AI outcomes) afforded by intelligent systems (Hemment et al., 2023, Hemment et al., 2023). Evaluation eschews single-metric approaches in favor of mixed methods:

• Legibility is multidimensional, a function of comprehension, emotional resonance, and narrative clarity.
• Agency depends on the degree to which systems can be intervened upon, contested, or reshaped by users within an experiential frame.
• Evaluative practices include qualitative interviews, think-aloud protocols, iterative workshops, and the proposal of pre-/post- intervention surveys (e.g., Likert-scale paired $f$8-tests on artist sense of agency) (Hemment et al., 2023).
• Standardized benchmarks and metrics remain an open research challenge; proposals include cognitive insight, affective engagement, behavioral shifts, collaborative outputs, and reach into public discourse (Hemment et al., 2019).

Evaluation Dimension	Example Assessment
Cognitive Insights	Post-intervention surveys, think-aloud protocols
Affective Engagement	Measurement of curiosity, surprise, criticality
Agency Steps	Count of user interventions, idea proposals
Behavioral Shift	Audits of subsequent AI-skepticism/action-taking
Public Reach	Media coverage, exhibition audience analytics

5. Case Studies, Prototypes, and Field Deployments

A diverse corpus of prototypes substantiates the scope and methodology of Experiential AI:

• “The Zizi Show”: A drag performance avatar system, retrained GANs on underrepresented datasets, used to highlight and remediate biases in body representation; directly engaged LGBTQ+ communities to strengthen agency and dataset dignity (Hemment et al., 2023).
• Climate-AI Installation (“The New Real Observatory”): Interactive platforms blending predictive climate models with generative AI and public-facing sliders, enabling participants to interrogate the data-dependency of environmental models.
• Normalization Investigations: Data remix works that surface statistical constructions of “normalcy” and embedded prejudice (Hemment et al., 2019).
• Poetic Code: Programmatic verse (e.g., Joy Buolamwini’s work) as both demonstration and contestation of commercial AI bias.
• Embodied Artifacts: Wearables that corporeally manifest AI decision boundaries (e.g., Donnarumma’s prosthetics, adversarial fashion).

The domain of educational technology, e.g., AI-literacy curricula integrating experiential labs or role-based simulation, further demonstrates the generalization of these principles for expanding public understanding of AI (Okpala et al., 2024, Warrier et al., 7 Nov 2025).

6. Theoretical and Conceptual Models

The underpinning conceptual model frames transparency not as a one-dimensional exposure of internal weights or feature saliency, but as a composition: $f$9 This formula emphasizes that genuine algorithmic transparency must involve measurable exposure, human-centric interpretive context, and ongoing lived interaction (Hemment et al., 2019).

The “4As” model—Aspect, Algorithm, Affect, Apprehension—encodes the iterative process and multifaceted evaluation needed for situated, embodied AI explanation (Hemment et al., 2023). The dynamic, cyclic mapping $g$0, $g$1 captures the recursive transformation of technical process into experience, and experience into knowledge and system evolution.

7. Limitations, Open Problems, and Future Directions

Experiential AI faces open challenges in scalability and generalization—most existing projects remain bespoke artworks or domain-specific interventions (Hemment et al., 2019, Hemment et al., 2023). The lack of formal, widely adopted evaluation metrics impedes systematic assessment and comparison. Questions of ethical provocation versus constructive engagement, resource constraints in sustaining interdisciplinary residencies, and the risk of instrumentalizing art practice are recognized as substantial concerns.

Research directions include:

• Developing rigorous, mixed-method evaluation protocols that can assess both cognitive and affective outcomes at scale.
• Expanding experiential frameworks into domains such as autonomous vehicles, healthcare, urban planning, and public policy to test generalizability and broaden impact.
• Building institutional infrastructures (funding, community networks, open-source toolkits) for sustainable, collaborative experiential-AI practice.
• Formulating policy and legal frameworks for “experiential audits,” further formalizing the role of lived encounter in AI oversight.
• Advancing theoretical integration with formal XAI metrics (e.g., faithfulness, completeness) and embedding participatory design principles for a new generation of AI governance (Hemment et al., 2019, Hemment et al., 2023).

---

In summary, Experiential AI constitutes a foundational shift from post hoc, reductionist explanation toward a plural, co-creative modality in which AI transparency, agency, and accountability are achieved through tangible, situated experience. It invites the technical and cultural reframing of AI systems as public phenomena to be encountered, challenged, and democratized across creative, scientific, and civic domains (Hemment et al., 2019, Hemment et al., 2023, Hemment et al., 2023).