The Lifetime of a Digital Artwork in the AI era

A significant and growing class of digital artwork — computational, generative, model-dependent, networked — is less a file than a process: code that executes, a model that infers, a runtime that must be assembled, a room that must be measured. The United States Copyright Office's second report on copyright and artificial intelligence, published in January 2025, treats AI-assisted works as copyrightable where the human contribution is sufficient, and finds prompts alone generally insufficient to confer authorship.^[1] That judgement concerns the moment of making. It does not speak to the decades that follow.

This study follows the whole lifetime of one such work — a generative, code-based, model-dependent piece realized differently in every room — through making, rights, commissioning, transfer, realization, exhibition memory, preservation and re-exhibition. At each stage it asks the same seven questions: what exists today, who holds the data or the decision, what travels, what is lost, who reconstructs it, whether AI can use it, and what remains unresolved.

The finding is not a shortage of documentation. Museums document time-based media in detail;^[3] provenance standards can cryptographically bind assertions to a file;^[4] preservation networks have formalized emulation, migration and reinterpretation as distinct strategies;^[5] platforms already move works and time-limited loans between accounts.^[7] Interpretation These records live in separate custody, are produced for different purposes, and — on the evidence reviewed here — no widely adopted art-world practice binds them into one persistent artwork-level context that travels with the work. The work moves. Its knowledge moves less reliably.

The study proposes no mechanism. It ends where its evidence ends: with an infrastructure question.

1. AI has entered the making of art. What happens when it enters the rest of an artwork's life?
2. Where does a computational artwork's lifetime remain continuous today, and where does it fragment?
3. What actually travels from one context to another — file, code, model, rights, intent, technical knowledge, provenance, curatorial context, display decisions, exhibition history?
4. What is lost at each handoff, and is the loss recorded anywhere?
5. Can an exhibition be understood as one manifestation in a continuing lifetime rather than an endpoint?
6. Can exhibition memory become part of the artwork's accumulating knowledge?
7. Can that accumulated lifetime become a sustainable digital knowledge asset?
8. If artworks become more adaptive and relational, what would allow their lives to remain whole?

Each lifecycle stage in Figure 01 is interrogated with the same fixed prompts, so that stages can be compared rather than narrated: what exists today · who holds it · what travels · what is lost · who reconstructs it · can AI use it · unresolved.

Citations point to primary, institutional, standards, legal or peer-reviewed sources at the exact page consulted; references carry title, issuing body, date where stated, and full URL. Where a primary source could not be verified — several regional and artist-studio cases among them — the gap is declared rather than filled; see Declared gaps. No quantitative claim appears here without a cited measurement, and none of the four figures reports measured data: each is a reasoning instrument and says so in its evidence note.

A file has an edge. A process does not. When a work is generated at runtime, its identity is distributed across at least six things: source code, model weights, training or reference data, a runtime environment with versioned dependencies, live input (sensors, feeds, visitors), and a display surface with its own physics. Change any one and the output changes. None of the six is the artwork; the artwork is the relation between them.

Preservation practice reached this position before the market did. The Variable Media Network's framework asks artists to declare, in advance, how a work should behave when its medium fails — and names storage, emulation, migration and reinterpretation as distinct, non-equivalent futures.^[5] Rhizome's ArtBase, an archive dedicated to born-digital and internet-based art, exists because works whose environment has disappeared cannot simply be re-opened.^[6]

AI-native work sharpens this. A model is a dependency that is itself authored elsewhere, versioned elsewhere, deprecated elsewhere. Interpretation For a work that keeps deciding, there is no single final state to archive; the question moves from what is the file to what is the behaviour, and who is entitled to define it.

On authorship the legal position is narrower than the practice. The Copyright Office's Part 2 report concludes that existing law is adequate to handle AI: human authorship remains the requirement, AI-assisted works can be protected where a human contributes sufficient expressive input, and prompting alone generally does not meet that bar.^[1]

Interpretation This is a test applied at one instant. An artwork that keeps generating produces new outputs long after that instant, in venues the author never saw, under conditions the author did not set. The law answers who authored it. The lifetime keeps asking who is answerable for it now — and this study found no single, agreed addressee for that second question.

Authorship, copyright, edition ownership, token ownership, exhibition licence, adaptation rights, commercial rights and provenance are eight separate claims. They are held by different parties, created by different instruments, and verified by different means. Interpretation Conflating them is a structural failure rather than a careless one: no one record in general use is designed to carry all eight.

A blockchain record can establish that a specific key signed a specific transaction at a specific time, and can order those transactions durably. The joint Copyright Office and Patent and Trademark Office study of NFTs and intellectual property examined this intersection, and the distinction it works through is the one the market blurs — acquiring a token is not, by default, acquiring the copyright in the underlying work.^[2]

Content provenance standards answer a different question again. C2PA binds signed assertions about creation and editing to the asset itself; validation establishes the integrity of the manifest and the identity of the signer — not the truth of what is asserted.^[4] Museum documentation answers a fourth question: not who owns or who signed, but how the thing is to be kept alive.^[3]

Open question Each of the four systems is well specified for its own purpose. What this study could not find is a widely adopted art-world practice binding them into one persistent artwork-level context: an institution's iteration report does not reference the manifest; the manifest does not reference the loan agreement; the chain references neither. Cross-reference is presently a human act.

When a work is a running system rather than an object, the curatorial act changes shape. Peer-reviewed research on curating AI-driven art examines exactly this: the actors involved, the institutional strategies adopted, and the organisational change that follows when institutions take on AI-driven work.^[10] The Serpentine's Future Art Ecosystems briefings were established to build shared strategic vocabulary for art and advanced technologies, with the fourth edition addressing art and public AI directly.^[8][9]

National institutions are staging the subject as well as collecting it: the MMCA in Korea presented What an Artificial World, an exhibition framed around artificiality and machine systems.^[12] In Japan, teamLab Borderless presents works that leave their rooms, interrelate and influence one another — a body of work whose identity is inseparable from a specific built environment.^[11]

Interpretation Commissioning of this kind produces two outputs: the artwork, and a body of technical and relational knowledge generated in making it work. On the evidence reviewed here, only the first is routinely treated as having a lifetime; the second is generally budgeted as project overhead and filed at the close of the show.

Moving a digital work between cities is no longer the hard part. Platforms operationalize transfers and time-limited loans of digital artworks between accounts, with defined terms and durations.^[7] The wider cultural sector has built interoperable ways to deliver and describe media across institutional boundaries.^[14] Graph data models exist for expressing relationships between resources in a form both machines and institutions can exchange.^[13]

What this study could not find in general use is a standardized realization layer: a machine-readable statement of how the work must be played — at what luminance, at what black level, in what aspect, for how long, with what sound, with what sensing, with what fallback when a dependency is unavailable — and which of those values the artist has fixed and which the venue may set. Institutional practice records much of this, but as installation instructions written for people.^[3]

Hypothesis Works of this class can be moved between venues with comparatively little friction; their knowledge moves less easily — not because it is secret, but because there is no shared form in which to send it.

A work specified for a calibrated OLED at close viewing distance is not the same encounter on a media façade seen across a plaza in daylight. Scale, peak luminance, black level, colour volume, crop, runtime, sound reproduction, sensor placement, viewing distance and ambient light are not delivery details; they are the terms under which the work exists in that room.

Institutional practice already treats these as recordable and negotiable rather than incidental: guidance on acquiring, exhibiting and sustaining time-based media centres on installation instructions and documented iterations precisely because each installation differs.^[3] Variable-media frameworks go further and ask the artist to pre-authorize the kinds of change that may be made when the original conditions are unavailable.^[5]

Open question Both mechanisms depend on the artist being reachable, or on a prior declaration being complete. Neither obviously scales to a work exhibited many times, in many venues, across decades, under conditions no one anticipated. The unresolved question is not may we adapt — it is who is entitled to decide, and where that decision is written down so the next venue inherits it.

An exhibition is normally recorded as an event: title, venue, dates, installation photograph. But a single install of a code-based work generates a dense body of specific knowledge — an approved display adaptation, a set of measured calibration values, an installation logic for that architecture, a list of what failed and how it was fixed, decisions the artist made on site, a curatorial reading, an observed pattern of audience behaviour, documentation, and a working relationship between a studio and a technical team.

Museum practice for time-based media already recognizes much of this as documentation worth producing.^[3] Interpretation The difficulty this study identifies is custody and direction rather than awareness: the record is produced by an institution, for that institution, and filed under the exhibition. No widely adopted art-world practice binds it to the work in a form the next venue inherits.

Hypothesis If each exhibition deposited its knowledge with the work rather than only with the venue, the second exhibition could begin where the first ended — and the tenth would be a substantially better-informed act than the first.

Conservation of variable media has long refused the idea that keeping a file is keeping a work. The Variable Media framework names storage, emulation, migration and reinterpretation as separable strategies with different consequences, and asks the artist to indicate which are acceptable.^[5] Archives dedicated to born-digital and internet-based art exist because network-dependent works stop working when their environment does.^[6] Institutional guidance for time-based media treats the artist interview and the iteration record as core conservation instruments rather than extras.^[3]

Interpretation AI-native work adds a dependency class that differs in kind: a model held by a third party, versioned on a commercial schedule, subject to deprecation, and not necessarily licensable in perpetuity by the institution that owns the artwork. Emulating an operating system is a known conservation problem; a substitute model is not expected to behave identically, which makes the question one of acceptable difference rather than technical recovery.^[5]

Open question If a future conservator must reconstruct a work, the decisive material may not be the file but the chain of decisions already made about it — what was changed, by whom, with whose consent, and why it was judged acceptable. That chain is exactly the material Figure 02 rates as least likely to survive a handoff.

AI has moved decisively into making, and the legal system has responded to making.^[1] Interpretation In the later stages, this study found no widely adopted art-world practice applying machine assistance to realization, exhibition memory or preservation — plausibly because those stages lack the structured, permissioned records such assistance would require, rather than because they are unsuited to it. Figure 04 maps eight later-lifetime functions against both the opportunity and the specific risk each carries.

Two risks recur and are worth stating plainly. First, authority: a system that proposes an adaptation is, in effect, making an artistic decision unless a human with standing signs it. Second, dependency: introducing a model into preservation makes that model a dependency of the work, which is the precise failure mode preservation exists to prevent.^[5]

• Hypothesis 01Works of this class may be becoming intelligent faster than their lifetimes are becoming connected.
• Hypothesis 02Such works can travel almost anywhere; their knowledge may not travel as easily.
• Hypothesis 03The present difficulty may not be a lack of documentation, but the fragmentation of context across systems designed for separate purposes.
• Hypothesis 04An exhibition may be better understood as one manifestation in a continuing lifetime than as an endpoint.
• Hypothesis 05Artwork lifetime knowledge may be capable of accumulating into a sustainable digital knowledge asset — with value to artists, institutions and future conservators alike.

• CustodyExhibition knowledge is produced by venues, for venues.^[3] The work has no standing to hold its own record.
• AuthorityAdaptation decisions require an answerable human. Works outlive their artists, and prior declarations cannot anticipate every room.^[5]
• Openness vs. controlA continuous lifetime record is only useful if it is readable across institutions — and only safe if the artist controls what is exposed.
• DependencyAny intelligence added to the lifetime becomes something the lifetime then depends on, including for preservation.
• PrivacyAudience interaction is real evidence of a work's realization and is also personal data. These cannot both be maximised.
• IncentiveThis study found no established funding line for maintaining an artwork's knowledge between exhibitions. Open question

This study's rule is that nothing appears as a factual claim without a source cited to the exact page consulted. The following research lines could not be verified to that standard and are therefore named as open gaps rather than written as evidence:

• UAE / DubaiDigital art programming and immersive display infrastructure — requires first-party programme documentation and technical specifications.
• SingaporeInstitutional media-art collecting and public-screen policy — requires primary institutional sources.
• Japan / KoreaBeyond the two cited exhibition sources [11] [12], conservation and re-exhibition practice for large-scale interactive environments.
• Platforms & specificationsDistribution-platform and display-specification initiatives for digital art beyond the single platform cited [7] — require current first-party specification documents.
• Artist practiceNamed artists working with simulation, agentic and long-duration systems — require first-party studio or gallery statements on realization and preservation terms.
• FrequencyNo claim in this study reports how often any described loss occurs. Establishing that would require a survey of installations, which this study did not conduct.