MCN 2026

Museums are increasingly experimenting with AI-driven interpretation tools, but what are the possibilities for the museum itself becoming a public laboratory for testing how visitors interact with AI?

At the MIT Museum, one of our strategic goals is to become a public lab for MIT. We have been developing and studying multiple approaches to conversational and embodied AI interpretation, including:

•               A multilingual AI-powered gallery guide

•               DebunkBot, an exhibit exploring misinformation and trust

•               An AI-enabled robotic docent and wayfinding experience based on Jibo technology

These projects explore how different forms of AI — handheld, embedded within exhibits, or physically present in gallery spaces — shape visitor inquiry, engagement, trust, and learning.

This interactive session invites participants to examine three distinct AI interpretation models and discuss questions such as:

•               Do visitors trust AI differently inside museums?

•               Can museums use AI to help visitors think critically about AI itself?

•               What forms of AI best support curiosity, accessibility, and conversation?

•               How can museums ethically conduct live visitor-centered AI experimentation in public spaces?

•               How can AI deepen engagement with exhibit content?

•               How can museums bring AI into museums to augment the visitor experience without infringing on staff presence and visitor interaction?

Participants will break into small groups to explore real-world scenarios drawn from active MIT Museum projects as well as projects from the home museums of session paerticipants. Group conversation will allow cross-pollination of ideas, best practices, and lessons learned across the museum field, as we collaboratively reflect on the opportunities, tensions, and future directions of AI in museum environments. Then each group will report out and share their experiences with AI.

Session Format: 45-minute interactive session

•    15 min framing presentation

•    20 min small-group discussion activity

•    10 min report-out and synthesis

Most AI-for-archives work in GLAM comes from engineers at vendor companies or well-resourced institutional AI teams. This session is a practitioner case study from a different corner: a public library manager — without an engineering background or an in-house AI team — building local-first AI infrastructure for a rights-sensitive regional broadcast archive.

The Northeast Ohio Broadcast Archives (NEOBA) are roughly 10,000–15,000 hours of regional broadcast footage. A digitized slice of 349 tapes from 1977–1982 served as the starting material. Cloud AI services were off the table — rights posture, institutional trust, and the sensitivity of local-community material all pointed to local-first processing, as did the budget and staffing reality of a public library.

Two artifacts will be demonstrated. The first is a local AI pipeline running on a single M-series Mac: scene detection, segment-level visual description (Gemma 4 E4B), transcription (mlx-whisper), face and entity detection (InsightFace), and Dublin-Core catalog synthesis (Gemma 4 26B A4B). The second is a public-facing Archive Explorer surfacing 14,242 classified items, 3,446 extracted entities, and 292 semantic clusters — with narrative threads and an AI-assisted chat interface for natural-language discovery that let someone unfamiliar with the collection begin exploring from any angle.

Both are examined honestly: what local-first bought in institutional trust, what it cost in throughput and fidelity, where contemporary AI tools extended what a small library team could do — and where they did not. The session also takes up a question the theme invites directly: as AI capability becomes cheaper and more locally runnable, who gets to build it, for whose collections, at whose institutions? A non-engineer practitioner working from a single workstation inside a public library is an unlikely builder profile for this work in 2026 — and that unlikelihood is part of the point.

Attendees leave with a reference implementation for local AI in a rights-sensitive archival setting, a set of trade-off conversations worth having before cloud-defaulting, and a practitioner perspective on AI-augmented work at institutions whose starting posture is not that of a cloud-first AI lab.

This session will provide an overview of the Museum of Modern Art’s implementation of new artificial intelligence tools in support of its access and accessibility priorities. The discussion will detail the decision-making involved in assessing and judiciously integrating AI into workflows, focusing on the identification and description of art assets. The presentation will recount three projects: identifying artworks in installation images; generating alt-text descriptions for images; and transcribing media assets with AI, as well as touching on available AI tools the Museum deemed as unsuitable or inappropriate for use.

This talk will address considerations applicable to other GLAM environments, from the specific challenges of working with a corpus of complex Modern Art material to assessing the suitability, accuracy, and resource expenditure involved in using artificial intelligence tools. The session will also detail the Museum-specific issues influencing the design and viability of the tools used, and consider the degree to which these models are replicable in other environments. In keeping with the conference's focus on Trust and Longevity, we will discuss the benefits of augmenting and repurposing Museum data to create a more expansive and trustworthy record of MoMA’s exhibition activity, and provide information for colleagues in the field weighing the costs and benefits of using AI in collections management.

This session would be well-suited for incorporation into a longer panel focused on institutional workflows, new platform adoption, or the experimental use of artificial intelligence tools in GLAM settings.

This summer, Crystal Bridges Museum of American Art launched the first version of its very own AI Companion App designed to help visitors identify works of art and have a rich conversation about what they’re seeing—all with artificial intelligence.

One major question we had to answer: can a conversational AI experience, even one designed in partnership with the museum’s curatorial staff, be trusted as a real-time visitor companion?

In this highly practical and engaging 45‑minute session, Stace Treat - Director of Interpretation - and Amanda Boomstra - Strategic advisor for museums - will share the product and institutional decisions required to make AI experiences trustworthy, deeply engaging, and built to scale beyond the current collection. Our starting point is always visitor value: we build what genuinely enhances the experience, not what simply showcases technology. 

We’ll cover the end-to-end approach: shaping a companion personality and conversation design that mirrors natural turns in human conversation; the challenges of grounding responses in legacy collections data; and designing accessible, inclusive interaction patterns that feel familiar to every user. Attendees will leave with clear takeaways for pairing data foundations, governance, and UX transparency to sustain trust in AI-driven interpretation over time.

As aquariums expand digital touchpoints across the guest journey, visitors expect fast, accurate access to information. At the Monterey Bay Aquarium, we introduced the Kelp Desk chatbot to help guests quickly find answers across our website. 

But deploying a chatbot isn’t just a technical issue; it’s a trust exercise. 

In this session, we’ll share what it takes to design, launch, and sustain a chatbot that supports the guest experience. We’ll cover how content quality impacts AI performance, what’s required to maintain accuracy over time, and how to manage a chatbot within existing resources. With real-life grounded lessons, this session focuses on both the opportunities and limitations of AI in a mission-driven environment. 

Attendees will leave with practical insight on how to implement and sustain a chatbot in a way that builds long-term trust with their digital audiences.

In early 2026, LACMA shipped Collections Online — a public search and discovery site for the museum's collection — in roughly 3.5 months from kickoff to launch. The team was small (LACMA's digital team of two) and the work was deeply collaborative across overlapping skills: LACMA's Senior Director of Publishing — portfolio across editorial, design, and web/digital — anchoring the experience perspective and the internal advocacy that kept the work aligned across the organization; an independent backend engineer on the data platform application; the Folder Design team behind the visual design system and reference demos; the Urban Insight team behind the Netlify-hosted Next.js front end — search integration, frontend architecture, performance, and the discovery experience; and LACMA's Director of Digital Strategy & Infrastructure, who architected the overall environment and built the systems around the data platform (image, AI, search, monitoring, admin tooling, data flows), iterated on intent, and held day-to-day validation, troubleshooting, and coordination.

At the center sits a custom data platform — a FastAPI on ECS application backed by PostgreSQL/RDS — that ingests eleven TMS SQL views (about 1.1 million rows per cycle) plus 119,000 AI-enriched records from DynamoDB, applies user-defined transformation rulesets, and publishes through REST, GraphQL, and a bulk OpenSearch index. Two VPCs joined by peering, schema/rules admin UI, 6-hour ingest cadence with admin-triggered refresh.

A Lambda + SQS image processing pipeline moves 68,000 image-bearing objects (and derivatives) into KMS-encrypted S3. An AI enrichment layer — a SageMaker-hosted SigLIP 2 vision model run ephemerally on g5.xlarge, plus a long-running EC2 service for visual descriptions, embeddings, and search-time text encoding — produces the metadata for hybrid vector search and accessibility. An OpenSearch cluster handles faceted keyword search. A pair of admin tools — diagnostic Lookups and editorial Overrides — keep curatorial, engineering, and the data pipeline honest about what is surfacing on the site.

On the frontend, Next.js + Netlify deliver the public experience, CDN-optimized for scale.

- End-to-End Architecture: the AWS/FastAPI stack, frontend delivery, and operations — with cost and performance numbers for each piece.

- The Reality of AI-Assisted Engineering: where Claude and Gemini delivered 10x acceleration, where they introduced subtle costly bugs, and where we reverted to traditional engineering.

- Operationalizing AI: prompt engineering, agent-based workflows, and the scaffolding that makes AI-assisted work repeatable across a hybrid team.

In this session, we break down both the technical architecture and the collection discovery experience behind the build, sharing key architectural decisions, tradeoffs, lessons learned, and our playbook, including:

The goal is to share a working playbook, with the rough edges visible, for any institution considering a similar build.

In March 2026, the RISD Museum opened Ways of Looking, an introductory gallery that orients visitors to the museum while offering tools for engaging with artwork (https://risdmuseum.org/exhibitions-events/exhibitions/ways-looking). This project departs from standard museum practice: it was led by interpretation and education staff rather than curators; it spans two galleries but presents only three artworks, prioritizing close engagement; and it positions the visitor’s response as not only valid, but essential. Works of art, collection research, and visitors’ contributions are presented with equal weight in the space, creating a collaborative experience. 

This session explores how power is constructed and distributed in museum collections, curation, and interpretation. Using Ways of Looking as a case study, participants will examine practical strategies for sharing authority with visitors. Attendees will learn how interpretation and technological interventions can disrupt traditional narratives and create more inclusive museum experiences.

In one gallery, we created a drawing space with projections, drawing prompts, and a pegboard for displaying visitor responses. In another, two audio pieces introduced perspectives from an artist and a security guard. Presented through spatial audio and a landline phone turned into an audio player/recorder, these works created a communal listening experience while inviting visitors to record their questions. We made explicit what museum professionals often leave implicit, including how to read labels and how to interpret works of art. We presented four key ideas as springboards for interpretation: process, meanings, origins, and connections. Curiosity and play were embedded throughout the design, encouraging engagement across ages and levels of familiarity.

Visitors often enter museums unsure how to look at art, what questions to ask, or how to do a museum “right,” leading to a lack of confidence or connection. The RISD Museum is one of the few publicly accessible spaces within a larger academic institution. We aim to share knowledge rather than gatekeep it. Visitors should understand that a museum is a network of people with varied skills and perspectives, not a monolith. Through this space, we aim to build trust between the institution and its visitors.

Trust is a two-way exchange: museum professionals must build credibility through accessible, transparent experiences, and Ways of Looking includes resources visitors can carry beyond the galleries. At the same time, institutions must demonstrate trust in visitors’ experiences, questions, and creative insight. This reciprocity is expressed through tools like the landline message system, as well as more formal feedback that directly informs future experiences.

This space operationalizes the values laid out in our strategic plan: becoming more welcoming, curious, collaborative, and bold. A strategic plan should not sit on a shelf; it should be felt in practice.