How Cincinnati Museum Center Built an AI Learning Community

Most museum staff don’t need to be convinced that AI is changing things. What they need is somewhere safe to figure out what that actually means for their work — and permission to be honest about how they feel. Cincinnati Museum Center decided to build that space themselves.

The museum — a natural history and science museum, a history museum, and a children’s museum, all housed within a National Historic Landmark art deco train station — has a staff of around 180. When Whitney Owens, Chief Learning Officer, and her colleagues began thinking about how to approach AI, they didn’t start with a policy or a training programme. They started with a question: how do you bring people who feel completely differently about something into the same room and get them to learn together?

At the Museum AI Summit 2026, Owens was joined by Jason Perkins, Vice President of Technology, and Katie Webb, Senior Director of Digital Assets Management, to share what they tried and what they found.

Start With the Policy, Not the Tools

Before any experimentation began, Cincinnati Museum Center developed what the team describes as a rapid-response AI policy — a set of guidelines designed to give staff enough structure to feel safe, without being so restrictive that genuine learning becomes impossible.

The policy names the tools the organisation considers acceptable: Microsoft Copilot, which was already rolling out within their Microsoft 365 environment; and free-tier versions of ChatGPT, Google Gemini, and Anthropic’s Claude, each with clear guidance on what can and cannot be uploaded. Paid-tier tools are permitted, but approved on a case-by-case basis, particularly where data security is a concern.

The free-tier approach was a deliberate choice. As Perkins explains, the built-in limits of free models kept both usage and cost in check while still being powerful enough to experiment with meaningfully.

Bring the Right People Together

For the working group itself, the team brought together fourteen people from eleven departments — Marketing, Information Technology, Exhibit Development, Guest Services, Finance, Human Resources, Education, and Collections and Research. The intention was to get a wide enough range of perspectives without the group becoming unmanageable.

The recommended size, based on their experience, is five to fifteen people. Large enough for genuine diversity of thought; small enough for real conversation.

Begin With Feelings, Not Facts

The decision that perhaps most defines Cincinnati Museum Center’s approach is where they chose to start. Their first meeting didn’t open with a presentation on large language models or a comparison of AI platforms. It opened with a simple question put to everyone in the room: what’s one thing that excites you about AI, and one thing that terrifies you?

The answers were candid. One staff member spoke about the potential to analyse large visitor datasets in ways that are simply impossible in a spreadsheet — and about fears that people would stop fact-checking and let AI become their only source of truth. Another raised the possibility of using optical character recognition to transcribe handwritten field notes, alongside genuine concern about the environmental impact of data centres on biodiversity. A third described the relief of finally having what felt like an admin assistant without needing budget approval — and worry about how fast the technology was moving and the racial and gender bias baked into training data.

“Robot overlords” came up more than once.

The exercise wasn’t just a warm-up. It was the foundation. By surfacing assumptions, anxieties and enthusiasms at the outset, the group created the conditions for honest learning rather than performative enthusiasm.

Build Shared Understanding Before Asking for Projects

Once the temperature of the room had been taken, Perkins and Webb shared foundational knowledge: what AI is and isn’t, how large language models differ from one another, what the environmental costs of AI use actually look like. They reviewed terminology, shared examples from museums and other industries, and walked through what the IT and digital development teams were already working on.

They also asked participants to read articles on AI from the July – August 2025 issue of Museum Magazine and share reflections. The goal was to give everyone a common reference point before moving into experimentation.

By the end of the first session, the group was asked to share one word describing how they were feeling. Responses included curious, cautiously optimistic, and wary. One participant described herself as apprehensive — but also, she noted, looking forward to seeing what AI might do for her.

Run Real Projects, Then Report Back

The second and third sessions focused on practical experiments. Staff were asked to identify a challenge in their own work area that AI might help with, then try something — and come back to the group to say what happened.

The results were instructive precisely because they weren’t uniformly successful.

Abijita Debata from Human Resources used AI to create customised management training materials, normally an external cost of $10,000–20,000 per year. The result, she reported, was a better product — more tailored to the museum’s specific needs, and produced in far less time and at no cost.

Makayla Dean from Collections attempted to use AI to transcribe handwritten field and preparation notes — records that represent the primary data for the museum’s specimens. She tried two tools. Transkribus, designed specifically for handwriting recognition, was good at identifying where writing appeared on a page but struggled badly with accurate transcription of abbreviated field records. ChatGPT produced far more accurate plain-text transcriptions, but fell apart when asked to transform that text into structured data for cataloguing — it lacked the field-specific knowledge to understand what each data point represented, and couldn’t grasp that a blank field is as meaningful as a populated one.

Valerie Horobik tackled a dataset of around 5,000 outreach education customer records collected over eight years, trying to identify pre-COVID regulars who hadn’t returned post-pandemic. The challenge began with the data itself — the same customer might appear under multiple spellings, punctuations and abbreviations. She used ChatGPT and Copilot within Excel to build a formula that standardised the entries, then used Copilot to generate a table of non-returning customers sorted by their pre-COVID visit frequency. The output gave her a clear, prioritised list for follow-up.

Emily Bertolo, working on camp materials, chose a project that carefully avoided her two main concerns — sensitive participant data and AI image generation. She tested Google Gemini and Google NotebookLM against a set of activity documents, asking each to compile a consolidated materials list. NotebookLM, she found, generally required less back-and-forth, handled missing information more intelligently, and provided direct links to the source documents when flagging where data had been drawn from. The prompt refinement and error-checking, she noted, was the most time-consuming part of the process — and also, she argued, the most necessary.

What the Group Learned

Across the projects and the reflection sessions that followed, several themes kept returning: the importance of data safety, questions about AI’s environmental impact, and the ongoing need for human oversight and fact-checking.

The message from participants wasn’t that AI is transformative across the board. It was more nuanced than that. Debata described it as a great tool, provided guardrails are in place, policies are clear, and humans remain in the loop. Horobik said she was impressed — but with a healthy amount of caution. Bertolo concluded that it should not be used without general knowledge of the risks, and should always have a human component to verify the outputs.

That kind of measured, considered response is, arguably, exactly what a good learning process is designed to produce.

A Model Others Can Follow

Cincinnati Museum Center is now planning a second working group this summer, with the intention of spreading knowledge further across the organisation and helping new staff go through the same process of structured experimentation.

For those looking to do something similar, the team offers a straightforward framework:

Identify a working group of five to fifteen people from a range of departments. Start the first session with feelings and perceptions, not facts. Use a credible outside resource to anchor shared reflection. Give people a layperson’s overview of AI and name the tools your organisation considers safe. Ask staff to brainstorm and then run real projects — and create space for them to report back honestly on what worked and what didn’t. Finally, empower working group members to become resources within their own teams, carrying the learning outward.

The approach won’t work for every museum. But the underlying principle — that navigating a genuinely uncertain technology requires both emotional honesty and hands-on learning — seems broadly applicable.

As Whitney Owens put it at the close of the session: “By convening an internal team and paying attention to both emotional responses and hands-on learning, we’ve tried to create a safe playground where staff can try things out, get feedback, and try again.”

This post is based on a presentation delivered at the Museum AI Summit 2026 by Whitney Owens, Jason Perkins and Katie Webb from the Cincinnati Museum Center.