AI generated interpretive labels for exhibits to support engagement and learning in science museums

The National Science Foundation Division of Research on Learning in Formal and Informal Settings awarded Trustees of Dartmouth College $900,000 on August 15, 2026, under the STEM Education program (CFDA 47.076) to develop adaptive, AI-generated interpretive labels for science museum exhibits.

The project investigates how exhibit interpretation can adapt in real time to visitors' different levels of understanding of artificial intelligence and other STEM topics. A team of university researchers and museum practitioners will develop and study adaptive exhibit labels that acknowledge, extend, or gently reframe what visitors notice and say about AI systems. The work focuses on strengthening public understanding of how AI systems work, including their capabilities and limitations, to support informed citizenship and workforce readiness. The project team will conduct three connected research phases across four interactive exhibits about AI at the Exploratorium, a science museum in San Francisco. The methods, datasets, and software developed will be openly shared so museums and community organizations of all sizes can adopt evidence-based, adaptive approaches that extend personalized learning to broad audiences.

The period of performance runs from August 15, 2026, through July 31, 2029. Work is performed in Hanover, New Hampshire, where Dartmouth College's Office of Sponsored Projects administers the award.

Abstract

Science museums are among the most trusted and widely used settings for lifelong learning in the united states, collectively hosting roughly 850 million visits each year. Even carefully designed exhibits face a basic challenge: they must support visitors who arrive with very different backgrounds, interests, and prior understanding, using labels and signage that are necessarily one-size-fits-all. This challenge is especially acute for exhibits about artificial intelligence (AI), a topic where public understanding varies widely and where building accurate mental models requires explanations tailored to what each visitor already knows. This project investigates how exhibit interpretation can adapt in real time to the different ways visitors make sense of science, technology, engineering, and mathematics (STEM) ideas. Working at the intersection of the learning sciences and computer science, a team of university researchers and museum practitioners develops and studies adaptive exhibit labels that acknowledge, extend, or gently reframe what visitors notice and say. By strengthening public understanding of how AI systems work, including their capabilities and limitations, the project supports informed citizenship and workforce readiness and advances the national interest in a scientifically and technologically capable public. Because the methods, datasets, and software developed are openly shared, museums and community organizations of all sizes will be able to adopt evidence-based, adaptive approaches that extend personalized learning to broad public audiences. The project addresses three research questions through three connected phases of work. Each phase is conducted across four interactive exhibits about AI with visitors to the exploratorium, a science museum in SAN francisco. First, the team establishes whether knowledge representations generated from brief museum conversations can accurately approximate visitors' conceptual understanding. To do this, the researchers model visitor understanding using three complementary representations: knowledge maps that locate concepts in a geometric space, knowledge graphs that capture dependencies among concepts, and visitor profiles that summarize each group's evolving understanding. These automated representations are compared against expert human coding of the same conversations. Second, an expert panel identifies recurring learning opportunities revealed by comparing visitor understanding with canonical models of AI concepts, and maps these opportunities to evidence-based scaffolding strategies drawn from research on museum learning. Third, a controlled study tests whether adaptive labels that are generated using large language models (LLMS) and constrained to remain accurate and responsive improve visitor engagement and conceptual understanding relative to static labels. Anticipated contributions include methods for capturing conceptual understanding in informal learning settings, techniques for aligning visitor-generated representations with reference models, and approaches for constraining language-model outputs to reduce errors while preserving responsiveness. Deliverables include an openly licensed and documented software toolkit, shared datasets of annotated visitor conversations, a taxonomy of learning opportunities and scaffolding strategies, design guidelines for adaptive interpretation, and peer-reviewed publications. This project is funded by the research on innovative technologies for enhanced learning (RITEL) program that supports early-stage exploratory research in emerging technologies for teaching and learning. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are planned for this award.