STEM Faculty Perspectives on Generative AI in Higher Education

AI summary

Overview

Research area: AI safety and ethics in education — specifically, qualitative study of generative AI (GenAI) adoption in STEM higher education, published under arXiv category cs.CY.

Technical level: Beginner-Friendly. The paper is a focus group study reporting faculty perspectives; no prior technical knowledge of AI systems is required to follow it.

Scope: A focus group study with 29 STEM faculty in the College of Science and Engineering at San Francisco State University, examining how they integrate GenAI into teaching, what benefits and challenges they observe for student learning, and what institutional support they say they need.

What This Paper Is About

GenAI tools entered higher education largely through student use, leaving instructors to react to a technology already present in their classrooms. Some faculty have adopted GenAI for tasks such as quiz generation, curriculum design, and feedback, while others worry about student learning, assessment validity, academic integrity, and the trustworthiness of AI outputs. This paper's goal is to document, in faculty's own terms, how STEM instructors at one large public, access-oriented, Hispanic-serving institution are actually using GenAI, what they see it doing to student learning, and what policies, training, and program changes they believe are required for responsible adoption.

Key Contributions

1. Provides insights into faculty perspectives on GenAI use in higher education from a large public Hispanic-serving and access-oriented university (San Francisco State University), enrolling a substantial portion of Pell grant eligible students.
2. Identifies pedagogical benefits and challenges of GenAI use as reported by STEM faculty, spanning course design, direct student learning support, and administrative work.
3. Proposes faculty-informed recommendations on institutional support, including training, policy, and curricular/program changes.
4. Reports a faculty-reported shift in instructional labor from content creation toward expert curation, alongside a proposed dual strategy for assessment under uncertainty.

Main Findings

• Instructional use is already widespread but varied: 27 of 29 participants (93%) reported using GenAI tools for teaching and learning activities, with only two reporting no prior use. Frequency varied: occasionally (n=12), frequently (n=9), rare use (n=6), no use (n=2).
• Specific instructional applications cluster into a few categories: facilitating class discussions (n=5), generating assessment or practice questions (n=4), providing feedback to students (n=2), supporting student research activities (n=1), and automating aspects of grading (n=1). Reported uses fall into three groups: course design and preparation, directly supporting students' learning, and administrative tasks and communications.
• Prior AI familiarity was high, GenAI familiarity more mixed: For AI broadly, 37.9% reported being very familiar (n=11), 34.5% moderately familiar (n=10), 24.1% slightly familiar (n=7), and 3.4% extremely familiar (n=1), with no participants unfamiliar. For GenAI, 40.0% were moderately familiar (n=12), 33.3% very familiar (n=10), 20.0% slightly familiar (n=6), 3.3% extremely familiar (n=1), and 3.3% no familiarity (n=1).
• Labor shifts rather than shrinks: Faculty reported spending less time creating content from scratch and more time reviewing, refining, verifying, and formatting AI-generated outputs, including iterative refinement of AI-generated content and careful verification of AI-generated solutions. The paper frames this as a shift from content creation to expert curation, and raises the question of whether GenAI produces genuine gains in faculty efficiency.
• GenAI can mask gaps in understanding: Faculty reported higher assignment submission rates when students used GenAI, but also that students struggle to debug AI-generated code because they did not develop the original logic and lack foundational understanding — described as an illusion of student competency under current assessment practices.
• Assessment practices are splitting in two directions: Some faculty are reverting to more controlled formats — in-class pen-and-paper exams, oral exams, and in-person explanations of submitted work — and moving away from attempts to "AI proof" assignments as unsustainable. Others are designing assignments that require students to compare, critique, and reflect on AI-generated outputs versus human-produced ones, or to conduct both a human-led and an AI-generated code review of the same code.
• Discipline-specific student uses were reported: In CS courses, students generate code snippets and orchestrate them into larger solutions; in chemistry, students use GenAI to produce Python code for data visualization and analysis (and to produce code for tasks like data processing and curve fitting); in engineering, students submit pseudocode and use GenAI to translate it into source code; in computer networking, students use GenAI to identify anomalous network packets or convert wireframes into HTML.
• Perceived benefits include a "private teaching assistant": Faculty said GenAI helps more students complete programming assignments, enables faster idea development and code implementation, supports faster and more on-time assignment submission, helps students articulate project ideas more clearly, offers immediate out-of-class responses (often before office hours or tutoring), is especially helpful for students balancing coursework with jobs or family responsibilities, and helps students overcome learning bottlenecks faster than traditional web searches.
• Concerns extend beyond cheating: Faculty cited the prompting burden on students, uneven quality of specialized GenAI results, the difficulty and time cost of detecting AI-generated work, unreliability of detection tools (false positives, missed sophisticated use), and difficulty proving misconduct in introductory technical courses where acceptable solutions are inherently similar. Over-reliance was described as potentially bypassing critical thinking and problem-solving.
• Institutional needs center on training, policy, and curriculum: Faculty asked for professional development workshops on how large language models work (stressing that the tools do not "think" or have emotions but predict text from learned patterns), prompt engineering, and task-specific training such as building rubrics or generating lecture materials; a centralized repository of reusable prompts and case studies of successful and unsuccessful uses; a dedicated consultation service or core AI support team; communities of practice such as "AI Squares" modeled on SFSU's existing "teaching squares" (each typically four multi-disciplinary instructors); and paid release time, professional development funds, and staffing investments.
• Policy gaps and equity concerns were named: Faculty described inconsistency in GenAI policies across courses and sequential classes, confusion about citing AI-generated content, unclear intellectual property norms, accessibility barriers created by some AI tools for students with disabilities, and the need for explicit guidance and training on bias related to race, language, and gender.
• Curricular proposals and caution coexist: Participants discussed a mandatory university-wide AI literacy course, possibly implemented through a required first-year seminar during upcoming degree restructuring and counting toward multiple General Education requirements, plus an "AI for Everyone" style course for all students. They also warned against basing permanent structural decisions on short-term trends and favored department-level curriculum change over top-down mandates.
• Critical thinking interventions are recommended: The paper argues educators should deliberately introduce pedagogical strategies that "force" students into developing critical thinking despite GenAI's prevalence, citing strategies from the literature: reframing GenAI responses as questions, facilitating conflict-filled group discussions, reflecting on the process, and gamifying a learning task (Lee et al. 2025).

Methodology in Plain English

The researchers ran a qualitative focus group study. They recruited 29 faculty members from the College of Science and Engineering at San Francisco State University — including lecturers and tenured/tenure-track faculty — via email invitations to the Department of Computer Science and CoSE during the Summer and Fall 2025 semesters. Participation was voluntary, informed consent was collected electronically via DocuSign, the protocol received IRB exemption approval in Summer 2025, and participants received a $75 gift card.

They held seven focus group sessions, each about 90 minutes, over Zoom, with group sizes ranging from three to five participants. Three sessions included only computer science faculty (three participants each, nine total); the remaining four were interdisciplinary, two of which included computer science faculty alongside other disciplines, enabling comparison between discipline-specific and cross-disciplinary perspectives. Each session had at least two researchers facilitating.

At the start of each session, participants completed a six-item demographic questionnaire via Qualtrics. Discussions followed a semi-structured protocol of eight open-ended questions organized around the three research questions, with facilitators encouraging follow-up questions, elaboration, and peer-to-peer dialogue. Sessions were recorded with Zoom's cloud recording (audio, video, auto-generated transcripts), researchers took detailed notes, and data were stored in a university-managed Box folder with access restricted to the research team.

For analysis, transcripts and AI-generated session summaries were anonymized, then imported into Google NotebookLM to surface and organize prominent topics and discussion patterns across sessions, producing an interactive hierarchical representation of themes (shown as Figure 1 in the paper). The team treated these AI-assisted outputs as analytic aids rather than definitive results, and manually cross-referenced and verified the themes against the original transcripts and researcher notes.

Participant departments were: Computer Science (n=11, 41.9%), School of Engineering (n=4, 12.9%), Psychology (n=4, 9.7%), Mathematics (n=3, 9.7%), Chemistry and Biochemistry (n=3, 9.7%), School of the Environment (n=2, 6.5%), Physics and Astronomy (n=1, 3.2%), and Biology (n=1, 3.2%). Among CS faculty, four were lecturers and seven were tenured or tenure-track; among the remaining STEM departments, three were lecturers and seventeen were tenured or tenure-track.

Why This Matters

The paper argues that integrating GenAI into higher education extends beyond tool adoption and requires deliberate reconsideration of pedagogical practices, evaluation methods, and governance structures. For research, it adds STEM-specific faculty perspectives from an access-oriented, Hispanic-serving institution — a context the authors say existing work offers limited insight into — and reframes GenAI's effect on teaching as a labor shift rather than a workload reduction.

Real-world applications:

• Assessment redesign: Informing how departments restructure evaluation toward in-class, oral, and work-explaining formats, and toward assignments that require students to critique AI-generated outputs.
• Institutional policy writing: Supporting university-wide guardrails paired with department-level autonomy, including syllabus templates, citation standards for AI-generated content, and consistency across sequential courses.
• Faculty development and support infrastructure: Motivating workshops on how large language models and prompting work, prompt repositories, case studies, and a dedicated core AI consultation team.
• Curriculum planning: Informing proposals for a mandatory AI literacy course or first-year seminar, and reconsidering the progression from introductory to advanced and lab-based courses.

Industry relevance: the paper notes that AI is now used across all industries and that fundamental questions remain about which skills will be most critical for the future workforce, likely varying across industry sectors. Employers and workforce planners have a stake in whether graduates develop durable critical thinking and problem-solving skills or become dependent on AI-generated output they cannot debug or adapt.

Future Directions

• Examine longitudinal changes in faculty practices and student learning outcomes as GenAI technologies continue to evolve.
• Investigate the effectiveness of AI-informed assessment and critical thinking interventions, including the strategies cited from prior work (reframing GenAI responses as questions, conflict-filled group discussions, process reflection, and gamified learning tasks).
• Explore how institutional policies can remain adaptive across disciplines and workforce contexts, given faculty skepticism toward static mandates and premature permanent changes.
• Incorporate student perspectives and conduct cross-institutional studies to inform equitable and sustainable GenAI integration.

Target Audience

STEM faculty and department chairs considering how to handle GenAI in their courses; university administrators, policy writers, and centers for teaching and learning responsible for GenAI guidance and faculty development; education researchers studying AI adoption in higher education, especially in access-oriented and Hispanic-serving institutions; and instructional designers building AI literacy curricula. The paper is also accessible to readers with no technical background in AI.

Authors’ abstract

Generative artificial intelligence (GenAI) tools are increasingly present in higher education, yet their adoption has been largely student-driven, requiring instructors to respond to technologies already embedded in classroom practices. While some faculty have embraced GenAI for pedagogical purposes such as content generation, assessment support, and curriculum design, others approach these tools with caution, citing concerns about student learning, assessment validity, and academic integrity. Understanding faculty perspectives is therefore essential for informing effective pedagogical strategies and institutional policies. In this paper, we present findings from a focus group study with 29 STEM faculty members at a large public university in the United States. We examine how faculty integrate GenAI into their courses, the benefits and challenges they perceive for student learning, and the institutional support they identify as necessary for effective and responsible adoption. Our findings highlight key patterns in how STEM faculty engage with GenAI, reflecting both active adoption and cautious use. Faculty described a range of pedagogical applications alongside concerns about student learning, assessment, and academic integrity. Overall, the results suggest that effective integration of GenAI in higher education requires rethinking assessment, pedagogy, and institutional governance in addition to technical adoption.