Welcome to CFM’s summer series on all things Artificial Intelligence. To catch up with our coverage so far, read three scenarios of potential AI futures—bright, dark, and muddling through—commentary by Nik Honeysett on how AI is changing how people encounter information; and a tour of the National Archives Museum’s new AI-powered exhibition. Today I’m sharing my working outline of key decisions museums face about the use of artificial intelligence.

–Elizabeth Merritt, VP Strategic Foresight and Founding Director, Center for the Future of Museums, American Alliance of Museums.

Museums face a daunting array of decisions about artificial intelligence, ranging from whether and when to integrate AI into their own practice to how to respond to AI’s impact on the relationship between people and information. 

These decisions are, by necessity, provisional, made with the best information currently available and updated as circumstances evolve. Given the speed of current developments, this effectively means continuous monitoring and agile response. Just in the past month we’ve seen the release of smaller, faster, cheaper AI models; new warnings that AI companies will need to significantly raise prices to offset their huge investments; and data from Gallup showing plunging public trust regarding AI.

Museums can lay the groundwork for this continual process by identifying critical decisions ahead of time, and preparing to respond. To help with that preparation, here is an outline that museums can use to guide their internal discussions around AI policies, along with some notes on options they might consider and questions that might inform their thinking.

I’ve grouped these decisions into two categories, those guided by values and those driven by practical considerations, while recognizing the messy imperfection of that division. The hardest decisions can be those in where these categories are in conflict—when operational advantage and institutional values collide. By starting with a values-based framework, museums can avoid being rushed into making decisions about AI based on purely short-term, practical concerns. 

Values-Based Decisions Regarding AI

Questions about values (dare I say ethics) are thorny, uncomfortable, and may require considerable discussion at all levels of the organization, including governing authority, staff, funders, and the public. Some of these values may be general—things a museum thinks any organization, or any nonprofit, should do—others might be specific to a museum’s mission. For example, an art museum might be particularly sensitive to the effect AI may have on working artists; a natural history museum may be especially concerned about environmental impact.

Provenance

How are AI products and services sourced, and when does this sourcing raise ethical red flags?

The major AI developers—Open AI, Meta, Anthropic—created their services by feeding vast amounts of data into Large Language Models (LLMs), training the algorithms to recognize and predict patterns in words and images. This data was for the most part harvested from the Internet without permission or compensation from books, journals, periodicals, blogs, personal web sites, and social media. Some of these actions may eventually be judged to have been illegal. (A number of lawsuits are currently making this case.) Others may fall into the category that researcher Richard Slaughter calls “careless non-legal innovation”—actions for which regulation does not yet exist but do immense harm.

All of us should consider whether we feel it is fair to use services built on stolen intellectual property. For museums it is a particularly relevant issue when that theft violates the rights and damages the careers and incomes of artists, scientists, or historians—the very people whose work museums are dedicated to preserving and interpreting. We as a sector are still making amends for decades of careless, non-legal, or illegal actions that built many of our most important collections. We have the opportunity to start off on the right foot when it comes to certifying the “clean provenance” of various AI tools.

With regards to ethical provenance, which tools will the museum allow staff to use, and under what circumstances? This might range from approving or proscribing tools such as ChatGPT, Claude, or Midjourney, or embedded AI functions such as Google AI Search or Microsoft Copilot.  It might involve seeking out an “ethically sourced” LLM, trained on material explicitly in the public domain, licensed content, and/or used with the permission of the creator.

These choices become even messier and more difficult when it comes to selecting software such as collections, customer relations, and membership management systems. What AI standards must a vendor meet to align with the museum’s values? Changing from one system to another can be an expensive and time-consuming effort. What will the museum do if, in the future, a vendor integrates AI in a way that conflicts with its values?

Environmental Impact

Data centers (buildings that host the servers that run AI programs) are notorious for consuming huge amounts of energy, straining local power grids and accelerating the use of fossil fuels. They can cause local “heat islands,” warming the surrounding area by up to 16 degrees Fahrenheit, and often generate noise levels that exceed 90 decibels round the clock (comparable to a leaf blower at close range, and way over the permissible limit for on-the-job exposure in an eight-hour workday).  Despite growing public concern, the data center boom is outpacing the ability or willingness of cities, states, or the country to regulate construction.

AI use is one of many factors a museum might consider when assessing its overall environmental footprint, along with energy used by HVAC, materials used in exhibition construction, shipping and travel, and the net impact of landscaping on water retention or flooding. On the positive side, AI can be used to monitor and adjust a museum’s energy efficiency in ways that reduce its overall power consumption. However, on a national scale, such savings are far exceeded by the additional pollution generated by AI.

The environmental impact of any one museum’s use of AI is negligible—but absent regulation, the impact we experience is the sum of individual decisions by individuals and organizations. Just as some museums are leaning into green practices to set a good example for the public, museums might decide to curb their use of AI as a way of raising public awareness of the environmental impact of the technology, and informing wise decision-making by individuals and policymakers.

Human Impact

A growing body of evidence documents some negative human impacts of AI. Museums might be particularly concerned about potential effects on:

• Perpetuation and amplification of bias that is built into commercially sourced LLMs. Like provenance, this is an issue that museums might want to get ahead of, given that our sector is still grappling with historical bias embedded in our records and interpretation.
• Employment, particularly the impact on specific sectors, including independent artists, illustrators, writers, editors, researchers, and craftspeople. 

How can museums prevent, detect, and remove bias from AI systems or the content they generate? How might a museums craft AI policies that acknowledge labor impacts related to their mission? (E.g., a museum of illustration might be particularly sensitive about the use of AI image generators.)

An Omnipotent Solution

One option museums might consider to address concerns about provenance and human and environmental impact is the creation of smaller, open-source AI models that run on the museum’s own servers. The training data can be drawn from material the museum owns, for which it has the rights, or can certify as copyright free. The code can be inspected and edited to minimize bias. And a museum could power its in-house systems with sustainable energy generated on site. (As a bonus, such localized systems make it easier to ensure data is kept private and secure.)

Practical Decisions Regarding AI

I’m categorizing as “practical” decisions that hinge on what works best for the museum, in terms of cost, sustainability, and overall impact on operations.

Capacity

As Nik Honeysett has eloquently explained, effective use of AI requires governance: the creation of policies, integration into planning, equitable and consistent staff training, continuous monitoring of outcomes, and vigilance about unintended consequences. This can be a heavy burden, particularly on small organizations that already have limited capacity to implement and maintain various digital technologies.

How can museums scale their engagement with AI to match their capacity to create, maintain, and implement these policies and procedures?

Agility and Adaptability

Given the speed at which the technology is evolving, organizations that integrate AI into their operations have to be agile enough to adapt to the rapid pace of change. As Axios CEO Jim VandeHel recently pointed out, organizations may need to rethink their focus and use of AI tools on a monthly basis. None of the major services are operating on a break-even basis yet, much less turning a profit, and any one of them may be gone or obsolete by next month or next year.

How can museums manage the volatility of AI, whether through devoting resources to continual monitoring and adaptation, or by identifying AI tools and applications that will be relatively stable over the life of a project? How can museums take advantage of services that are useful—for now—without becoming dependent on any one service that may disappear or become obsolete in the fairly near term? How can museums avoid decisions that are difficult to undo/change/reverse? (See this essay by Nik Honeysett on that topic.)

Long-Term Costs

After years of wooing users with access to free or low-cost AI tools, Open AI and other major providers are beginning to pivot to paid services and higher prices. These price changes, which are causing even large, for-profit companies to rethink their use of AI, may be an even greater, and unexpected, burden for small non-profits. A museum that has used free or low-cost AI to increase productivity in communications, development, or other key areas of operations may find itself in a bind when it needs to lower expectations or increase staff to achieve the same goals.

How can museums identify the potential long-term costs of integrating AI into their processes and operations, avoid becoming dependent on free-for-now services, or preemptively budget for eventual cost increases?

Impact on Staff

Companies that rush the implementation of AI often experience decreased efficiency, and heightened uncertainty, stress, fear, and confusion among staff. Overuse of AI can erode staff’s sense of autonomy, make them feel disempowered, and reduce their satisfaction at work.  Younger workers, in particular, show increasing anger and anxiety about the impact of AI on their work and lives. On the other hand, AI that is perceived to be useful and reliable can increase engagement at work and job satisfaction.

How can museums monitor the impact of AI on staff wellbeing, and calibrate its use so as to minimize negative effects and maximize benefits?

Impact on Reputation

Less than 10 percent of the public feel that AI does more good than harm, and four times as many feel the technology, on balance, is harmful. According to 2026 data from the Annual Survey of Museum-Goers, 70 percent of the general public want museums to use no AI at all when it comes to developing exhibitions, and 43 percent felt museums shouldn’t even use AI to write emails or website text. Many survey respondents feel that museums should be “dedicated to showcasing and celebrating humanity,” and that AI use would run counter to that mission. (The depth of such feelings is illustrated by the backlash to AI-generated images in social media posts from the British Museum earlier this year.)

How can museums provide the transparency the public expects around the organization’s use of AI, identify applications that are clearly aligned with mission, and eschew applications that erode public trust?  (These decisions will, quite properly, be made by each museum in light of their mission and circumstances. But I confess that one item in my “big box of dark imaginings” is that careless use of AI by individual organizations could over time, prove to be Kryptonite to museums’ superpower of trust.)

What Museums Can Do

Work through these issues in a series of conversations that include representatives of a wide variety of stakeholders—the governing authority, staff, museum visitors, the general public, subject specialists (artists, historians, scientists) who engage with your work. Ground the conversation in the museum’s mission and values, and in a shared understanding of the impact you want to have on the world. Look for areas of consensus, identify areas of disagreement, surface new issues. Assess your financial and staff capacity for managing AI. Finally, embed a framework for decision making into your planning, policies and procedures—even if the decisions themselves need to be rethought every morning when you open your newsfeed.

Yours from the future,

Elizabeth