Museums have more stories than they have the time or resources to tell them.
They live in collection records, photographs, letters, oral histories, and uncataloged boxes. Some are known to curators and researchers but have never found their way into exhibits or exhibitions. Others are still waiting to be discovered when someone finally has the time to trace them out.
Gaps between what museums hold and what they can share have always been part of the reality of museum life. The Smithsonian, for example, displays less than 1% of its collections at any given time. Of course, not every object warrants display, but if research and interpretation capacity were greater, much more could be made accessible.
This is why AI’s potential is so consequential for museums. It can help museum staff accelerate their work. With custom tools, agentic sub-agents, and generative capabilities, AI can help museums find connections, develop stories, and invite people deeper into their collections.
I know this isn’t always an easy proposition for everyone in the museum world. Some have serious concerns about AI, particularly its environmental impacts and the uncomfortable history of using copyrighted work to train models. I’ll circle back to those concerns, but first, I think we need to see the real size of the opportunity.
Multiplied Paths Through a Collection
Imagine a local history museum with thousands of photographs, property records, maps, and recorded interviews. Its staff may know their way around these materials and could tell compelling stories about a community’s changing streets, the founders of historic local businesses, and all the people whose names are obscured in the archives. Researching and publishing such stories takes significant time and effort.
But what if museums used AI to transcribe interviews, extract names and places from records, and identify connections across collections? AI agents can easily run a first pass over the material. Then curators can concentrate their efforts on verifying the evidence. With the help of AI, this work can scale significantly to levels previously considered out of reach.
And once a story has been developed, we can give it multiple presentations. The same source material could support a chronological account, an interactive map, or a study guide for a school group, or form the basis for deeper exploration by researchers. Visitors could dig deeper into the people, places, and objects that interest them most.
That is a different kind of access. It gives audiences more ways to find meaningful connections to the collection, while keeping the museum responsible for the ultimate interpretation.
The Work Behind a Better Digital Collection
In addition to developing new stories, AI can deepen and enrich the digital use of collections.
Consider IIIF—the International Image Interoperability Framework. It lets institutions present and share digital objects in ways that support deep zooming, comparing images, and navigating the pages of manuscripts. Implementing IIIF requires more than a directory full of image files. It needs reliable object information, rights statements, and structures that tell a viewer how the parts of an object fit together. IIIF packages that information in manifests.
For museums with thousands of objects, prepping that information can be a substantial task. But building structured data from draft descriptions is a perfect use of AI, which is more than capable of doing it. AI can reconcile inconsistent fields, identify likely page sequences, and map existing collection data into manifests for staff review. It could also help develop interpretive layers around those objects once they are available through IIIF.
Rights statements still need authoritative sources, and curators have to check the output. But when AI removes the repetitive preparation work, many more museums could bring their collections into viewing experiences that have typically demanded significant staff time and specialized technical support.
Small Museums Can Now Think Big
Ambitious museum digital projects have often depended on unusual resources. The Williams College Museum of Art’s Digital Project, for example, was supported by a three-year Mellon Foundation grant and drew on expertise across the college. Its Collection Explorer allows visitors to move from a bird’s-eye view of 12,400 objects and zoom down to each individual artwork. Such a visualization required intensive effort by students and faculty in the Computer Science department. It’s a compelling example of a marriage between museum grant funding and nearby technical talent.
Consider another example, the Chicago Blues and Jazz StoryMap. It takes photographs, recordings, interviews, and field notes from a 1977 documentation project and provides a rich scroll-reveal interface that places them in the clubs and neighborhoods where they unfolded. A visitor can follow the path of the music while hearing and seeing the artists who made it.
Both projects were significant undertakings. Staff had to research the collections, make interpretive decisions, prepare data, and then finally build out the experience. Such interactives have been out of reach for most smaller museums. But now AI can help smaller museums take on these kinds of initiatives. They can draw from their archival records, organize material around places, quickly prepare map data, prototype the experience, and finally code the website. Museum professionals still need to verify the places and tell the story. The difference is that a project like this no longer requires the same depth of specialized resources to get off the ground.
It’s not that every museum needs projects of this scale. Rather, it’s that every museum now can ask a similar question: What could people discover if we had better ways to work with the material we already hold?
Putting Environmental Impacts in Proportion
The barriers to museums broadening the reach and impact of their collections, programs, and resources are no longer just time and technology. But there are still barriers, and some of those come from staff discomfort around AI’s environmental impact. While such concerns need to be taken seriously, they also need to be properly measured.
For example, Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed about 17 billion gallons of water directly in 2023. That sounds like a very big number. However, when compared to something as ordinary as golfing, we start to bring that concern back to a reasonable measure. U.S. golf facilities applied about 531 billion gallons for irrigation in 2024—roughly 30 times as much. And compared to agricultural irrigation, the problem shrinks down even further: the U.S. Geological Survey’s most recent nationwide estimate put irrigation at about 118 billion gallons per day in 2015. These figures come from different years, but still, they help put the scale of direct data center water use into proper perspective.
A museum should be able to care about these impacts without treating every use of AI as environmentally disqualifying.
Electricity demand is also a significant concern. The International Energy Agency projects that global data center electricity use could roughly double between 2025 and 2030, reaching about 3% of global electricity demand. And AI is a major driver of that growth.
Data centers are extremely unpopular at the moment. However, the majority of data center electricity use is not for AI. Gartner estimates that in 2026 AI-optimized servers account for about 31% of worldwide data center electricity consumption. Conventional servers, along with cooling and other infrastructure, account for the rest. These data centers support our ordinary use of digital services like shopping on Amazon and streaming Netflix. The AI share is growing quickly, but putting a number on it helps us see its growth in the context of the larger digital infrastructure we use every day.
And let’s remember that as technology creates some short-term problems, it also starts to solve these problems. Cerebras, for example, says its wafer-scale systems can perform AI workloads with substantially less energy than the GPU systems in use today. As AI use grows, efficiency gains may be accompanied by innovators whose inventive solutions are likewise advancing rapidly through their own use of AI. Today’s hardware use is not static, and the trend lines may not necessarily continue on their current trajectory.
There are more speculative ideas, including data centers in space. Who knows what other technological solutions will emerge? The practical work now is to improve chips, software, cooling, and energy supply—and to choose providers that report clearly on their impact.
The Copyright Question
Copyright concerns are also a serious matter, particularly for institutions entrusted with artists’ and communities’ work. But the phrase “AI was trained on stolen content” conflates several different questions into one.
There’s a clear difference between questions about how AI companies obtained the works they trained on and questions about the lawful use of those works in training. Those questions have different answers. The U.S. Copyright Office’s analysis does not treat all AI training uses in the same way.
Think of it this way. Imagine walking into a Barnes & Noble, reading a chapter of a book, then leaving with an idea you incorporate into a blog post. Compare that to walking out with the book without paying and then doing the same thing. What you learned from the book is completely different from how you obtained the book. Some AI companies have obtained copyrighted material from pirated sources. In Bartz v. Anthropic PBC, the court found that the use of books to train the AI model was indeed fair use, but that did not excuse the company’s acquisition and retention of pirated copies. This analogy helps distinguish real ethical breaches in obtaining material from the question of whether training on it infringes copyright.
Lastly, there are concerns about AI errors and bias. Some collection materials are carefully restricted, and some contain culturally sensitive elements; the fact that AI can find and describe them does not give a museum permission to publish them directly. Staff must review public interpretation, respect community authority, and be clear about substantive use of AI.
The Museums Association’s ethics code calls on museums to use AI responsibly and transparently—assessing both the risks and benefits of emerging technologies. That’s a reasonable and balanced standard.
Mission Is Part of Ethical Calculations
Museums already make judgments about how to use resources in service of their mission. They preserve collections in climate-controlled spaces, transport objects, digitize records, and decide which stories they bring to the public.
Using AI calls for that same fundamental application of judgment. Its costs and risks are real, but so is the opportunity cost of leaving valuable material inaccessible for lack of the time and resources to transcribe, connect, interpret, and share it.
Museums can begin by applying AI tools and methods to just one collection or story. Then they can explore other ways AI can help with research, IIIF preparation, interpretation, and building interactive experiences. Yes, by all means, keep people responsible for the final decisions while measuring results against the museum’s standards and mission.
If we do that well, AI will help museums tell more of the stories they’ve been entrusted to preserve—and give far more people access to them.
