MCN 2026

In recent years there has been much productive discussion both at MCN and within institutions on AI-generated visual descriptions for accessibility: how to generate descriptions, how good they need to be, and how institutions should oversee their use. Some museums have already successfully released AI-generated descriptions as accessibility text. 

The Met has been treating AI-generated descriptions as something different: not (yet) as a final user-facing product, but as raw search material. With descriptions of what artworks depict, embedded into vector space and indexed alongside traditional metadata, one can build a discovery experience that works not only for experts, but also for the roughly two-thirds of users who arrive at our collection without a specific artwork in mind. Users searching for "year of the horse," "mythical beasts," or "eating food together" can find relevant works for the first time. Users searching in Chinese, Spanish, or any of dozens of languages can now find the same works as users searching in English. Users without art-historical vocabulary find what they are looking for without needing to know an artist's name or an artwork’s title. And yes, the search still works for our experts and researchers who search via fields like accession number.

We will share what we learned shipping this to a production system of The Met’s 550,000 objects used by millions of visitors a year. We will cover the technical pipeline (vision-language models, embeddings, hybrid lexical-and-semantic search), the evaluation problem (with different concerns when descriptions are an input to search rather than user-facing text), the UX work (in many ways more difficult than the engineering), and what our post-launch analytics and user feedback have actually shown. We will be candid about what worked, what surprised us, what still fails, and what we would do differently. We will also share what transfers to other institutions and what is specific to working at this scale.

The brutal reality in 2026 is that there is scant effort being made in the cultural heritage sector to engage with generative systems outside of being simple consumers. Maybe one of the big vendors will give us discounted rates for a limited time and we can achieve "great things". At the end of the day though few, if any, organizations will be in a position to control their own destiny. Nor will they  have done anything to change the power dynamic of surviving on the benevolence of someone else's handouts. We simply don't have the knowledge, skills or means to deploy these systems and little to no collective work is happening right now to change that.

SFO Museum has been developing a suite of open source tools designed to run on consumer-grade hardware (even old and crusty hardware) for producing, storing, querying and visualizing vector embeddings using as many different open-weight large language models as possible. Specifically we are using vector embeddings to implement image similarity features for collections imagery and other image-based artifacts related to museum programming; to offer avenues of suggestion for investigating our various efforts derived from the properties "seen" by different ML models in the images of those works.

This idea of using vector embeddings to "bridge" different aspects of museum programming – collections imagery, photos posted to social media, exhibition and installation photos - has led SFO Museum to publish our vector embeddings for anyone else to use and to encourage the broader cultural heritage sector to do the same. Cross-institutional collections search continues to be something of a “holy grail” in museums and other cultural heritage institutions. Vector-based image similarity, while imperfect, offers a meaningful first step to finally achieve it. It is a messy, or at least fuzzy, solution and tilts more towards the “discovery and surprise” end of the spectrum rather than informed academic research. While it may not be a tool for scholars and experts, it can still provide meaningful avenues for non-experts to investigate and discover the relationship between different collections. This was still the stuff of fantasy a few years ago and now it is within our reach using nothing more than consumer-grade computer hardware.

This presentation will demonstrate the work SFO Museum has done to date. It will discuss the theory practice and gotchas of the work to date and solicit feedback, criticism and hopefully increased participation from the broader cultural heritage sector. The goal is to demonstrate that we have agency beyond a monthly-recurring fee to third-party vendors.

This session rapidly introduces 15 experimental visitor-facing GenAI prototypes we developed and tested at The Met over the course of six months. These sketches are the product of The Met’s inaugural “Technologist in Residence,” a position focused on quickly and iteratively exploring emerging tech with our visitors. The goal of these past six months was to explore the potential for AI in the on-site experience: what’s working, what’s not, and what questions does this work raise—within the organization, and with our visitors.

Our goal is to move beyond theory and provide a practical library of examples that other museums can use to inform their own AI experiments, and that show what is currently possible with GenAI. We’ll cover novel ideas, tracking nascent visitor AI backlash, small successes and outright failures.

The prototypes cover a wide range of use-cases across the museum experience, including physical navigation, tools for close-looking single artworks, collaborative games, and experiments connecting art to the outside world. For each prototype we will share our most important findings, including what visitors found useful, what they ignored, where the technology failed, and where great ideas are butting up with current technical limitations.

We will end our presentation by taking a step back and sharing our large-scale conclusions, open questions, and next steps towards a sustainable implementation of our findings.

Building on a history of metadata innovation, the Mead Art Museum at Amherst College has deepened its commitment to digital accessibility following the launch of a new internal and public-facing online collections database. This initiative centers on a strategic collaboration between the museum and the college’s Academic Technology Services (ATS). Utilizing an investigative workflow involving generative AI models, undergraduate students—under the joint supervision of museum and technology staff—produced descriptive narrations for art and material culture objects.

In their roles as coders, editors, reviewers, researchers, and narrators, undergraduate students refined AI-generated outputs, simultaneously developing critical AI literacy. Their application of "Responsible AI" was guided by a human-centered ethic, requiring rigorous fact-checking, vigilance against algorithmic bias, and the calibration of descriptive style to align with established accessibility guidelines. The preliminary results from this student cohort underscore the necessity of "Human-in-the-Loop" strategies. Furthermore, the project demonstrates that expert museum intervention remains essential to ensuring factual accuracy and maintaining social and cultural sensitivity.

This work-in-progress establishes a model for peer institutions where trust in museum metadata and the transparency of automation tools are integrated directly into the professional workflow. Beyond digital discovery and broadened access, these detailed descriptions serve as a vital institutional archive for inventory purposes and as a resilient analog resource. By providing printed or QR-coded visual descriptions, the museum ensures that objects on display remain accessible to low-vision visitors, bridging the gap between digital innovation and physical gallery experiences.

As museums and cultural institutions explore AI-driven discovery tools, a critical question emerges: how do we design generative experiences that enhance discovery while sustaining institutional trust over time? This session shares early exploratory work from the National Gallery of Art rethinking search through the lens of Generative UI (GenUI), an emerging interaction model where interfaces dynamically adapt, synthesize information, and guide users through more personalized discovery journeys rather than relying solely on static keyword-based search results.

Drawing from ongoing cross-collection search initiatives and AI-assisted prototyping work at the National Gallery of Art, this session will explore how generative search experiences may help visitors navigate complex cultural content, begin searching without knowing exact terminology, and engage with collections through more intuitive and exploratory pathways. At the same time, these systems introduce important questions around transparency, interpretive authority, bias, user expectations, and the long-term stewardship of institutional knowledge.

Through a short case study and examples of generative search concepts and prototypes, the session will examine how UX design and human-centered research can help museums experiment responsibly with emerging AI technologies while maintaining clarity, credibility, and public trust. Rather than presenting a finalized solution, this talk offers a framework for how cultural institutions can thoughtfully evaluate and prototype AI-mediated discovery experiences during a period of rapid technological change.

Robots, fever dreams, and artworks with escape fantasies: for more than ten years, the Harvard Art Museums have been asking what really happens when we trust AI to look at our art. What began as a practical effort to improve search and discovery by augmenting curatorial descriptions with machine‑generated synthetic metadata has grown into a long‑term experiment in co‑creating knowledge with machines.

Over that time, we’ve processed more than 500,000 images (depicting artworks, archival material, and exhibitions and galleries) and amassed over 120 million synthetic annotations to enrich our collections catalogue. Today, new images entering our systems are automatically queued for processing by a “panel” of commercial services: Amazon Nova and Rekognition, Clarifai, Imagga, Google Gemini and Vision, Microsoft Cognitive Services, OpenAI GPT, Anthropic Claude, Meta Llama, Mistral, Qwen, and others. Together, these models tag features, write descriptions, identify colors, read text, detect faces, and more. Their outputs are compared, combined, and surfaced through our public API (hvrd.art/api) and AI website (ai.harvardartmuseums.org) alongside human‑written cataloguing. This processing is not a temporary experiment, but part of how we work. It’s baked into our data pipeline and designed so that new models can be added over time, recognizing that describing collections is a forever project. Opinions, interpretations, and experiences of art are always in flux, for humans and machines alike.

Along the way, we’ve learned that the real questions are not just about accuracy. What happens when humans and machines look at art together? How much does “accuracy” matter for material that is inherently subjective? How can we embrace the inconsistencies, hallucinations, and disagreements between models (and between models and people) to broaden how collections are described and experienced?

This solo session will share concrete examples from over a decade of practice, including short collective looking exercises where attendees attempt to reconcile their own responses to artworks with those of AIs. We’ll explore how this work has changed our thinking about trust, authorship, and longevity in collections data, and how we are designing systems where AI is a long‑lived, transparent tool in the hands of people, rather than a faceless, monolithic black-box of implied authority.

Many museums are already doing AI engagement work, including exhibits, programs, community conversations. But what's the full range of what's possible, and how do you know where your institution fits? This session introduces a typology of 13 roles museums can play in AI public engagement, organized across four domains: building AI awareness and knowledge, facilitating public input into AI development, strengthening AI-related infrastructure, and exploring AI's social impact and meaning. Developed by the Association of Science and Technology Centers through a cross-sector workshop process, the framework offers a shared language for a field that is moving fast and often in parallel. Come map where your work already lives, and where you might want to go next.

Museums are being asked to make decisions about AI and take on new roles and practices at a moment when both the technologies themselves and the social expectations surrounding them are evolving rapidly. This session shares the synopsis of the Museums and AI convening funded by the Institute of Museum and Library Services in spring 2026, which brought together museum professionals, technologists, educators, researchers, and institutional leaders to reflect on the current state of AI in the museum field, imagine possible futures, and identify what support museums need in order to engage with AI responsibly and effectively.

Drawing from facilitated discussions, scenario-building exercises, and thematic synthesis conducted after the convening, this session presents what museum professionals currently see as the most promising and most likely trajectories for AI adoption in museums, as well as concerns and ethical issues. Rather than framing AI as either an inevitable solution or an existential threat, the session examines the nuanced perspectives that emerged across the convening.

One theme that emerged was exploring AI use for internal workflow efficiency, particularly where these tools could help address museums’ chronic capacity limitations, in contrast to visitor-facing AI experiences, with concerns about the trust the public has in Museums, hallucinations and misinformation, intellectual property, transparency, and the potential erosion of human interpretation and expertise. The convening also revealed significant concerns about uneven institutional capacity across the field. Leaders raised the issue that AI adoption could widen existing inequities between large, well-resourced museums and smaller institutions lacking staffing, technical expertise, policy guidance, or experimentation capacity. 

Many other themes arose, including AI policies, ethical use, improvements to accessibility, environmental concerns, use of Open Access and other content as AI training sets, complexities in working with AI vendors, AI literacy in the public and in museum staff, rapidly changing legal implications,  leadership skills necessary for meeting the moment, and how funders can best understand and support museums in this moment. 

Finally, participants identified significant needs for shared field-wide support, including ethical frameworks, implementation guidance, professional development, peer-learning networks, and stronger forms of AI literacy grounded in museum contexts and values.

Rather than focusing on what AI is, specific tools, or individual implementations, this session offers attendees a high-level synopsis of how museums across the field are currently thinking about AI, unpacking these themes to document where major tensions and opportunities are emerging, where there was consensus, and when viewpoints differed, and how institutions can approach AI adoption in ways that align with mission, public trust, equity, and long-term institutional values.

Description coming soon