Research Article
AI Resources in Higher Education: What Instructors Need and How We Deliver
Synthesis: Iskandarova, Yi, Taggart and colleagues asked Virginia instructors what they actually want from an open resource on AI in teaching, built one, and then asked instructors across the state to judge the prototype. The three-phase design ran from March 2024 to September 2025: a statewide needs assessment with 149 instructors, development of the open educational resource Fostering AI literacy: A guide for educators in higher education in Pressbooks, and a prototype evaluation with 75 instructors. The needs assessment is the most useful part of the paper for anyone who builds faculty development. Instructors wanted AI literacy, ethics, and pedagogy, in that order, and they ranked AI fundamentals last; they asked for interactive, self-paced, openly licensed material with classroom-ready examples rather than theory. The evaluation found strong stated intent, a majority satisfied with what they reviewed, and a consistent complaint about volume and navigation rather than content. The authors read that combination as a design brief: resources have to be modular, maintainable, and neutral about adoption, presenting selective use and refusal as legitimate pedagogical decisions rather than treating AI integration as the destination.
Key Findings
- Instructors are interested and anxious at the same time. 94% reported at least some interest in integrating AI into their teaching, including 54% who were very or extremely interested, and their goals split between efficiency (course design, quizzes, automating grading to free time for feedback) and student learning (engagement, critical thinking, creativity). Their stated challenges were academic integrity, student overreliance, weakened critical thinking, inaccurate content, and a lack of training and institutional backing.
- The content priorities are AI literacy, ethics, and pedagogy, with fundamentals last. Mean importance ratings on a five-point scale were 4.46 for AI literacy, 4.43 for AI ethics, and 4.37 for AI pedagogy and assignments, ahead of 4.17 for AI tools, 4.14 for AI policy, and 3.30 for AI fundamental concepts (n = 114).
- The preferred format is interactive, self-paced, and social. Interactive online modules were selected by 76%, videos by 56%, structured online courses by 50%, text-based resources by 46%, in-person courses by 42%, and podcasts by 27%. 77% wanted a discussion thread or community feature, and 79% rated adaptability, meaning open licensing and permission to modify, as moderately to extremely important.
- Practical classroom material outranks theory by a wide margin. Open-ended answers asked for discipline-specific examples, adaptable templates, authentic scenarios, and step-by-step guides. One participant put the preference in terms of their own students: "Just like my students, I will be interested in the practical 'how to' material in your online book."
- The prototype cleared the intent bar. After reading only the table of contents, 81% expected to find useful information, 80% were interested in reading it, 77% saw it as relevant to their own AI literacy, and 75% expected it to help them foster students' AI literacy. Among the participants who reviewed a chapter, 80% reported being somewhat or very satisfied, and 91% expressed at least slight interest in reading the full resource, with 62.5% very or extremely interested.
- The defects are volume and navigation, not substance. 64% found the resource somewhat or very easy to navigate, but 11% found it somewhat difficult and 4% very difficult. Reviewers described too much clicking, too many external links, and too much length: "most teachers (K-12 and higher education) are not likely to read a 20-chapter textbook on AI."
- Neutrality about adoption is a design requirement. One reviewer read the draft as assuming that AI use is the future and that instructors should get on board, and another questioned the credibility of vendor-produced videos. The authors conclude that a resource must present adoption, selective use, limitation, and non-use as legitimate pedagogical decisions, which makes critical agency part of usability rather than an add-on.
How the study was done
The design is a multiphase mixed methods study in three connected phases, embedded in a statewide initiative on generative AI in higher education. Phase one was a needs assessment: an anonymous Qualtrics survey distributed between March and May 2024 through centers for teaching and learning and university listservs across Virginia. Recruitment was voluntary and nonprobability, invitations were redistributed through institutional networks, and the authors therefore report no sampling frame and no response rate. 149 instructors completed it. The sample was 51.7% public four-year, 29.5% community college, and 8.1% private four-year, with the remainder not answering; 34.9% tenured or tenure-track, 25.5% full-time non-tenure-track, and 18.8% part-time or adjunct; humanities and the arts (32.2%) and STEM and information technology (30.2%) were the largest disciplinary groups. The instrument combined Likert items, multiple choice, ranking, and open-ended questions, and demographic items were optional, so totals vary by question.
Phase two was development. The team built Fostering AI literacy: A guide for educators in higher education in Pressbooks, with interactive elements in Genially, Padlet, and H5P, released under a CC BY-NC-SA license and organized as an overview, three competence levels (know and understand, use and apply, evaluate and create), and a closing section, for 21 chapters in all. Ethics and academic integrity were embedded at every level rather than kept as a separate topic, and each chapter carried a Reflect and Apply educator's toolkit. Generative AI was used for brainstorming, summarization, and format adaptation, with human editing and approval at each round.
Phase three was a formative evaluation between May and September 2025, again an anonymous Qualtrics survey through the same channels. 75 instructors completed it; 56 reviewed at least one completed chapter and answered the five-item assessment, 43 answered the first open-ended question and 39 the second. Analysis was descriptive statistics for the closed items and codebook thematic analysis in an interpretive frame for the open responses, with four researchers coding a shared subset and reconciling differences in discussion. No intercoder agreement statistic was calculated, by design. The study was approved by the George Mason University IRB and funded by the State Council of Higher Education for Virginia and a SoTL collaboration grant.
What this means for practice
- Faculty developers: survey your own instructors before you choose a format. The Virginia data are a template as much as a finding. A short instrument that asks about goals, barriers, content priorities, format preference, and willingness to adapt will produce a local design brief, and the questions here are reusable.
- Instructional designers: build for the second visit, not the first. Reviewers wanted material they could find, evaluate, and implement quickly, and they resented clicking. Short modules with visible entry points, written summaries beside video, and a clear distinction between core and optional material address the failure the evaluation actually found.
- Librarians and OER teams: license for remixing from the start. The stated preference for adaptability is the argument for open licensing, but the paper's harder point is maintenance. Modular structure, version histories, and update schedules are what keep an AI resource from going stale within a semester.
- Centers for teaching and learning: make non-adoption a legitimate answer. A resource that reads as advocacy loses the instructors who are skeptical, and the paper records exactly that reaction. Framing adoption, selective use, limitation, and refusal as pedagogical choices keeps the skeptical instructor in the conversation.
- Institutional leaders: fund the maintenance, not only the launch. The authors treat community contribution and periodic revision as what separates a living resource from a static one, which means staff time, contributor guidelines, and a named owner after the grant ends.
Limitations
- The sample selected itself. Recruitment was voluntary and nonprobability, with no sampling frame and no calculable response rate, and the authors note the sample probably over-represents instructors already interested in or worried about AI, which 94% interest makes hard to disentangle from the finding itself.
- The evaluation measures impressions, not effects. 75 respondents reviewed a draft, most of them only the table of contents or a single Level 1 chapter. Nothing here shows whether instructors changed a course, taught differently, or improved student learning.
- The evidence is one state and one moment. Data come from Virginia institutions between 2024 and 2025, in a period when AI tools and institutional policies were changing quickly, and the authors are explicit that transfer to other regions should not be assumed.
- The sample skews experienced and community college. 40% of evaluation respondents had taught for 16 years or more, and community colleges were well represented, which the authors treat as valuable but note may under-represent early-career staff and research-intensive institutions.
- Coding reliability was not quantified. Four researchers used an interpretive thematic analysis and reached understanding through discussion, with no intercoder agreement statistic, so the themes are a defensible reading rather than a reproducible coding.
Citation
Iskandarova, S., Yi, F., Taggart, J., Mickel, B., Bayraktar, B., & Henry, D. (2026). AI Resources in Higher Education: What Instructors Need and How We Deliver. Frontiers of Digital Education, 4(1), article 1.