Tomáš Marušiak
Faculty of Arts, Masaryk University, Brno, Czech Republic
Visiting Researcher, Department of Digital Humanities, King’s College London, United Kingdom (Summer 2026)

 

Abstract

This study develops the Model of Experimental Speculative Aesthetic Processes in AI Art (MEPAA) as a framework for examining how artistic strategies can anticipate, without deterministically predicting, future aesthetic relations among humans, computational systems, and environments. It critically revises an earlier formulation that treated human-centred artificial intelligence (HCAI) as a transitional stage towards artificial general intelligence (AGI) and assigned a privileged date to AGI feasibility. HCAI is instead defined as a normative and design orientation, while AGI is treated as a contested capability construct and modelling horizon. The study distinguishes five forms of prediction frequently conflated in AI discourse: computational prediction, mechanistic scientific prediction, expert elicitation, market-based forecasting, and scenario-based critical anticipation. Blaise Agüera y Arcas’s account of prediction as constitutive of intelligence is interpreted as a theoretical claim about adaptive systems, not evidence for an AGI timetable. Carlo Rovelli’s relational and operational analysis of temporal prediction provides a limited epistemological corrective: prediction proceeds from asymmetric information available to situated users and does not reveal a uniquely fixed future. A critical document analysis of Mario Klingemann, Refik Anadol, Lauren Lee McCarthy, and Sofia Crespo applies five authorial heuristic indicators—recurrent processing, global availability, metarepresentation, relevance selection, and limited auditable reportability. The profiles indicate comparatively stronger evidence for temporal interaction and relevance selection, but persistent limits concerning metarepresentation and causal auditability. The indicators neither form a validated scale nor establish machine consciousness. The revised MEPAA protocol replaces date-led horizons with capability- and governance-led regimes and specifies decomposition, recomposition, intervention, boundary testing, source triangulation, and rival explanations. It also separates the completed conceptual analysis from a prospective empirical phase involving logged artistic experiments, participant interpretation, and expert horizon scanning. Artistic research thereby functions as disciplined anticipation: it constructs contestable situations in which possible futures can be experienced, criticised, and empirically constrained.

Keywords: AI art; artistic research; anticipation; prediction; human-centred AI; artificial general intelligence; artificial consciousness; speculative design; MEPAA; agency

1. Introduction

AI art is not adequately described as the production of images by generative models. It is a heterogeneous field in which datasets, models, interfaces, institutions, artists, audiences, labour, material infrastructures, and ecological costs participate in the formation of aesthetic events. Its research significance therefore lies less in whether a machine can be designated an “artist” than in how aesthetic agency, perception, selection, and responsibility are redistributed across a sociotechnical arrangement (Coeckelbergh, 2023; Manovich & Arielli, 2024; Zylinska, 2020, 2025).

The study addresses four linked research questions:

• RQ1: How can contemporary artistic research strategies in AI art be analysed as anticipatory models of future aesthetic processes without treating artworks as technical forecasts?

• RQ2: Which dimensions of recurrent processing, information availability, metarepresentation, relevance selection, and auditable reportability can be documented in present artistic protocols, and which remain inferential or unassessable?

• RQ3: How can MEPAA relate artistic anticipation to debates about agentic AI and artificial consciousness while maintaining empirical and ontological restraint?

• RQ4: Under what capability, governance, and boundary conditions can an existing artistic strategy be experimentally recomposed into a testable scenario of a future aesthetic process?

The central thesis is that artistic research can prefigure and test relations that may become more consequential as AI systems acquire persistent memory, tool use, multimodal perception, longer planning horizons, and greater operational autonomy. Prefiguration, however, is not prophecy. An artwork can materialise a conditional model of surveillance, delegated judgement, ecological synthesis, or distributed authorship without showing that AGI will arrive, that a system is conscious, or that a given future is probable. MEPAA therefore treats anticipation as a structured research practice for exploring possibility, consequence, and contestability under uncertainty (Auger, 2013; Barendregt & Vaage, 2021; Dunne & Raby, 2013; Miller, 2018).

Three corrections are essential to this formulation. First, AGI has no single accepted operational definition. Capability breadth, performance, and autonomy can vary independently, and claims about “AGI” frequently change with the benchmark or institutional interest at stake (Morris et al., 2024). Second, HCAI is not a lower rung on an inevitable ladder to AGI. It is a normative and methodological orientation concerned with reliable, safe, accountable, and humanly meaningful systems; it can govern narrow, general, or hybrid intelligence arrangements (Capel & Brereton, 2023; Shneiderman, 2022). Third, intelligence, agency, creativity, linguistic self-description, and phenomenal consciousness are not interchangeable properties. Evidence for one does not by itself establish another (Butlin et al., 2026; Chatterjee, 2022; Cogitate Consortium et al., 2025).

The article consequently replaces the earlier claim of “AGI by 2029” with a plural, capability-led horizon. Current forecasts are included only as dated and definition-dependent observations. They calibrate uncertainty; they do not determine the ontology or chronology of MEPAA.

The contribution is threefold. Conceptually, the study differentiates functional agency, intelligence, creativity, consciousness, prediction, forecasting, and anticipation. Analytically, it provides a profile-based reading of four artistic practices and identifies the evidential limits of public project documentation. Methodologically, it translates the MEPAA concept into a prospective experimental protocol based on decomposition, recomposition, scenario construction, boundary testing, and transparent uncertainty reporting. The result is not a claim to have discovered artificial consciousness in art. It is a framework for testing how future-oriented aesthetic relations are technically organised, socially attributed, and institutionally governed.

2. Conceptual Framework

2.1 AI art as a research field rather than a teleological stage

AI art can be distinguished analytically from earlier forms of digital and generative art by the growing role of learned statistical models, large datasets, natural-language interfaces, multimodal generation, and semi-autonomous workflows. Yet it should not be described as a stage whose historical purpose is to culminate in AGI. Such a description mistakes technological succession for necessity and imports a teleology that the evidence cannot support. Generative and computational art have long distributed decisions between human rules, stochastic processes, and machine execution; contemporary AI changes the scale, opacity, adaptivity, and institutional concentration of those relations rather than beginning them ex nihilo (Boden & Edmonds, 2009; Manovich, 2019; Taylor, 2014).

The more defensible proposition is that current AI art offers a laboratory for studying changing distributions of agency. Here, “agentic” denotes the capacity of a component to initiate or select actions within a defined environment and protocol. It does not imply unrestricted autonomy, general intelligence, moral personhood, or consciousness. Evans et al. (2026) theorise future intelligence as socially distributed across agents, human–AI ensembles, and institutions. This account is productive for aesthetics because it shifts attention from a singular superintelligence to the organisation of collective cognition. It remains, however, a theoretical Science perspective rather than empirical evidence of an impending intelligence explosion.

2.2 Agentic Entity, latency, and distributed agency

The concept of the Agentic Entity names a functional position within an artistic system rather than a new ontological class of being. An entity is agentic to the extent that it can initiate or select actions, maintain task-relevant state, adapt to feedback, or alter the causal trajectory of an environment within specified constraints. Degrees and kinds of agency must therefore be described separately. A model may select among outputs without defining its own goals; an installation may adapt to participants while remaining dependent on fixed sensors and rules; and a human–machine collective may produce effects that no individual component controls. None of these configurations requires consciousness.

This functional definition follows the insight that agency can be distributed across heterogeneous networks (Latour, 2005), while resisting the stronger conclusion that all participants possess equivalent autonomy, intention, or moral status. Barad’s (2007) account of intra-action is useful for showing that relata and measurement conditions are co-constituted within a phenomenon. For MEPAA, however, relational constitution does not erase causal asymmetry. The artist, software provider, institution, audience, dataset, and model possess different capacities to set objectives, refuse participation, access records, or bear harm.

The earlier concept of latency is retained in a restricted sense. Latency does not mean that AGI or consciousness already exists secretly inside current artistic systems. It refers to a set of relations, functions, and cultural expectations that may be partially staged before their technical or institutional stabilisation. A work can model delegated judgement, persistent adaptation, machine-mediated care, or synthetic ecology while the underlying system remains narrow and highly scaffolded. Latency is therefore an anticipatory property of the artistic strategy and its reception, not evidence of a hidden machine subject.

This clarification also distinguishes the Agentic Entity from three adjacent concepts. It is not identical with authorship, because causal contribution does not automatically entail cultural or legal authorship. It is not identical with creativity, because selection or novelty can occur without self-directed problem formation. Finally, it is not identical with consciousness, because flexible behaviour and self-description can arise from mechanisms that do not establish subjective experience. The analytical task is to locate agency in a concrete protocol rather than infer it from the fluency, novelty, or anthropomorphic presentation of an output.

2.3 Relational aesthetic processes and two modes of organisation

MEPAA defines an aesthetic process as a temporally extended and relational trajectory produced within a loop linking:

1. the human participant—embodiment, perception, cognition, affect, memory, values, and social position;

2. the computational system—data structures, learned representations, memory, inference, selection, tools, and interface behaviour; and

3. the environment—space, institutions, material infrastructure, audiences, legal rules, social signals, and ecological conditions.

The earlier phrase “interconnected human and AI biologies” is conceptually misleading. Contemporary AI does not constitute a second biology in the ordinary scientific sense. A more precise formulation distinguishes two modes of aesthetic ecosystem organisation: biological and computational. They are coupled but not equivalent. Zylinska’s (2025) concept of Bio-AI is valuable precisely when read as a critical account of biomachines, mediation, and data animism, not as proof that computational models are organisms. Barad’s (2007) relational ontology and Latour’s (2005) analysis of distributed action can illuminate the formation of the assemblage, but they do not remove differences in embodiment, vulnerability, ownership, energy use, or causal architecture.

This distinction also prevents “relation” from becoming an excuse for flattening power. A dataset is not merely another participant; it can encode extraction, exclusion, copyright disputes, and classificatory violence. A model provider, museum, artist, and viewer do not occupy symmetrical positions. Any account of future aesthetic processes must therefore examine who can define goals, inspect logs, stop a process, refuse data collection, receive credit, and absorb environmental or social costs (Costanza-Chock, 2020; Crawford, 2021; Crawford & Paglen, 2021).

2.4 HCAI is a normative regime, not a precursor to AGI

HCAI should be retained in MEPAA as a set of evaluative commitments rather than a technological phase. These commitments include meaningful human control, safety, reliability, auditability, accessibility, contestability, and explicit responsibility (Capel & Brereton, 2023; Shneiderman, 2022). They become more important as systems become more capable, but their relevance does not depend on AGI arriving.

AGI, by contrast, is a disputed family of constructs. Morris et al. (2024) distinguish performance, generality, and autonomy, thereby showing why a single threshold is analytically weak. A system can outperform humans in several domains while remaining unreliable across apparently simple tasks; it can be broadly competent but require human scaffolding; or it can act autonomously within a narrow domain. The International AI Safety Report 2026 similarly emphasises uneven or “jagged” capabilities, evaluation gaps, and limited reliability in real-world deployment (Bengio et al., 2026). MEPAA therefore uses observed capacities and governance conditions as scenario variables rather than presuming a unitary transition from HCAI to AGI.

AGI narratives also carry political histories and normative assumptions. Gebru and Torres (2024) show how influential superintelligence narratives can be entangled with the TESCREAL bundle of long-termist, transhumanist, and related ideologies. Their critique should not be generalised to every operational taxonomy of generality or autonomy, but it prevents MEPAA from treating a trajectory towards AGI as politically neutral or historically inevitable.

2.5 Functional indicators do not establish consciousness

Research on artificial consciousness has become methodologically more precise but remains deeply uncertain. Butlin et al. (2026) propose deriving indicators from multiple scientific theories and using them to update credences about particular systems. They do not offer a consciousness test, and the presence of an indicator is neither necessary nor sufficient evidence of phenomenal experience. Butlin and Lappas (2025) accordingly argue for responsible organisational policies even under uncertainty and warn against misleading public communication.

The large adversarial collaboration conducted by the Cogitate Consortium et al. (2025) tested predictions of global neuronal workspace theory and integrated information theory in 256 human participants using fMRI, MEG, and intracranial EEG. Its results challenged important predictions of both theories without conclusively refuting either. The study concerns human consciousness; it cannot be transferred directly to current AI systems. Its relevance to MEPAA is methodological: consciousness science itself contains competing theories, difficult measurements, and underdetermined interpretations. Strong claims about artificial consciousness are therefore premature.

MEPAA uses five authorial heuristic indicators informed by, but not identical to, theory-derived consciousness indicators:

Indicator	Operational question in an artistic protocol	Permissible evidence	What it does not establish
P1 Recurrent processing and temporal integration	Does earlier state causally affect later processing across an episode?	Versioned state, memory traces, feedback loops, intervention tests	Subjective duration or experience
P2 Global availability or broadcast	Can information from one process alter several other functional modules or actors?	Cross-module propagation, tool calls, shared state, interaction logs	A human-like global workspace or consciousness
P3 Metarepresentation	Does the protocol maintain causally active models of its own, another actor’s, or the task’s state?	Explicit state models, counterfactual probes, interventions on representations	Self-awareness or theory of mind in the phenomenal sense
P4 Attention and relevance selection	How are inputs, goals, and outputs prioritised under conflict or scarcity?	Selection rules, uncertainty, salience changes, ablations	Human interest, intention, or aesthetic judgement
P5 Limited auditable reportability	Can a claim about a decision be linked to logged evidence and tested against alternatives?	Provenance, uncertainty, decision traces, reproducible explanation	Introspection or truthful first-person access

The profile is deliberately heterogeneous. Indicators are not summed into a score, and P5 is especially restricted: fluent explanation can be post-hoc rationalisation, prompted performance, or imitation. Reportability counts only when it is causally and procedurally connected to independently inspectable evidence. The profile is thus a tool for designing and comparing research protocols, not diagnosing minds.

This restriction is reinforced by Pennartz’s (2026) challenge concerning the validation of theory-derived indicators beyond the biological cases from which they originate. MEPAA cannot solve that extrapolation problem by renaming artistic features as indicators; it must test the causal organisation of each protocol and state where evidence is unavailable.

2.6 Artistic research as an epistemic and anticipatory operation

Artistic research is not treated here as scientific prediction by other means. Its epistemic contribution arises from making a problem materially and experientially available through a designed situation. Borgdorff (2012) locates artistic research in practices where the process and product of art participate in the production of knowledge. Biggs and Karlsson (2012) similarly show that research in the arts involves heterogeneous relations among practice, documentation, interpretation, and public articulation. Rheinberger’s (2023) phenomenology of experimentation adds an important qualification: experiments do not merely verify fully specified propositions; they can create conditions in which previously unavailable questions and epistemic objects emerge.

Within MEPAA, an artistic research strategy becomes an epistemic vector when it meets four conditions. First, it articulates a problem that exceeds formal style or the production of a singular artefact. Second, it translates that problem into operations, constraints, or interactions that can be described and modified. Third, it produces traces that permit reflection on what occurred, including failure and ambiguity. Fourth, it makes its inferential limits explicit. An immersive installation, participatory performance, or synthetic image series is not research solely because it uses AI. It becomes research-relevant when its protocol supports a disciplined relation among question, intervention, evidence, and interpretation.

Schwab’s (2018) notion of transpositionality is useful here because artistic research can transpose methods across domains without pretending that their epistemic standards are identical. A consciousness indicator, a futures scenario, or an ablation test can enter an artistic context, but its meaning changes with the new material and institutional arrangement. MEPAA therefore does not import neuroscience as an aesthetic authority. It translates selected functional questions into artistic experiments and then subjects those translations to critical evaluation.

Anticipatory potential is consequently not an intrinsic quality that can be read directly from an artwork. It is a relation among a strategy, an articulated future condition, present evidence, and a method of recomposition. The study operationalises this relation rather than presuming it.

3. Method

3.1 Research design and scope

The study combines a critical literature synthesis with a comparative qualitative case-study design. Its purpose is analytical generalisation: to develop concepts and test the internal usefulness of a framework across contrasting practices. It does not seek statistical generalisation to all AI art. Four core cases were selected because they produce maximum theoretical variation within the human–system–environment loop:

• Mario Klingemann: iterative generation, audience feedback, and delegated selection;

• Refik Anadol: large-scale datasets, immersive display, and institutional spectacle;

• Lauren Lee McCarthy: social automation, surveillance, care, consent, and delegated decision-making; and

• Sofia Crespo: synthetic morphology, biological imagination, taxonomy, and ecological representation.

The temporal focus is primarily 2018–2026, although earlier projects are included where they establish a strategy subsequently developed in the focal period. Selection criteria were: sustained engagement with computational systems; a process that extends beyond one-off image generation; sufficient public documentation to reconstruct at least part of the protocol; variation in interaction and institutional form; and relevance to at least two P1–P5 questions. Fame, market value, or an artist’s own use of consciousness terminology was not a selection criterion.

A secondary contextual corpus—used for transferability rather than full P1–P5 profiling—includes work by Sougwen Chung, Ian Cheng, Pierre Huyghe, Holly Herndon and Mat Dryhurst, Michael Sedbon, and Christa Sommerer and Laurent Mignonneau. These practices extend the comparison towards robotic co-production, live simulations, autonomous environments, collective voice models, biohybrid computation, and interactive artificial-life traditions. They are discussed in Section 4.6 but are not treated as equivalent in evidential depth to the four core cases.

3.2 Unit of analysis and three-level decomposition

The principal unit of analysis is the artistic research strategy: a repeatable configuration of questions, technical operations, interaction rules, institutional conditions, and modes of public presentation. An individual work is evidence of that strategy only insofar as documentation supports claims about the protocol. A visually coherent output does not reveal which component generated, filtered, selected, arranged, or explained it.

Each strategy is decomposed at three levels.

1. Conceptual level: the problem articulated by the work, its metaphors, declared research aims, model of agency, and relation to histories of art and technology.

2. Process-technical level: datasets, sensors, models, memory, feedback, selection, human intervention, interfaces, display systems, and documented failure modes.

3. Reception-epistemic level: what participants can perceive or know about the process; how agency is attributed; how the institution frames the system; and whether claims can be challenged, reproduced, or audited.

The three levels prevent a common inferential collapse. A metaphor such as “the machine dreams” belongs initially to conceptual framing. Recurrent state or generative iteration belongs to the process-technical level. A viewer’s experience of an apparently continuous subject belongs to reception. Evidence at one level cannot be used automatically to establish a claim at another.

3.3 Source hierarchy and verification procedure

Sources were assessed according to the claim they can legitimately support:

1. peer-reviewed research and authoritative scientific reports for capability, consciousness, HCAI, and forecasting claims;

2. primary artist, museum, and project documentation for dates, interfaces, datasets, and declared procedures;

3. scholarly criticism for institutional, political, ecological, and aesthetic interpretation;

4. interviews and artist statements as situated self-description, not independent validation; and

5. prediction platforms only as time-stamped observations of heterogeneous public beliefs.

The review privileges DOI records, journal or conference pages, books from established scholarly presses, official reports, and first-party project documentation. Artist websites are suitable for identifying titles, dates, and stated procedures, but not for verifying hidden model mechanisms or the success of ecological and social claims. Press coverage is used only when it documents otherwise unavailable institutional context and is not treated as mechanistic evidence.

Every analytical claim is assigned one of four statuses:

• D—documented: directly supported by a traceable source or observable protocol feature;

• I—inferred: a theoretically reasoned interpretation consistent with the available evidence but not directly verified;

• S—scenario proposition: a conditional claim introduced for anticipatory modelling; or

• U—unassessable: the required evidence is unavailable, inaccessible, or too ambiguous.

Where internal logs, model versions, prompts, datasets, or selection histories are inaccessible, the study treats mechanistic attribution as underdetermined. Rival explanations—including human curation, interface framing, stochastic variation, audience projection, maintenance labour, and institutional marketing—are considered before agency is attributed to a system.

3.4 P1–P5 coding and decision rules

Evidence is coded by episode, defined as a bounded interaction or generative sequence with identifiable inputs, state changes, decisions, and outputs. Coding an entire career or installation as simply “high” in an indicator would conceal differences among versions, interfaces, and moments of use. Each indicator therefore requires a minimal causal criterion.

• P1 requires evidence that information from an earlier state alters later processing. Mere animation, looping video, or chronological display does not qualify.

• P2 requires propagation from one local process to at least two other functional components or actors. Visual scale or multiscreen presentation alone does not qualify.

• P3 requires a representation of a state that is used to regulate further action. A portrait, mirror image, or first-person sentence does not qualify by itself.

• P4 requires a consequential prioritisation among competing inputs, goals, or outputs. Human curation must be distinguished from model-level selection.

• P5 requires a report linked to contemporaneous records, provenance, or interventions that permit an independent reconstruction of at least part of a decision. Fluent natural-language explanation without causal linkage is insufficient.

The qualitative code for each case is reported as documented, plausible but indirect, weak, or unassessable. These categories describe evidential support, not the quantity of consciousness. No indicator is weighted, and no total score is calculated.

3.5 Comparative analysis and validity safeguards

The analysis proceeds in five steps: within-case reconstruction; coding of episodes; identification of rival explanations; cross-case comparison; and translation of findings into prospective MEPAA interventions. Validity is strengthened through source triangulation, negative-case reasoning, explicit uncertainty, and separation of artistic self-description from independent criticism. A claim is downgraded when it depends on a single promotional source, when the system version is unknown, or when the same effect can be explained by human curation.

Reliability in a future empirical application would require a documented codebook, two or more independent coders for a subset of episodes, resolution of disagreements by recorded discussion, and publication of anonymised coding examples. Because the current article is a conceptual synthesis conducted by a single researcher, it cannot claim intercoder reliability. Its contribution is to specify coding decisions sufficiently clearly for later replication and criticism.

3.6 Status of the present evidence

This study does not report a completed laboratory or field experiment, participant data, direct access to proprietary model internals, or causal interventions in the four cases. Its “findings” are therefore findings of critical document analysis: they concern what can responsibly be inferred from the available record. The P1–P5 profiles are provisional hypotheses that guide a second, prospective research phase.

That prospective phase will select an artistic strategy, implement a documented recomposition, expose it to controlled scenario variations, and collect process logs, participant reports, observational records, and expert assessment. The distinction between the completed conceptual study and proposed empirical validation is maintained throughout the article.

4. Critical Comparative Analysis of Artistic Strategies

4.1 Mario Klingemann: selection, feedback, and delegated evaluation

Klingemann’s practice is significant because it repeatedly moves the aesthetic problem from the generation of an individual image to the organisation of a selection process. Neural Glitch/Mistaken Identity intervenes in learned visual structures and treats computational error as a productive operation rather than a defect (Klingemann, 2018). Memories of Passersby I presents an apparently continuous stream of portraits, foregrounding temporal generation and the impossibility of reducing the work to a single authorised image. Uncanny Mirror introduces real-time portrait transformation and makes the participant’s appearance part of the generative relation. These works establish a trajectory from latent-space exploration towards systems in which generation unfolds publicly and viewers attribute continuity or intention to a sequence (del Campo, 2024).

The distinction among artistic framing, actual recurrence, and viewer attribution is crucial. A continuous display may be implemented as independent samples rather than as a system that retains a causally active history. P1 can therefore be documented only where prior state, feedback, or accumulated evaluation alters later processing. Publicly available documentation supports temporal iteration but does not always reveal whether the portrait stream contains persistent model memory. The relevant MEPAA experiment would compare an unchanged generative sequence with a version in which state is deliberately reset, then test whether participants and logs register a difference in continuity.

Circuit Training provides stronger evidence of a feedback architecture. Audience evaluations are incorporated into a cycle of production, learning, and presentation, making the social distribution of selection part of the artwork (Klingemann, 2019). P4 is documented at the level of the protocol because candidates are ranked and some are privileged over others. P2 is plausible where audience input propagates across generation, evaluation, and display. Yet the value criterion remains hybrid: the system does not independently establish what “interesting” means; human ratings, interface design, and the artist’s prior choices structure the selection space.

Botto extends distributed evaluation through a community that votes on generated fragments and thereby affects subsequent production (Onkaos, n.d.-b). Its importance for MEPAA lies not in whether Botto is “the artist”, but in the governance of aesthetic relevance. Generation, machine-level filtering, community voting, economic incentives, and narrative framing form a multi-actor decision system. The work makes authorship a protocol-level question. It also shows why more agency does not necessarily mean more autonomy: greater complexity can increase dependence on governance rules and platform infrastructure.

Appropriate Response stages linguistic output as an oracular encounter, while Neural Decay makes degradation and loss into generative principles (Onkaos, n.d.-a, n.d.-c). In both cases, viewers may attribute interiority to outputs whose causal history remains opaque. A.I.C.C.A. intensifies this problem by embodying machine criticism as a robotic performance (Klingemann, 2023). The work redistributes institutional authority by asking what happens when critique is delivered through an apparently autonomous object. It does not demonstrate that the critic possesses a reflective model of its own aesthetic standards.

Klingemann’s provisional profile is therefore strongest for P4 and, in feedback-based works, P1. P2 is plausible when local evaluations modify several components. P3 remains indirect because responsive portraiture and generated criticism do not by themselves establish causally active self- or other-models. P5 is weak: procedural descriptions and visible votes improve legibility, but they rarely allow reconstruction of why a specific candidate was generated, rejected, or explained. The anticipatory potential lies in the explicit governance of selection. A future MEPAA recomposition should expose candidate histories, selection thresholds, uncertainty, community influence, and human overrides, then introduce conflicts between model salience and public preference.

4.2 Refik Anadol: data immersion, scale, and the problem of infrastructural opacity

Anadol’s practice translates image, sensor, architectural, and natural-history datasets into spatially extended, temporally unfolding installations. His concepts of “data painting” and “data sculpture” frame machine-learning processes as architectural and perceptual events (Anadol, 2020, 2022). Melting Memories connects neural measurement, cultural ideas of memory, generative transformation, and large-scale display. Machine Hallucinations organises very large image corpora into continuous audiovisual environments. Unsupervised uses material from the Museum of Modern Art’s collection as a generative archive and positions the museum itself as both dataset and institutional frame (Museum of Modern Art, 2022).

These works make P1, P2, and P4 appear intuitively strong, but the intuition must be decomposed. Temporal animation is observable, yet it may result from a precomputed sequence rather than recurrent state. Coordination across projection, sound, architecture, and data processing supports a system-level account of P2, but perceptual unity for the viewer is not equivalent to computational broadcast. Selection is undeniable at the level of dataset construction, model development, filtering, and exhibition. The unresolved question is where this selection occurs and who controls it.

Anadol’s metaphors are epistemically productive and risky. “Memory”, “dream”, and “hallucination” make otherwise invisible computations culturally legible, but they also import biological and phenomenological associations. The fact that a system transforms an archive does not show that it remembers; the fact that outputs are unfamiliar does not show that it hallucinates in a clinical or experiential sense. MEPAA therefore codes such terms as conceptual framing unless the protocol supplies independent evidence of state retention, prediction error, or self-modelling.

The progression from museum archives to environmental datasets intensifies these questions. Winds of Yawanawá links generated imagery with collaboration and environmental data, while later projects associated with DATALAND and the Large Nature Model frame machine learning as a means of interpreting natural archives (DATALAND, n.d., 2026; Possible Futures, 2023). Primary documentation can establish the declared collaboration, data sources, and presentation. It cannot by itself establish that a computational system represents ecological agency, benefits the communities invoked, or offsets the material costs of computation.

This evidential distinction is central to the critique by Zeilinger and Johns (2025), who argue that environmental spectacle may absorb ecological difference into a visualising system and thereby foreclose rather than extend non-human agency. Their interpretation is not a direct measurement of every Anadol project, but it supplies a necessary rival account to celebratory institutional narratives. A scientific analysis must therefore include dataset provenance, consent and benefit-sharing, compute and energy, model ownership, cultural authority, and the difference between representing an ecology and participating responsibly in one.

P3 is weak or unassessable because the works do not normally document a stable, causally active model of the system’s own state or the participant as a distinct agent. P5 is similarly limited: pipeline descriptions provide a broad account of production but rarely expose rejected outputs, uncertainty, state transitions, or the causal role of curatorial intervention. Anadol’s anticipatory value lies in staging a future in which cultural and environmental archives are processed as responsive media at architectural scale. A MEPAA recomposition would make that future testable by exposing provenance, energy, uncertainty, and the consequences of removing a data source, community authorisation, or display modality.

4.3 Lauren Lee McCarthy: social prediction, surveillance, and contestability

McCarthy’s practice provides the clearest case for analysing AI as a social protocol rather than an image generator. Her projects ask what happens when prediction, care, surveillance, and decision-making are delegated to technical systems—or appear to be delegated. Follower hires a person to follow a participant, converting a platform-like request into an embodied relation (McCarthy, 2016). LAUREN places the artist in the role of a human smart-home system, while SOMEONE distributes remote observation and assistance among performers and domestic participants (McCarthy, 2017, 2019). These projects deliberately blur automation and human labour.

The ambiguity is methodologically valuable. If participants behave differently because they believe a system is watching, predicting, or caring for them, the aesthetic effect cannot be reduced to the actual level of automation. Perceived agency becomes causally effective. MEPAA must therefore record both technical agency and attributed agency, then examine how labels, interface design, institutional framing, and hidden labour produce the difference.

Unlearning Language organises a social situation in which machine recognition pressures participants to modify or conceal communication (McCarthy et al., 2022). P1 appears in adaptation across an episode: participants learn from previous responses and change their tactics. P2 operates sociotechnically when a detected signal alters communication, spatial behaviour, normative expectations, and group strategy. P3 is distributed rather than located solely in the model: participants model how the system reads them, other participants model one another, and the technical protocol classifies selected signals. P4 becomes visible through the competition between machine-readable relevance and human counter-strategies.

Voice in My Head makes externalised guidance an explicit dramaturgical device, while AUTO develops questions of bodily autonomy and delegated action (McCarthy, 2023, 2025). These works anticipate agentic systems less by presenting technologically independent agents than by staging the social conditions under which people surrender, negotiate, or resist decision authority. The critical variable is not simply system accuracy. It is whether participants know when an intervention is automated, who can interrupt it, and how dependency changes interpretation.

P5 is therefore more than a transparency feature in McCarthy’s practice; it is an aesthetic and political problem. A procedure may be declared in advance while particular inferences remain inaccessible. Conversely, a fully visible rule may still exert coercive pressure. A robust analysis must distinguish procedural transparency, local explanation, causal auditability, and meaningful contestability. Participants should be able to know what data were collected, challenge an intervention, withdraw where feasible, and identify the responsible human or institution.

McCarthy’s profile is comparatively strong in P1, distributed P2, P3 as reciprocal modelling, and P4. P5 remains partial but is unusually central to the works‘ declared questions. The principal rival explanation is that human performers and dramaturgy supply much of the apparent intelligence. Rather than disqualifying the work, this exposes a likely future condition: agentic AI will often be a hybrid organisation of models, interfaces, remote labour, organisational rules, and user belief. A MEPAA recomposition can therefore test matched versions of a protocol with different degrees of actual automation, disclosure, appeal, and human override.

4.4 Sofia Crespo: synthetic morphology and the danger of biological metaphor

Crespo’s Neural Zoo, Artificial Natural History, Critically Extant, and Perpetual Present use learned visual models to construct speculative organisms, fragments, and taxonomies (Crespo, 2018–2022, 2020–2024, 2021–2022, 2024). The projects are relevant not because they demonstrate computational life but because they test how machine-learned morphology reorganises human expectations of species, extinction, evidence, and the natural archive.

Neural Zoo presents forms that appear biologically plausible while remaining taxonomically unstable. Their perceptual force derives from learned regularities in images of organisms and from the human tendency to complete ambiguous forms. Novelty is therefore relational: the image may be statistically recombined, curatorially selected, and biologically interpreted at once. A claim of “new species” belongs to speculative framing unless the work supplies an evolutionary environment, reproduction, inheritance, selection pressure, and a criterion of persistence.

Artificial Natural History extends the logic of the scientific plate and collection. It does not merely generate images; it simulates an epistemic apparatus through which organisms are isolated, classified, and made visible. The strategy anticipates a future in which generative models participate in scientific imagination, but it also exposes the risk that synthetic plausibility may be mistaken for empirical observation. The distinction among document, model output, and speculation must therefore remain visible.

Critically Extant connects generative representation with vulnerable or endangered life. This raises questions of evidential and moral substitution: can a synthetic image amplify attention to a threatened organism, or does it displace the organism with an aesthetically optimised proxy? Perpetual Present further complicates the temporal relation between biological record and computational production. In MEPAA terms, the works are strongest as experiments in P4—what morphological features, species, or patterns become relevant—and as reception-level tests of categorisation.

P1 is plausible in iterative production workflows but remains unassessable as persistent model memory without technical documentation. P2 may describe relations among local traits, global form, dataset, and curatorial taxonomy, yet perceptual coherence is not sufficient evidence of architectural broadcast. P3 is weak because a system that produces organism-like images need not model itself, an organism, or a viewer as an agent. P5 is limited to documentation of datasets, workflows, and selection unless decision traces and rejected candidates are retained.

Crespo’s work clarifies why MEPAA distinguishes representation, simulation, and living process. Computational models can participate in a biological imaginary without becoming organisms. The anticipatory question is not whether AI “evolves” new life, but how synthetic morphology may influence scientific expectation, ecological concern, and aesthetic value. A MEPAA recomposition would compare different provenance disclosures, involve biological expertise, introduce constraints derived from endangered species data, and test whether participants distinguish empirical, simulated, and speculative claims.

4.5 Cross-case findings

The comparative profile summarises the strength of available evidence, not the magnitude of a mental property.

Case	P1 temporal integration	P2 availability/broadcast	P3 metarepresentation	P4 relevance selection	P5 auditable reportability	Dominant rival explanation
Klingemann	Plausible; documented in feedback structures, uncertain in continuous portrait streams	Plausible in audience–generation–display loops	Weak to indirect	Documented in ranking and voting protocols	Weak to partial	Human curation, community governance, interface attribution
Anadol	Plausible as temporal generation; persistent state often unassessable	Plausible across data, media, and display components	Weak/unassessable	Documented at pipeline and curatorial levels; model share unclear	Weak	Curatorial selection and immersive engineering create apparent system unity
McCarthy	Documented at the human–system episode level	Documented sociotechnically	Plausible as reciprocal human/system modelling	Documented in surveillance and intervention rules	Partial and explicitly problematised	Human performance and dramaturgy supply apparent automation
Crespo	Plausible in iterative workflow; model memory unassessable	Indirect at morphological and pipeline levels	Weak	Documented in feature and curatorial selection	Weak to partial	Human taxonomy and selection produce apparent synthetic ecology

Across the cases, P1 and P4 are the easiest indicators to document because iteration and selection are often explicit. P2 is frequently plausible but risks conflating technical propagation with institutional or perceptual coherence. P3 is commonly over-attributed from portraits, first-person language, or biological metaphors. P5 is the recurrent bottleneck because public documentation rarely exposes causal state transitions, uncertainty, overrides, and unrealised alternatives.

The comparison also reveals four recurring transformations of the artwork. First, the artwork shifts from a stable object to a process trajectory. Second, authorship shifts from individual intention to governance of selection. Third, spectatorship shifts towards participation, data provision, and strategic adaptation. Fourth, curatorship shifts towards the design of attention: deciding what enters the dataset, which outputs become visible, and whose interpretation receives authority.

The result is asymmetrical. Artistic practices offer strong evidence that protocols redistribute attention, authorship, and social expectation. They offer much weaker evidence about internal model organisation and no direct evidence of phenomenal consciousness. Their anticipatory value lies in turning the consequences of possible agency into experiential and criticisable situations—not in functioning as prototypes of conscious AGI.

4.6 Contextual transfer to a wider field, 2024–2026

The four-case framework is not exhaustive. Recent and historically continuous practices indicate how MEPAA may transfer to other modes of AI art. Sougwen Chung’s robotic drawing research foregrounds embodied co-production, gesture archives, and recursive human–machine learning (Chung, 2026a, 2026b). Ian Cheng’s live simulations, from Emissaries to BOB and later work, offer a long-duration laboratory for virtual agents, narrative contingency, and audience attribution (Cheng, 2015–2017; Serpentine Galleries, 2018). Pierre Huyghe’s Liminal and Camata extend agency across machine learning, sensors, robotics, organisms, and exhibition environments (Huyghe, 2024a, 2024b).

Holly Herndon and Mat Dryhurst’s The Call makes collective voice, consent, training data, and community governance central to generative music (Herndon & Dryhurst, 2024). Michael Sedbon’s biohybrid projects connect living processes with computation and therefore demand especially strict separation of biological agency, computational control, and metaphor (Sedbon, 2024, n.d.). Christa Sommerer and Laurent Mignonneau’s interactive artificial-life tradition, including Flâneur, demonstrates that relational and evolving aesthetic systems predate the current generative-AI wave (Sommerer & Mignonneau, 2026).

These examples support the transferability of the human–system–environment unit, but they also warn against a universal profile. Robotic gesture, live simulation, collective voice, and biohybrid systems require different evidence and ethical safeguards. They should enter MEPAA as new purposive cases, not as illustrations forced into conclusions derived from the initial four.

4.7 Provisional answers to RQ1 and RQ2

For RQ1, artistic strategies possess anticipatory potential when they operationalise a future-oriented relation—such as delegated judgement, adaptive care, synthetic ecology, or distributed authorship—in a protocol that can be described, modified, and criticised. Anticipatory potential is strongest where consequences are experienced and assumptions can be challenged. It is weakest where futurity is supplied only by promotional language.

For RQ2, the document analysis indicates that temporal integration and relevance selection are most accessible to present study. Information broadcast can be examined when system components and propagation paths are documented. Metarepresentation requires much stronger causal evidence than responsive imagery or language. Auditable reportability remains the most serious limitation and the most productive design target. No case supports an inference to phenomenal consciousness.

5. Prediction Is Not One Method

The word prediction names several epistemically different operations in AI research and futures discourse. Conflating them produces category errors—for example, inferring an AGI date from a model’s next-token prediction, or treating a prediction-market price as an empirically validated scientific probability. MEPAA distinguishes five forms.

Form	Object and method	Legitimate output	Principal limitation	Role in MEPAA
A. Computational prediction	A model estimates a next state, token, reward, sensory input, or error from data and an objective function	Conditional model output or policy update	Depends on training distribution, representation, objective, and environment; may fail out of distribution	A process variable that can shape an artwork
B. Mechanistic scientific prediction	A specified theory and measurement model predicts an observable outcome under controlled conditions	Testable, replicable hypothesis with uncertainty	External validity and theory underdetermination	Tests present protocol claims, including P1–P5
C. Expert elicitation	Researchers report calibrated beliefs under an operational definition	Distribution of expert judgement	Sampling, framing, definitions, incentives, and long tails	Broad uncertainty calibration only
D. Market or platform forecast	Prices or aggregated forecasts track beliefs under resolution criteria	Time-stamped collective estimate	Liquidity, selection effects, legal constraints, and non-comparable questions	Dated horizon scan, never a master timeline
E. Scenario-based critical anticipation	Researchers or artists construct conditional worlds and artefacts to expose consequences and choices	Plural, contestable “what-if” models	Not a probability estimate; quality depends on explicit assumptions and inclusion	Core futures method of MEPAA

5.1 Computational prediction: Agüera y Arcas’s ambitious thesis

Agüera y Arcas (2025) argues that prediction is fundamental to intelligence across evolution, brains, and contemporary AI. The thesis connects several levels: organisms anticipate conditions relevant to survival; nervous systems reduce discrepancies between expectation and sensory input; reinforcement-learning systems update expectations of future value; and generative models estimate conditional continuations. It resonates with predictive-processing accounts in cognitive science, in which hierarchical generative models minimise prediction error (Clark, 2013; Friston, 2010).

This synthesis is theoretically generative for AI art. Aesthetic interaction can be modelled as reciprocal prediction: participants anticipate system behaviour, systems estimate likely inputs or preferred outputs, and institutions structure both sets of expectations. Surprise is not the absence of prediction but a relation between expectation, error, and revision. An adaptive artwork can therefore be designed to expose whose priors are encoded, which errors matter, and who is authorised to update the model.

Nevertheless, three cautions are necessary. First, the term prediction spans mechanisms that are not demonstrably identical across living systems, brains, and machine-learning architectures. Functional analogy is not evidence of biological equivalence. Second, predictive success does not by itself establish understanding, agency, or consciousness. A system may accurately estimate continuations while lacking persistent goals, counterfactual self-models, embodied stakes, or phenomenal experience. Third, a theory that makes prediction constitutive of intelligence does not yield a calendar forecast of AGI. It addresses what intelligence may do, not when a contested sociotechnical category will be declared achieved.

The same restraint applies to the public dialogue between Agüera y Arcas and Carlo Rovelli in 2026 (Futurology & Berggruen Institute, 2026). Claims in a public intellectual conversation—such as the proposition that AGI has already arrived or that intelligence is intrinsically social—should be treated as philosophical positions, not as peer-reviewed empirical findings. Evans et al. (2026) provide a more formal account of socially distributed intelligence, but their article remains a perspective and not a measurement of an “intelligence explosion”.

5.2 Rovelli: prediction from situated information, not access to a fixed future

Rovelli does not offer a theory of AI forecasting. His work is relevant as an epistemological corrective. In The Order of Time, time is not presented as a single universal flow independent of physical relations and scale (Rovelli, 2018). More specifically, Di Biagio et al. (2021) show that the time orientation of operational quantum formalisms—predicting the future from records of the past—does not necessarily reflect a fundamental temporal orientation in microscopic physics. It is built into the informational situation of users: some variables are known, others unknown.

MEPAA draws a limited methodological inference from this analysis. Forecasts are relations among a model, a target, available records, a time of assessment, and a resolution rule. They are not views from nowhere. Changing the information set, scale, operational definition, or institutional observer changes what can responsibly be predicted. This does not make all forecasts arbitrary, nor does quantum physics directly determine futures studies. It means that a forecast’s conditions of production must remain visible.

The contrast with Agüera y Arcas is therefore productive rather than oppositional. Agüera y Arcas foregrounds prediction as an adaptive operation internal to intelligence; Rovelli foregrounds the relational and informational conditions under which temporal prediction becomes possible. Together they support a central MEPAA distinction: systemic prediction is a functional process inside an agent or model, whereas historical forecasting is a situated claim about future sociotechnical events. Neither should be confused with artistic anticipation, which constructs possible relations in order to examine their consequences.

5.3 Expert surveys and the instability of AGI definitions

Grace et al. (2025) report responses from 2,778 AI researchers. Under the survey’s definition—unaided machines outperforming humans in every possible task—respondents assigned a 10% probability by 2027 and a 50% probability by 2047. The distributions are wide, and estimates for full automation of labour extend much further into the future. These results are evidence about researchers‘ beliefs at the time of elicitation, not evidence that the predicted system is technically feasible on those dates.

Several inferential limits follow. Survey respondents do not share a complete causal model of AI development; their expertise differs by domain; definitions and question order matter; and aggregate medians conceal disagreement. Moreover, “high-level machine intelligence”, “transformative AI”, “human-level AI”, and “AGI” are not interchangeable labels. MEPAA should use such surveys to preserve uncertainty and identify competing assumptions, not to choose a deadline.

5.4 Forecast platforms as dated and non-comparable observations

On 20 August 2026, three public platforms displayed materially different numbers:

Platform and question	Snapshot on 20 August 2026	Resolution object
Metaculus: first general AI devised, tested, and publicly announced	Community estimate: March 2033	A platform-specific AGI definition and adjudication process
Kalshi: OpenAI achieves AGI before 2030	Approximately 44%	A regulated event contract concerning OpenAI and specified criteria
Polymarket: OpenAI announces AGI before 2027	Approximately 9%	An official OpenAI announcement by 31 December 2026

These numbers are not contradictory measurements of one variable. The Polymarket contract concerns an announcement, the Kalshi market concerns a specific organisation and threshold, and Metaculus uses a broader adjudicated definition. Participation, liquidity, regulation, and resolution rules also differ. The values are volatile and should always carry a retrieval date (Kalshi, n.d.; Metaculus, n.d.; Polymarket, n.d.). Their scientific use is limited to analysing public expectations and changes in those expectations. They cannot justify the earlier assertion of a 2029 transition.

5.5 Critical anticipation as the appropriate futures method

Scenario-based anticipation does not ask only when a technology will arrive. It asks what combinations of capacity, infrastructure, institutional authority, values, and resistance could produce a regime, and what would follow for differently situated actors. Speculative design supports this inquiry by materialising propositions that can be inhabited and contested rather than accepted as extrapolated fact (Auger, 2013; Barendregt & Vaage, 2021; Dunne & Raby, 2013). Futures literacy similarly treats anticipation as a capacity to use imagined futures to perceive and act differently in the present (Miller, 2018).

Anticipation becomes research-grade only when its assumptions, mechanisms, affected groups, and disconfirming conditions are explicit. A beautiful speculative image without a stated causal pathway is not a forecast. Conversely, an empirically constrained scenario need not assign a probability to be useful. Its validity lies in internal coherence, plural perspectives, traceable premises, critical leverage, and the possibility of testing selected claims in present systems.

The widely discussed AI 2027 project illustrates this distinction. It is an explicit, conditional scenario supported by forecasts and technical assumptions, but it is not a peer-reviewed scientific prediction (Kokotajlo et al., 2025). Subsequent changes in its authors‘ timelines further demonstrate that such work should be read as an updateable model rather than a revelation of a fixed future (Lifland et al., 2026). MEPAA may use it as one scenario input only when rival scenarios, assumptions, and disconfirming evidence are also included.

6. The Revised MEPAA Framework

6.1 From calendar horizons to capability-and-governance regimes

The original H1–H3 model attached date ranges to a dominant present, a transition, and an emergent future. This implied greater chronological knowledge than the evidence supports and created an inconsistency when the text later referred to H1–H5. The revised model retains three regimes but defines them by observable properties. Dates may be added as scenario metadata and updated without changing the model.

Regime	Capability conditions	Governance and aesthetic conditions	Research question
R1 Bounded operational systems	Episodic or limited memory; narrow tool use; significant human prompting and curation; uneven reliability	Responsibility remains human and institutional; provenance and consent are often incomplete	Which apparent forms of agency are produced by the interface, workflow, and audience attribution?
R2 Agentic sociotechnical ecologies	Persistent state; multimodal sensing; coordinated tools or agents; longer task horizons; partial adaptation	Control becomes distributed; override, monitoring, data rights, labour, and institutional alignment become central aesthetic materials	Can the protocol remain legible, contestable, and safe while initiative is redistributed?
R3 Contested emergent regimes	Broader transfer and autonomy may occur, but generality and consciousness remain separately assessed and uncertain	Legitimacy depends on public oversight, responsibility, ecological limits, cultural plurality, and moral uncertainty	What new aesthetic relations become possible, and which forms should be refused or constrained?

R1–R3 are not inevitable stages. A practice may combine properties from several regimes; systems may regress; regulation can prevent a deployment; and cultural futures can diverge. R3 does not presuppose AGI or machine consciousness. It identifies a space in which stronger autonomy, social dependence, and moral uncertainty would require more demanding evidence and governance.

6.2 Conditions of realisability

MEPAA evaluates a future aesthetic process through conditions of realisability rather than a predicted arrival date. A scenario becomes more than fiction when its necessary conditions can be stated and at least some can be tested in the present.

Condition domain	Minimum condition for an experimental scenario	Stronger condition for future transfer	Disconfirming or boundary observation
Temporal organisation	State from an earlier episode can be stored and causally reused	Persistent, context-sensitive memory across sessions with controlled forgetting	Reset produces no difference; continuity is only display-level
Perception and action	At least one input channel changes a consequential system action	Multimodal perception and tool use coordinated over a longer task	Inputs are decorative or all meaningful actions are pre-scripted
Relevance	Competing inputs or outputs are ranked by an explicit rule	Priorities adapt under conflict while remaining inspectable and governable	Selection is wholly post-hoc human curation
Metarepresentation	A model of task, participant, or system state changes action	Counterfactual self/other models support correction and perspective-sensitive interaction	First-person language persists after the putative state model is removed
Reportability	Decisions can be reconstructed from provenance and logs	Reports remain calibrated under intervention and identify uncertainty and alternatives	Explanations are fluent but unrelated to causal traces
Human governance	Roles, consent, override, responsibility, and stop rules are explicit	Appeal, refusal, redress, and shared authority are institutionally supported	No actor can identify who is responsible for harmful output
Cultural legitimacy	Affected actors can interpret and contest the scenario	Multiple communities participate in defining values and success	One institutional narrative suppresses conflicting interpretations
Material sustainability	Compute, data sources, and infrastructure are documented	Ecological limits and benefit distribution shape design choices	Environmental claims cannot be reconciled with unreported extraction

These conditions are conjunctive only within a specified scenario. For example, a work can investigate relevance selection without implementing metarepresentation. The framework therefore avoids treating an incomplete profile as a failed attempt at a human-like mind. The research question is whether the implemented conditions are sufficient for the claimed aesthetic process.

6.3 From decomposition to recomposition

Recomposition is the methodological step recovered from the original MEPAA proposal. It translates a documented artistic strategy into a modified protocol while preserving its research problem. The purpose is not to imitate the artist’s style or claim authorial continuity. It is to isolate a relational operation and expose it to variables that the original work did not or could not make testable.

Decomposed element	Recomposition question	Example intervention	Evidence generated
Artistic problem	Which future relation is being tested?	Reformulate “machine memory” as a test of state-dependent continuity	Explicit construct definition
Input and dataset	Whose records shape the process?	Remove, substitute, or permission-gate a data source	Provenance and sensitivity trace
State and memory	What information persists?	Reset, corrupt, or partition memory between episodes	Causal evidence for P1
Broadcast architecture	Which component can influence which other components?	Block one communication path	Causal evidence for P2
Actor models	Is a self/other representation causally used?	Replace or perturb the represented state	Causal evidence for P3
Selection	Who determines relevance under conflict?	Introduce incompatible human and system priorities	Causal evidence for P4 and governance
Explanation	Can a decision be reconstructed?	Compare contemporaneous logs with generated retrospective explanations	Causal evidence for P5
Institution and audience	Who may consent, refuse, or appeal?	Add a participant veto or collective review	Evidence on HCAI and aesthetic reception

Recomposition must be documented as a new research artefact, not attributed to the original artist. When an artist participates, authorship and permission should be agreed explicitly. When the strategy is reconstructed only analytically, the resulting scenario should be identified as a researcher-created derivative model.

6.4 Experimental speculative protocol

A MEPAA study should proceed through the following eight operations.

1. Define the research object. Specify the artistic strategy, episode, participants, system version, dataset, tools, institution, location, and temporal boundary. Avoid treating “the AI” as a unitary actor.

2. Construct an evidence ledger. For every material claim, record the source type, date, peer-review status, operational definition, and whether the source supports description, mechanism, interpretation, or forecast. Dynamic forecasts receive retrieval dates and archived snapshots.

3. Decompose the protocol. Map inputs, state changes, model calls, human decisions, environmental events, outputs, and stop conditions. Identify where apparent autonomy may be supplied by curation, performance, or interface design.

4. Profile P1–P5. Code evidence as documented, inferred, absent, or unassessable. Do not aggregate indicators. For each positive code, specify an intervention that could falsify the inference—for example, resetting memory, blocking a broadcast channel, altering a state model, inducing salience conflict, or comparing an explanation with logged causal traces.

5. Select a regime and vary capabilities. Use R1–R3 to specify capacities and governance conditions, then modify one or two variables at a time. The goal is not to “upgrade” the artwork but to test how its aesthetic relations change.

6. Build at least three conditional scenarios. Each scenario should state drivers, assumptions, actors, power distribution, resource requirements, desired and undesired consequences, early indicators, and disconfirming observations. At minimum include: (a) a continuity scenario, (b) a conflict or failure scenario, and (c) a transformation or refusal scenario.

7. Stress-test boundary conditions. Introduce modality loss, memory corruption, adversarial input, conflicting relevance signals, strategic self-presentation, human override, ambiguous consent, unequal access, energy constraints, and an ethical stop. Observe whether the aesthetic process remains coherent and whether responsibility remains assignable.

8. Evaluate and report uncertainty. Report process traces, participant accounts, rival explanations, harms, and unresolved uncertainty. Separate findings about behaviour from inferences about architecture and from claims about phenomenality. Publish sufficient provenance for critique without exposing personal data or enabling harm.

6.5 Agent-3 as a scenario, not a forecast

Within the broader doctoral project, Agent-3 can be retained as a named scenario for an advanced but non-AGI artistic ecology. It denotes a configuration in which systems may combine persistent task state, multimodal input, tool use, extended planning, high coding capacity, and coordination with other agents, while remaining dependent on human and institutional governance. The label does not certify a third universal stage of AI development and does not imply consciousness.

Any date previously associated with Agent-3, including September 2027, must be treated as an archival scenario parameter rather than a research finding. The scenario remains analytically useful if its capacities, dependencies, and failure conditions are stated. It should be retired or revised if those conditions cease to discriminate it from R1 or R2 systems.

For artistic research, Agent-3 focuses attention on a plausible qualitative change: the system may begin to modify parts of the production protocol, not merely outputs within a fixed protocol. MEPAA should then test who approves such modifications, whether their causal consequences remain inspectable, and whether an artist or participant can refuse the revised goal.

6.6 Example hypotheses

The revised framework supports falsifiable hypotheses at the protocol level:

• H1—Temporal continuity: If a system’s persistent state is causally necessary to an aesthetic trajectory, resetting that state should produce a measurable disruption not explained by audience expectation or display continuity.

• H2—Distributed relevance: When human and system salience signals conflict, the final trajectory will reflect the actor with effective control over selection rules, revealing where agency is institutionally located.

• H3—Auditability: Explanations linked to contemporaneous logs and intervention results will allow more accurate reconstruction of decisions than ungrounded natural-language self-reports.

• H4—Anthropomorphic framing: Consciousness-oriented labels such as “dreaming” or “remembering” will increase attributed agency even when the underlying technical protocol is unchanged.

• H5—Governance as aesthetics: Adding meaningful refusal, appeal, and override mechanisms will alter participants‘ affective and interpretive experience, not merely the safety profile of the work.

These hypotheses concern observable relations. None treats P1–P5 as a proof of consciousness.

6.7 Prospective empirical phase

The next phase of MEPAA should be organised as a qualitative experimental study rather than presented as already completed. A pilot can recruit approximately 12–20 adult participants with relevant experience in art, design, creative coding, digital culture, or AI-supported production. The final sample should be justified by information power and the diversity of cases rather than by claims of population representativeness. If quantitative hypothesis testing is added, a separate power analysis and preregistered outcome measures will be required.

Participants would work with one documented recomposed protocol across three conditions:

1. Continuity condition: the system operates within declared capabilities and stable governance;

2. Conflict/failure condition: memory, modality, or salience is disrupted, or human and system priorities conflict; and

3. Refusal/transformation condition: participants can stop, appeal, reassign authority, or redesign a rule.

The order of conditions should be counterbalanced where carry-over effects are likely. The system version, prompts, tools, datasets, temperature or stochastic settings, and human interventions must be recorded. Where commercial systems prevent reproducibility, the report should state this limitation and preserve the fullest legally permissible trace.

Data collection should combine interaction logs, screen and event recordings, versioned outputs, field observation, brief post-episode ratings, and semi-structured stimulated-recall interviews. Concurrent think-aloud should be used selectively because it can alter the aesthetic process. Participant reports document experience and attribution; they do not reveal model states. Model-generated explanations document system output; they do not establish introspection.

Analysis should integrate four layers. Directed qualitative coding applies P1–P5 and the D/I/S/U status. Interaction-sequence analysis reconstructs how inputs, state changes, selections, overrides, and outputs unfold. Reflexive thematic analysis examines participants‘ accounts of agency, control, surprise, dependence, and authorship. Descriptive statistics may summarise ratings, but small-sample exploratory values should not be presented as confirmatory evidence. Qualitative and technical traces are compared for convergence and contradiction.

An expert-elicitation component may be added after the pilot. A small Delphi panel of artists, AI researchers, HCI specialists, consciousness researchers, curators, and governance experts would assess conditions of realisability rather than predict an AGI date. Horizon scanning would identify capability, regulatory, infrastructural, and cultural changes that require scenario revision. Anticipation remains plural and updateable (Inayatullah, 2008; Miller, 2018; Poli, 2010).

6.8 Ethics and research integrity

Ethical review is required before participant recruitment. Consent materials must distinguish research observation from public exhibition and specify data retention, recording, withdrawal, publication, and the role of automated systems. Special care is required when works involve faces, voices, homes, health-related signals, cultural archives, or communities whose data may be reused beyond the original context.

The study should avoid deceptive anthropomorphism. Participants may examine how consciousness-oriented framing affects attribution, but any experimental manipulation of labels must be disclosed during debriefing and justified by minimal risk. No system should be described as conscious on the basis of P1–P5. Unexpected claims of distress, dependency, or moral status should be recorded and handled under a predefined escalation protocol rather than improvised for dramatic effect.

Research integrity also requires model and prompt versioning, disclosure of researcher intervention, reporting of negative results, and separation of exploratory from confirmatory analysis. Ecological reporting should include the available information on compute, hardware, duration, data transfer, and display infrastructure. An aesthetic benefit does not cancel a material cost; both belong to the same relational process.

6.9 Planned outputs and criteria of success

The empirical phase should produce: a documented protocol map; episode-level P1–P5 profiles; an evidence ledger; scenario comparisons; a record of failures and overrides; participant and expert interpretations; and a set of revised conditions of realisability. Success is not defined by producing a more autonomous or impressive artwork. It is defined by whether the study makes causal organisation, uncertainty, aesthetic transformation, and responsibility more intelligible.

The principal criterion for theoretical success is discriminatory value: MEPAA should permit researchers to distinguish two protocols or scenarios that ordinary descriptions would conflate. The principal criterion for empirical success is traceability: another researcher should be able to understand how a conclusion follows from documented events, even when the work itself cannot be exactly reproduced. The principal ethical criterion is contestability: affected participants must have meaningful ways to question, refuse, or stop consequential operations.

7. Discussion

The revised MEPAA changes the status of anticipation in three ways. Epistemically, it rejects a single AGI chronology and uses forecasts only to map uncertainty. Ontologically, it distinguishes computational organisation from biology and functional agency from phenomenality. Politically, it makes ownership, labour, ecological cost, consent, and contestability constitutive of the aesthetic process rather than external ethical additions.

7.1 Artistic knowledge and the limits of technological description

These corrections strengthen rather than diminish the role of artistic research. A technical benchmark can measure performance under specified conditions, but it cannot by itself show how delegated judgement feels, how a public negotiates uncertainty, or how an institution aestheticises extraction. Artistic strategies can construct those situations. Their knowledge is experiential, situated, and materially articulated; it becomes academically robust when the protocol and inference are open to criticism (Borgdorff, 2012; Rheinberger, 2023).

The analysis also clarifies the value of “a medium that thinks”. Manovich (2026) uses the phrase to mark a significant change in media cognition: generative media respond, recombine, and participate in cultural production. The phrase is analytically useful if “thinks” names operational participation in a distributed process. It becomes misleading if it is read as established evidence of a unified subject or conscious medium. MEPAA keeps the productive metaphor while demanding mechanistic and phenomenological restraint.

Agüera y Arcas’s predictive account and Rovelli’s relational account contribute at different scales. The first encourages experiments in reciprocal prediction, error, and adaptation; the second requires that the observer, information set, and time of assessment be included in any claim about the future. MEPAA joins them through reflexive anticipation: it studies how systems and people predict within an artwork while also examining how forecasts about AI shape funding, design, institutional behaviour, and audience expectation. Forecasts are not neutral descriptions; they can become performative components of the system they purport to describe.

7.2 Answer to RQ3: agency without premature consciousness claims

MEPAA can engage with artificial-consciousness research only through disciplined translation. P1–P5 direct attention to temporal continuity, information propagation, state modelling, relevance, and reportability because these functions can organise an aesthetic process. They do not establish that the process is experienced by the system. The same architecture can be relevant to creativity, control, or interaction without settling the metaphysics of consciousness.

This position avoids two symmetrical errors. The first is inflation: treating fluent output, a face, or a first-person statement as evidence of a machine subject. The second is dismissal: assuming that because consciousness is unverified, functional changes in agency have no aesthetic or moral significance. A system need not be conscious to classify a participant, redirect attention, generate dependency, consume resources, or modify a cultural archive. Governance is therefore justified by consequential agency, not only by moral status.

The Agentic Entity is most useful at this intermediate level. It identifies where causal initiative appears, how it is constrained, and how it is attributed. If later research supplies stronger evidence concerning artificial consciousness, the framework can update its interpretation. It does not need to presuppose that conclusion in order to study present transformations.

7.3 Answer to RQ4: from future image to testable relation

An existing artistic strategy can be recomposed into a testable future scenario when four requirements are met. The claimed future relation must be operationally defined; its necessary capabilities and governance conditions must be stated; at least one causal component must be open to intervention; and the results must be reported with rival explanations. Without these requirements, scenario construction remains valuable as cultural imagination but does not yet constitute the experimental layer of MEPAA.

Boundary conditions are especially important. A protocol that works only when memory is accurate, participants comply, and model and human priorities coincide reveals little about future robustness. Failure and resistance expose whether continuity is genuinely system-dependent, whether relevance can be negotiated, and whether human authority remains meaningful. The refusal scenario is therefore not a pessimistic supplement. It is an epistemic test of where power resides.

7.4 Anticipated transformations of aesthetic processes

The study supports five conditional transformations rather than predictions of inevitable outcomes.

First, the artwork may shift further from object to adaptive trajectory. Its identity would depend on the persistence and governance of change rather than on a fixed perceptual form. Second, authorship may shift from producing outputs to designing and contesting protocols, including the allocation of memory, goals, tools, and stop conditions. Third, curatorship may become a form of attention governance: curators and communities would regulate data entry, salience, model versions, and acceptable forms of adaptation.

Fourth, spectators may increasingly become participants, data subjects, co-trainers, and counter-agents. Their aesthetic experience will depend not only on what a system displays but on what it infers and permits them to control. Fifth, the institution may become part of the artwork’s functional architecture. Procurement, access to compute, data agreements, safety review, conservation, and public accountability would shape the aesthetic process as directly as visual and sonic form.

These transformations are already partially visible in the cases, but their future extent is uncertain. MEPAA does not infer a smooth progression towards more machine autonomy. Some of the most significant artistic futures may emerge from limiting automation, designing collective control, preserving opacity for vulnerable participants, or refusing datasets and model infrastructures that cannot meet cultural or ecological conditions.

7.5 Implications for the aesthetics of AI art

The broader implication is that an aesthetics of AI art requires more than a theory of machine-made objects. It requires an account of how perceptual form, computational operation, human attribution, and institutional power become temporarily stabilised as one aesthetic event. This does not dissolve formal analysis; it situates form within the process that selects and sustains it.

MEPAA contributes a middle-range framework for that task. It is narrower than a total philosophy of AI and more expansive than a technical model audit. Its proper object is the experimental aesthetic relation: a bounded but revisable configuration in which humans and computational systems predict, select, explain, resist, and redistribute agency. Its scientific credibility will depend on whether future studies can use the framework to produce discriminating, traceable, and ethically contestable findings.

8. Limitations

The four cases are purposively selected and cannot represent the full field of AI art. Most evidence comes from public documentation rather than direct access to source code, logs, datasets, curatorial negotiations, participants, or compute accounting. P1–P5 are an authorial heuristic and have not been psychometrically or experimentally validated. Their theoretical proximity to consciousness research creates a continuing risk of anthropomorphism even when caveats are explicit.

The literature component is a critical, source-verified synthesis rather than a systematic review. It does not claim exhaustive database coverage, a preregistered search strategy, or formal risk-of-bias scoring. The selected cases were chosen for theoretical relevance and documentary accessibility, which may overrepresent internationally visible, institutionally supported practices. Future research should include less visible, non-Western, community-led, and deliberately low-compute work, as well as negative cases in which an apparently agentic process is fully explained by automation theatre or hidden human labour.

The forecast snapshot is valid only for 20 August 2026 and depends on platform-specific rules. The expert survey records beliefs rather than technological facts. The use of Rovelli is an explicitly limited methodological analogy; quantum theory is not a model of AI development or cultural history. Agüera y Arcas’s general theory of intelligence is likewise interpreted as a theoretical programme, not consensus. Finally, this article proposes a protocol but does not yet report its prospective application with participants. Such an application would require ethics review, preregistered hypotheses where appropriate, data governance, and domain-specific technical collaboration.

Rapid model and platform change will complicate replication. Even when prompts are preserved, provider-side updates, unavailable weights, safety-policy changes, and stochastic sampling may alter results. MEPAA can mitigate but not eliminate this problem through versioning, archived outputs, local models where appropriate, and reporting of dependencies. The framework itself should be revised when empirical use shows that an indicator fails to discriminate relevant processes or produces systematic anthropomorphic bias.

9. Conclusion

Future aesthetic processes in AI art should not be organised around a countdown to AGI. The more defensible research object is the changing relation among biological participants, computational systems, and sociomaterial environments under uneven capability, uncertainty, and power. HCAI supplies normative criteria for these relations; it is not an evolutionary stage destined to disappear when systems become more general.

The four core cases show that artistic research already redistributes generation, selection, attention, observation, and decision authority. The strongest present evidence concerns temporal interaction and relevance selection; evidence for global availability is protocol-dependent; metarepresentation is frequently over-attributed; and auditable reportability remains the principal limitation. These conclusions concern functional organisation and reception. They do not demonstrate phenomenal consciousness.

MEPAA turns artistic research into a disciplined form of anticipation. It profiles functional organisation without diagnosing consciousness, distinguishes internal prediction from external forecasting, and replaces fixed dates with capability-and-governance regimes. Its scenarios are valuable when they expose assumptions, distribute voice, support intervention, and remain answerable to empirical evidence. The decisive question is therefore not simply whether AI will become an autonomous or conscious artist. It is which aesthetic ecologies are being assembled now, who can shape or refuse them, what evidence justifies their descriptions, and what futures those descriptions help to make possible.

As a full research programme, MEPAA now contains a completed conceptual layer and a specified but not yet completed empirical layer. Its next scientific test is prospective: whether decomposition, recomposition, boundary scenarios, and P1–P5 coding can generate traceable differences among artistic protocols and improve the intelligibility and contestability of future aesthetic processes.

Declarations

Ethics statement. The present article reports critical document analysis and conceptual-methodological development; it does not report research with human or animal participants. The prospective empirical phase described in Section 6.7 must not begin before approval by the relevant research ethics body and completion of consent and data-management procedures.

Data availability. No new empirical dataset was generated for the present article. The analysed evidence consists of the scholarly, institutional, and primary project sources listed in the reference section. Dynamic forecast values are explicitly dated.

Funding and competing interests. These declarations must be completed by the author in accordance with the requirements of the target journal before submission.

Generative-AI disclosure. Any use of generative AI for language editing, research organisation, or manuscript development should be disclosed in accordance with the target journal’s policy. Responsibility for source verification, argument, and the submitted text remains with the human author; an AI system must not be listed as an author.

References

Agüera y Arcas, B. (2025). What is intelligence? Lessons from AI about evolution, computing, and minds. MIT Press.

Anadol, R. (2020). Synaesthetic architecture: A building dreams. Architectural Design, 90(3), 76–85. https://doi.org/10.1002/ad.2572

Anadol, R. (2022). Space in the mind of a machine: Immersive narratives. Architectural Design, 92(3), 28–37. https://doi.org/10.1002/ad.2810

Auger, J. (2013). Speculative design: Crafting the speculation. Digital Creativity, 24(1), 11–35. https://doi.org/10.1080/14626268.2013.767276

Barad, K. (2007). Meeting the universe halfway: Quantum physics and the entanglement of matter and meaning. Duke University Press.

Barendregt, L., & Vaage, N. S. (2021). Speculative design as thought experiment. She Ji: The Journal of Design, Economics, and Innovation, 7(3), 374–402. https://doi.org/10.1016/j.sheji.2021.06.001

Bengio, Y., Clare, S., Prunkl, C., Murray, M., Andriushchenko, M., Bucknall, B., Bommasani, R., Casper, S., Davidson, T., Douglas, R., Duvenaud, D., Fox, P., Gohar, U., Hadshar, R., Ho, A., Hu, T., Jones, C., Kapoor, S., Kasirzadeh, A., . . . Mindermann, S. (2026). International AI Safety Report 2026 (DSIT 2026/001). Department for Science, Innovation and Technology. https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026

Biggs, M. A. R., & Karlsson, H. (Eds.). (2012). The Routledge companion to research in the arts. Routledge.

Boden, M. A., & Edmonds, E. A. (2009). What is generative art? Digital Creativity, 20(1–2), 21–46. https://doi.org/10.1080/14626260902867915

Borgdorff, H. (2012). The conflict of the faculties: Perspectives on artistic research and academia. Leiden University Press.

Butlin, P., & Lappas, T. (2025). Principles for responsible AI consciousness research. Journal of Artificial Intelligence Research, 82, 1673–1690. https://doi.org/10.1613/jair.1.17310

Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. https://doi.org/10.1016/j.tics.2025.10.011

Capel, T., & Brereton, M. (2023). What is human-centered about human-centered AI? A map of the research landscape. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Article 359, pp. 1–23). Association for Computing Machinery. https://doi.org/10.1145/3544548.3580959

Chatterjee, A. (2022). Art in an age of artificial intelligence. Frontiers in Psychology, 13, Article 1024449. https://doi.org/10.3389/fpsyg.2022.1024449

Cheng, I. (2015–2017). Emissaries [Live simulation trilogy]. https://iancheng.com/emissaries

Chung, S. (2026a). RECURSIONS at Hong Kong. https://sougwen.com/2026/recursions-hong-kong

Chung, S. (2026b). Research: Rewilding the Machine. https://sougwen.com/research

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204. https://doi.org/10.1017/S0140525X12000477

Coeckelbergh, M. (2023). The work of art in the age of AI image generation: Aesthetics and human–technology relations as process and performance. Journal of Human-Technology Relations, 1. https://doi.org/10.59490/jhtr.2023.1.7025

Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U., Henin, S., Hirschhorn, R., Khalaf, A., Lepauvre, A., Liu, L., Richter, D., Vidal, Y., Bonacchi, N., Brown, T., Sripad, P., Armendariz, M., Bendtz, K., Ghafari, T., Hetenyi, D., Jeschke, J., Kozma, C., . . . Melloni, L. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. https://doi.org/10.1038/s41586-025-08888-1

Costanza-Chock, S. (2020). Design justice: Community-led practices to build the worlds we need. MIT Press.

Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press.

Crawford, K., & Paglen, T. (2021). Excavating AI: The politics of images in machine learning training sets. AI & Society, 36, 1105–1116. https://doi.org/10.1007/s00146-021-01162-8

Crespo, S. (2018–2022). Neural Zoo. https://sofiacrespo.com/neural-zoo/

Crespo, S. (2020–2024). Artificial Natural History. https://sofiacrespo.com/artificial-natural-history/

Crespo, S. (2021–2022). Critically Extant. https://sofiacrespo.com/critically-extant/

Crespo, S. (2024). Perpetual Present. https://sofiacrespo.com/perpetual-present/

DATALAND. (n.d.). Large Nature Model. https://dataland.art/about/large-nature-model

DATALAND. (2026). Machine Dreams: Rainforest. https://dataland.art/exhibitions/machine-dreams-rainforest

del Campo, M. (2024). Art beyond mechanical reproduction: In conversation with AI artist Mario Klingemann. Architectural Design, 94(3), 62–69. https://doi.org/10.1002/ad.3056

Di Biagio, A., Donà, P., & Rovelli, C. (2021). The arrow of time in operational formulations of quantum theory. Quantum, 5, Article 520. https://doi.org/10.22331/q-2021-08-09-520

Dunne, A., & Raby, F. (2013). Speculative everything: Design, fiction, and social dreaming. MIT Press.

Evans, J., Bratton, B. H., & Agüera y Arcas, B. (2026). Agentic AI and the next intelligence explosion. Science, 391(6791), Article eaeg1895. https://doi.org/10.1126/science.aeg1895

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Futurology & Berggruen Institute. (2026, July 7). We merged with machines a long time ago | Google’s Blaise Agüera y Arcas & physicist Carlo Rovelli [Video]. YouTube. https://www.youtube.com/watch?v=tDccIoz-SFI

Gebru, T., & Torres, É. P. (2024). The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence. First Monday, 29(4). https://doi.org/10.5210/fm.v29i4.13636

Grace, K., Sandkühler, J. F., Stewart, H., Weinstein-Raun, B., Thomas, S., Stein-Perlman, Z., Salvatier, J., Brauner, J., & Korzekwa, R. C. (2025). Thousands of AI authors on the future of AI. Journal of Artificial Intelligence Research, 84, Article 9. https://doi.org/10.1613/jair.1.19087

Herndon, H., & Dryhurst, M. (2024). The Call: New rituals for collaboration with AI [Exhibition and interactive sound installation]. Serpentine Galleries. https://www.serpentinegalleries.org/whats-on/holly-herndon-mat-dryhurst-the-call/

Huyghe, P. (2024a). Camata [Autonomous film environment using machine learning, robotics, and sensors]. Pinault Collection. https://www.pinaultcollection.com/en/boursedecommerce/pierre-huyghe-1

Huyghe, P. (2024b). Liminal [Exhibition]. Pinault Collection. https://www.pinaultcollection.com/palazzograssi/en/pierre-huyghe-liminal

Inayatullah, S. (2008). Six pillars: Futures thinking for transforming. Foresight, 10(1), 4–21. https://doi.org/10.1108/14636680810855991

Kalshi. (n.d.). When will OpenAI achieve AGI? Retrieved August 20, 2026, from https://kalshi.com/markets/kxoaiagi/openai-achieves-agi/oaiagi

Klingemann, M. (2018, October 28). Neural Glitch/Mistaken Identity. Quasimondo. https://quasimondo.com/2018/10/28/neural-glitch/

Klingemann, M. (2019). Circuit training [Interactive installation]. Barbican Centre. https://artsandculture.google.com/story/circuit-training-machine-made-art-for-the-people-barbican-centre/ngWRdP9M5scyLQ?hl=en

Klingemann, M. (2023). A.I.C.C.A. [Robotic performance sculpture]. https://aicca.me/about/

Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dean, R. (2025). AI 2027. https://ai-2027.com/

Latour, B. (2005). Reassembling the social: An introduction to actor-network-theory. Oxford University Press.

Lifland, E., Kokotajlo, D., & Halstead, B. (2026, January 27). Clarifying how our AI timelines forecasts have changed since AI 2027. AI Futures Project. https://blog.aifutures.org/p/clarifying-how-our-ai-timelines-forecasts

Manovich, L. (2019). AI aesthetics. Strelka Press. https://manovich.net/index.php/projects/ai-aesthetics

Manovich, L. (2026). A medium that thinks: Generative AI and media cognition. Emerging Media, 4(2), 225–232. https://doi.org/10.1177/27523543261454458

Manovich, L., & Arielli, E. (2024). Artificial aesthetics: Generative AI, art and visual media. https://manovich.net/index.php/projects/artificial-aesthetics

McCarthy, L. L. (2016). Follower [Participatory performance]. https://get-lauren.net/Follower

McCarthy, L. L. (2017). LAUREN [Networked performance and installation]. https://get-lauren.com/

McCarthy, L. L. (2019). SOMEONE [Distributed performance and installation]. https://get-lauren.net/SOMEONE

McCarthy, L. L. (2023). Voice in my head. https://lauren-mccarthy.com/Voice-In-My-Head

McCarthy, L. L. (2025). AUTO [Participatory performance]. https://get-lauren.net/Auto

McCarthy, L. L., McDonald, K., & Yamaguchi Center for Arts and Media. (2022). Unlearning language [Interactive installation and performance]. https://www.ycam.jp/en/events/2022/unlearning-language/

Metaculus. (n.d.). When will the first general AI system be devised, tested, and publicly announced? Retrieved August 20, 2026, from https://www.metaculus.com/questions/5121/date-of-artificial-general-intelligence/

Miller, R. (Ed.). (2018). Transforming the future: Anticipation in the 21st century. Routledge.

Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). Position: Levels of AGI for operationalizing progress on the path to AGI. Proceedings of Machine Learning Research, 235, 36308–36321. https://proceedings.mlr.press/v235/morris24b.html

Museum of Modern Art. (2022). Refik Anadol: Unsupervised. https://www.moma.org/calendar/exhibitions/5535

Onkaos. (n.d.-a). Appropriate Response. https://onkaos.com/work/appropriate-response/

Onkaos. (n.d.-b). Botto. https://onkaos.com/work/botto/

Onkaos. (n.d.-c). Neural Decay series. https://onkaos.com/work/neural-decay-series/

Pennartz, C. M. A. (2026). How can we validate theory-derived indicators of consciousness in artificial intelligence? Trends in Cognitive Sciences, 30(7), 573–574. https://doi.org/10.1016/j.tics.2026.01.011

Poli, R. (2010). An introduction to the ontology of anticipation. Futures, 42(7), 769–776. https://doi.org/10.1016/j.futures.2010.04.028

Polymarket. (n.d.). OpenAI announces it has achieved AGI before 2027? Retrieved August 20, 2026, from https://polymarket.com/event/openai-announces-it-has-achieved-agi-before-2027

Possible Futures. (2023). Refik Anadol × Yawanawá: Winds of Yawanawá. https://www.possiblefutures.one/nft

Rheinberger, H.-J. (2023). Split and splice: A phenomenology of experimentation. University of Chicago Press.

Rovelli, C. (2018). The order of time (E. Segre & S. Carnell, Trans.). Riverhead Books.

Schwab, M. (2018). Transpositionality and artistic research. In M. Schwab (Ed.), Transpositions: Aesthetico-epistemic operators in artistic research (pp. 191–214). Leuven University Press.

Sedbon, M. (n.d.). Living computations. https://michaelsedbon.com/Living-Computations

Sedbon, M. (2024). Bio economics [Biohybrid installation]. Hybrid Forms Lab. https://hybridformslab.com/projects-ongoing/bio-economics/

Serpentine Galleries. (2018). Ian Cheng: BOB. https://www.serpentinegalleries.org/whats-on/ian-cheng-bob/

Shneiderman, B. (2022). Human-centered AI. Oxford University Press. https://doi.org/10.1093/oso/9780192845290.001.0001

Sommerer, C., & Mignonneau, L. (2026). Flâneur. https://interface.ufg.ac.at/christa-laurent/Flaneur.html

Taylor, G. D. (2014). When the machine made art: The troubled history of computer art. Bloomsbury Academic.

Zeilinger, M., & Johns, D. (2025). AI art and the foreclosure of ecological agency: A critique of Refik Anadol’s Echoes of the Earth. In J. Markelj & C. C. Bueno (Eds.), Vectoral agents: Power in the age of planetary computation (pp. 21–31). Institute of Network Cultures. https://networkcultures.org/wp-content/uploads/2025/11/Network-Notions-4.pdf

Zylinska, J. (2020). AI art: Machine visions and warped dreams. Open Humanities Press. https://www.openhumanitiespress.org/books/titles/ai-art/

Zylinska, J. (2025). Bio-AI: The aesthetics and ethics of data animism. Culture Machine, 24. https://culturemachine.net/vol-24-aesthetics-biomachine/bio-ai-the-aesthetics-and-ethics-of-data-animism/

Appendix A. Citation and Claim Audit

Earlier element	Verification result	Revision made
AI Impacts (2024) preprint used as survey authority	Superseded by a peer-reviewed article	Replaced with Grace et al. (2025), JAIR, DOI 10.1613/jair.1.19087
Butlin and Lappas (2025) cited as arXiv preprint	Peer-reviewed version published in JAIR	Updated journal, volume, pages, and DOI
Butlin et al. (2023) treated as the current indicator framework	Substantially updated and peer reviewed	Replaced for principal claims by Butlin et al. (2026), Trends in Cognitive Sciences
P1–P5 presented near the language of consciousness indicators	No validation establishes this exact five-item scheme as a diagnostic scale	Explicitly designated an authorial heuristic profile; no sum score or consciousness inference
HCAI described as a transitional stage towards AGI	HCAI literature defines a normative/design orientation, not an AGI stage	Reframed HCAI as persistent evaluative commitments
AGI by 2029	No verified source supports this as a robust scientific conclusion	Removed; forecasts are definition-dependent distributions and dated snapshots
Metaculus question 3479 and February 2028 estimate	Link and estimate were obsolete	Corrected to question 5121; snapshot updated to March 2033 on 20 August 2026
Prediction platforms combined as if measuring one event	Current contracts have different resolution criteria	Added a comparison table and prohibited direct aggregation
“H1–H5” reference in a three-horizon model	Internal inconsistency	Replaced H1–H3 dates with R1–R3 capability-and-governance regimes
“Human and AI biologies”	Treats computational systems as biological without scientific warrant	Replaced with coupled biological and computational modes of organisation
Artist and museum statements used as evidence of mechanisms	Such sources document projects and declared intentions, not hidden causal architecture	Added source hierarchy, rival explanations, and auditability requirements
Metaphors such as machine memory, dreaming, or hallucination	Metaphors do not establish metarepresentation or consciousness	Retained only as aesthetic framing and marked their evidential limits
Bazin (1967) used to support technological prefiguration	The citation is poorly matched to that empirical-historical claim	Removed from the argument
Chen (2024) used as a major authority for speculative design	Source adds limited methodological authority relative to established scholarship	Replaced by Auger (2013), Dunne and Raby (2013), Barendregt and Vaage (2021), and Miller (2018)
Agüera y Arcas used to connect prediction with AGI timing	His 2025 argument concerns prediction as a function of intelligence	Separated computational prediction from sociotechnical forecasting
Rovelli implicitly treated as an AI prediction theorist	Rovelli’s primary research concerns time, physics, and relational information	Used only as an explicitly limited epistemological corrective; no AI forecast attributed to him

Appendix B. Status of Evidence

This article is a critically revised conceptual study prepared from the supplied research proposal and the updated AI Art Strategies 2026 source corpus. Bibliographic metadata for the principal scientific claims was checked against publisher, journal, conference, or official-report records available on 20 August 2026. Dynamic platform values require rechecking at every future submission. Before journal submission, the four case profiles should be supplemented by archived project documentation, interviews where feasible, a transparent coding appendix, and a prospective application of the MEPAA protocol.