Existing AI governance work provides several ways of approaching that question. The NIST AI Risk Management Framework addresses the management of risks associated with AI systems, while ISO/IEC 42005:2025 provides guidance for assessing the effects AI systems may have on individuals, groups, and society across their lifecycle. In education, UNESCO’s AI competency frameworks for students and teachers focus attention on the knowledge and competencies needed to engage with AI responsibly. Accessibility standards establish requirements for digital access, while research on participatory AI and co-design asks who participates in shaping technological systems and how much influence participants actually have (Steen, 2013; Steen et al., 2007, 2008). These approaches do not address the same problem, but collectively they show why examining an AI system only in terms of what it can technically accomplish provides an incomplete basis for institutional decision-making.

The library and information science profession is already having practical conversations on how to confront this problem. The American Library Association’s 2026 Guidance on the Use of Artificial Intelligence in Libraries asks libraries to consider when and whether to use AI systems, AI-enabled features, and vendor integrations. The guidance highlights privacy, accessibility, equity, human oversight, community needs, public trust, and vendor accountability. Recent research in library and information science has also examined whether libraries are institutionally ready for AI. Scholars ask whether organizations have adequate infrastructure, policies, staff knowledge, leadership, and other conditions for adoption. However, the study of libraries’ readiness for AI systems raises another parallel question. While a library may be ready to adopt AI, it may still need to determine whether the technology under consideration is appropriate for the people, purpose, and environment for which it is being considered.

Answering that question requires empirical evidence one may not find easily. Consider an AI-powered educational tool being evaluated by a school or library. Decision-makers can ask whether the tool works and whether they can afford it. They can also ask what populations and contexts informed its development and testing. Which languages were represented? What evidence supports claims about accessibility for people with disabilities? What conditions of access does the technology assume? Were students, educators, librarians, or affected communities involved in its development, testing, or adaptation? If they participated, what were they able to influence? The crux of these questions is not to assume that a technology will necessarily fail when such evidence is unavailable, but to expose what the adopting institution knows, what it does not know, and what evidence supports the claims on which its decision will depend.

This makes pre-adoption auditing an information problem as much as a technology problem. Vendors have information about how systems were developed and tested. Institutions have knowledge about their infrastructure, policies, users, and local conditions. Educators and information professionals contribute professional knowledge about learning and information practices. Communities bring knowledge about their own languages, experiences, needs, and circumstances. Independent assessments may provide evidence that neither vendors nor adopting institutions can produce themselves. A serious pre-adoption review would have to consider these different sources of evidence and ask where important information is missing. A vendor’s statement that a product is accessible, for example, is not the same as evidence showing what standards were tested, how testing occurred, and what limitations remain. Likewise, stating that users were consulted tells an institution little about when they participated or whether they could influence consequential decisions.

Pre-adoption auditing should not be expected to predict everything that will happen once a technology enters practice. Technologies can behave differently across environments, and consequences may emerge only through use. Post-adoption monitoring therefore remains necessary. But uncertainty after adoption is not a reason to avoid scrutiny beforehand. It is a reason to identify what can reasonably be known before a decision is made, document what remains uncertain, and determine whether the available evidence is sufficient for the proposed use. An assessment may lead an institution to adopt a technology, request additional evidence, conduct further accessibility testing, begin with a limited pilot, or reconsider the proposed use. The important point is that these choices follow examination rather than assumption.

As AI becomes more deeply embedded in educational and information environments, information professionals have reason to pay attention to this earlier part of the technology lifecycle. Our concern should extend beyond evaluating information produced by AI systems to evaluating information available about the systems themselves. Before asking communities to live with the consequences of an AI adoption decision, institutions can ask a simpler set of questions: What do we know about this technology? What evidence supports what we know? Whose knowledge is represented in that evidence? What remains unknown? These questions will not tell an institution everything about the future performance of an AI system. They can help ensure that adoption is not the point at which serious scrutiny begins.