When the Museum of Modern Art installed Refik Anadol's "Unsupervised" in its lobby in November 2022, the acquisition triggered a round of internal debate that curators had not faced in quite the same form before: how do you evaluate a work whose visual output is generated in real time by a machine learning model trained on the museum's own collection data? The question was not whether the piece was interesting. It was whether existing review frameworks, built for painting, sculpture, video, and even earlier computer-based art, could absorb a new category of submission without being rewritten from scratch.
Museums have handled machine-made imagery before. Harold Cohen's AARON program produced drawings from 1973 onward, and Vera Molnar was writing algorithmic rules for generative compositions in the 1960s. What has changed since roughly 2021 is the volume of submissions using large pretrained models, the speed at which the underlying software changes or disappears, and the legal uncertainty around authorship. This article describes how institutions are adjusting their review criteria, what curators ask for before a work reaches a gallery wall, and where published policy still leaves gaps.
Why Museums Needed New Review Criteria
Traditional acquisition and exhibition review rests on a handful of stable assumptions: the artist can describe their process, the object or file is fixed once finished, and the materials used have a known conservation profile. AI-assisted work breaks at least one of these assumptions in most cases. A diffusion model's output depends on a checkpoint version, a random seed, and a prompt, none of which are self-evident from the image alone, and none of which guarantee that the same result can be reproduced six months later if the model is updated or taken offline.
Several institutions responded by adding a technical review stage ahead of the usual curatorial and conservation review. The Centre Pompidou, which acquired eighteen NFT-based works in 2021 as part of a broader push to collect digital-native art, created an internal checklist asking for the smart contract address, the hosting arrangement for any off-chain assets, and a plan for what happens if the hosting service shuts down. ZKM Center for Art and Media in Karlsruhe, which has worked with computer-generated art since its founding in 1989, extended its existing software-art documentation forms to cover model weights and training data provenance rather than building an entirely separate track.
The practical effect is that AI-based submissions now often take longer to clear review than a comparable painting or photograph, not because the artistic judgment is harder, but because the technical due diligence has more unknowns. Curators interviewed for institutional reports describe review timelines stretching from a few weeks for conventional media to several months for AI-based installations requiring live compute infrastructure.
Process Documentation as a Curatorial Requirement
Most museums that have published guidance now ask for a documentation packet alongside the work itself. This is treated as part of the artwork's record, comparable to a conservator's condition report, rather than as optional background material.
- Model and version identification: the specific model or checkpoint used, including version number, since outputs from "the same" model can differ across releases.
- Prompt and parameter logs: the text prompts, seed values, guidance scale, and any post-processing steps applied after generation.
- Training data disclosure: whether the artist trained or fine-tuned a model on a specific dataset, and whether that dataset's licensing terms permit exhibition and commercial display.
- Human intervention points: where in the process a human selected, edited, composited, or rejected outputs, since this affects both authorship claims and conservation planning.
- Reproducibility statement: a plain statement of whether the exact output can be regenerated, and under what conditions it cannot.
The Whitney Museum's Artport program, running since 2001 for net-based and software art, had already built similar documentation habits for earlier digital work, and several curators note that AI submissions are being folded into that existing infrastructure rather than treated as a wholly new discipline. The difference is the training data question, which did not arise with earlier generative or algorithmic art in the same legal form.
Conceptual Framing Versus Technical Novelty
A recurring point in published curatorial statements is that technical novelty is not treated as an independent criterion. A museum is not evaluating whether the model is state of the art; it is evaluating whether the artist's use of the tool produces a coherent conceptual or visual statement. Several curators describe rejecting submissions that amount to a demonstration of a new model's capability rather than an artistic argument.
This distinction mirrors how museums historically treated early video art: the fact that a Portapak camera was new technology in 1965 did not by itself justify exhibition. What mattered was whether an artist like Nam June Paik used the medium to make a claim about perception, feedback, or broadcast culture. Curators now apply a similar test to AI-generated work, asking what the piece says about authorship, labor, data, or representation that could not be said as clearly with another medium.
In practice this means a submission letter or artist statement is read closely for whether it treats the model as a tool subordinate to a concept, or as the subject of the work itself. Both approaches can clear review, but they are evaluated against different comparison sets: a tool-use piece is compared against other conceptual or process-based art, while a piece about the model itself is compared against the history of art that interrogates its own medium, such as structural film or process art from the 1960s and 1970s.
Exhibition Criteria Versus Acquisition Criteria
Institutions draw a sharper line between showing a work temporarily and acquiring it permanently than they do for most physical media, because the long-term risks differ substantially.
| Criterion | Exhibition (temporary loan or commission) | Acquisition (permanent collection) |
|---|---|---|
| Reproducibility requirement | Not always required; a one-time live generation may be acceptable | Usually required, or an agreed fallback format (fixed video export, print) |
| Training data licensing proof | Requested but sometimes waived for short runs | Typically mandatory before accession |
| Software dependency plan | Addressed per-exhibition with the artist present or on call | Requires a written migration or emulation strategy for decades-long storage |
| Compute infrastructure cost | Budgeted per show, often borne by the artist or a grant | Becomes a recurring line item in collection maintenance budgets |
| Conservation documentation | Basic installation notes | Full technical dossier comparable to time-based media conservation standards |
This gap explains why a museum can show a generative AI installation during a festival or temporary program while declining, for the time being, to add it to the permanent collection. The Centre Pompidou's NFT acquisitions were explicitly framed as a pilot program with a stated intention to monitor how the works aged technically before expanding the practice further.
Examples From Published Institutional Policies
Few museums have released a single public document titled "AI art policy." Instead, guidance tends to appear as amendments to existing digital art or new media collecting policies.
- ZKM's new media conservation unit has published methodology papers describing how software art documentation, originally built for 1990s interactive installations, was extended to cover model checkpoints and dependency versioning for machine learning works shown from 2018 onward.
- The Barbican Centre, which has hosted several AI-themed exhibitions including "AI: More than Human" in 2019, applies a curatorial framework that groups works by whether the AI component is generative, analytical, or robotic, and reviews each category against different safety and technical criteria, since a robotic installation raises physical safety questions that a generative print does not.
- Ars Electronica, through its Prix Ars Electronica awards running since 1987, added a computer animation and later an AI-specific jury track, and its published jury statements function as de facto criteria documents even though they are not framed as institutional policy.
A consistent feature across these sources is that none treat "made with AI" as a single category requiring one uniform rule. Each separates works by what kind of model is involved, what data it was trained on, and whether the output is static or ongoing, and applies criteria accordingly.
What Curators Say They Still Struggle to Evaluate
Published interviews and conference panels from 2022 through 2024 point to a few unresolved problems rather than a general sense that the field has stabilized.
Authorship attribution remains contested in legal terms, which complicates wall labels and catalog entries. The U.S. Copyright Office's 2023 decision regarding Kris Kashtanova's graphic novel "Zarya of the Dawn" found that the individually AI-generated images could not be copyrighted, while the human-arranged narrative structure could. Museums citing or displaying such works face a practical labeling question with no single settled convention: some institutions credit the artist alone, others credit "artist, with [model name]," and a few include a footnote describing the legal ambiguity directly.
Long-term technical obsolescence is the second recurring concern. A model hosted via an API can be deprecated by its provider with little notice, which is a different risk profile than a VHS tape degrading slowly over decades. Several conservators describe this as closer to the preservation problem for early CD-ROM art from the 1990s, where entire works became unplayable within a decade, and say that migration strategies borrowed from that era are being tested but not yet proven over a comparable timespan.
A third issue is valuation for insurance and loan purposes, since market pricing for AI-assisted work has fluctuated sharply since the 2021 NFT boom and subsequent correction, making appraisal inconsistent across institutions. None of these problems have a single correct resolution at this time, and museums with legal or tax exposure on valuation questions generally involve outside counsel or a specialized appraiser rather than resolving the matter internally.
Common mistakes
Artists submitting AI-based work to institutional review commonly run into a short list of avoidable problems: omitting the specific model version used, assuming a cloud-hosted generation process will remain available for the full run of an exhibition without a service agreement, treating training data licensing as the gallery's problem rather than providing proof in advance, and writing an artist statement that describes the model's capabilities rather than the work's conceptual argument. Curators also flag submissions that arrive without a fallback static format in case live generation fails during installation, which can delay or derail an otherwise strong proposal.
Preparing a Submission for Institutional Review
Artists approaching a museum with AI-based work can reduce review friction by assembling a documentation packet before first contact rather than after a curator requests it.
- Write a one-page technical summary listing the model or models used, version numbers, training data sources, and whether any fine-tuning was performed.
- Prepare prompt and parameter logs for the specific outputs being proposed, not a general description of the workflow.
- Obtain or request written confirmation of training data licensing terms if the model was custom-trained, since this is increasingly a mandatory checklist item for acquisition and sometimes for exhibition.
- Draft a fallback plan: a fixed export, print, or recorded loop that can substitute for live generation if compute access fails during the exhibition period.
- Separate the artist statement into two parts, one addressing concept and one addressing process, since curators read these against different evaluation criteria.
- List infrastructure requirements explicitly, including GPU specifications, network bandwidth, and any ongoing hosting costs, so the receiving institution can budget accurately before agreeing to a loan or commission.
Institutions vary in how formal this process needs to be, and a smaller gallery may accept a shorter packet than a national museum with a dedicated conservation department. Artists working on acquisition-track proposals, where licensing or copyright status is unclear, are generally better served consulting an intellectual property attorney before submission rather than relying on the museum's legal team to resolve ownership questions on their behalf.
