🔬 physics

Faculty Orientations Shape Adoption of AI in Research and Teaching

This mixed-methods study of 90 STEM faculty reveals that a specific "AI pedagogical orientation"—defined by beliefs about AI's role in disciplinary thinking and expertise development—is a stronger predictor of AI adoption in research and teaching than institutional context, demographics, or general attitudes, suggesting that existing technology-adoption models may be insufficient for contexts where AI directly interacts with knowledge production.

Original authors: Timothy J. Atherton, Ian Descamps, Tova R. Holmes, Christina L. Vizcarra, Ning Sui, Max Webel, Jay J. Foley IV

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Timothy J. Atherton, Ian Descamps, Tova R. Holmes, Christina L. Vizcarra, Ning Sui, Max Webel, Jay J. Foley IV

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of 90 science professors (from physics, chemistry, and astronomy) who are known for being early adopters of new teaching methods. They are like the "tech-savvy neighbors" of the academic world. Researchers asked them a simple question: "How are you using Artificial Intelligence (AI) in your work, and why?"

The paper's main discovery is that the answer isn't about whether they have access to AI or if they think it's "cool." Instead, it's about how they view the role of AI in the thinking process itself.

Here is the breakdown of their findings using simple analogies:

1. The "Tool vs. Partner" Divide

The researchers found that professors fall into two main camps, not based on whether they like AI, but on how they think it fits into their mental workflow.

The "AI as a Thinking Partner" Group:
These professors see AI as a co-pilot or a sparring partner. They use it to brainstorm ideas, debug code, or draft outlines. To them, AI is like a calculator for complex thoughts. They believe it can handle the "boring" parts of thinking so they can focus on the deep, creative parts. Even though they worry AI might make mistakes, they see those risks as something to manage, not a reason to stop using it.
- The Metaphor: They treat AI like a gym spotter. The spotter helps you lift heavier weights (do more work) safely, but you still have to do the lifting.
The "AI as a Shortcut" Group:
These professors see AI as a crutch or a fast-forward button that skips the necessary struggle of learning. They worry that if students (or they themselves) use AI to get answers quickly, they miss out on the "deep neural networks" of the brain that only fire up during slow, difficult thinking. To them, AI is like ordering a meal instead of cooking it; you get the result, but you don't learn the recipe or the skill.
- The Metaphor: They treat AI like a cheat code in a video game. It gets you to the end, but you never actually play the game or build the skills needed to win on your own.

2. What Didn't Matter (The Surprising Part)

The researchers expected that certain things would drive adoption, but they found these factors were surprisingly weak:

Institutional Rules: Whether the university had a fancy new AI policy or a degree program didn't really change how much a professor used AI.
Access: Having the "best" tools or knowing where to find them wasn't the deciding factor.
Demographics: It didn't matter if the professor was young or old, worked at a small college or a big research university, or taught physics vs. chemistry.

The Analogy: Imagine two people in the same gym. One has a personal trainer, a brand-new machine, and a membership card (Institutional Support). The other has an old bench and no trainer. You might expect the first person to work out more. But in this study, the person who worked out the most was simply the one who believed working out was essential to their identity, regardless of the equipment they had.

3. The "Orientation" Concept

The paper coins a term called "AI Pedagogical Orientation." Think of this as a compass that points in one of two directions:

North (Pro-Integration): "AI is a tool that helps us think better if we use it carefully."
South (Pro-Caution): "AI is a tool that might stop us from thinking deeply if we rely on it too much."

The study found that this compass was the strongest predictor of whether a professor actually used AI in their research or teaching. If your compass pointed North, you used AI. If it pointed South, you didn't, even if you had access to the same tools as the North-pointers.

4. The "Middle Ground" Reality

It's important to note that nobody was 100% for or 100% against.

Even the professors who used AI the most were worried about it. They knew it could lie (hallucinate) or make students lazy. They just decided to use it carefully anyway.
Even the professors who used it the least admitted it had some potential uses, but they felt the risks to deep learning were too high.

The Bottom Line

The paper concludes that we can't just hand professors AI tools and say, "Here, use this!" or pass a rule saying, "You must use this."

The real driver is philosophy. Before a professor will use AI, they have to answer a deep question: "Does this tool help us think like scientists, or does it replace the thinking process entirely?" Their answer to that question determines everything else.

Technical Summary: Faculty Orientations Shape Adoption of AI in Research and Teaching

Problem Statement
Despite the widespread availability of Large Language Models (LLMs) in higher education, faculty adoption and usage vary significantly. Existing literature documents this variation and the coexistence of multiple interpretations of AI (as a tool, a knowledge source, or a threat), yet current technology-adoption models fail to fully explain why instructors respond differently to similar affordances. Traditional frameworks, such as Rogers' Diffusion of Innovations (DoI) and the Unified Theory of Acceptance and Use of Technology (UTAUT), emphasize external conditions, social influence, and perceived utility. However, these models may be insufficient in contexts where technologies interact directly with disciplinary reasoning and knowledge production. The authors posit that variation in AI use cannot be explained solely by access, information sources, or general attitudes, suggesting a need to investigate how instructors conceptualize the epistemic role of AI within their specific disciplines.

Methodology
The study employed a mixed-methods survey design targeting a specific population of 572 STEM faculty members who are awardees of the Research Corporation for Science Advancement (RCSA) Cottrell scholarship. This group was selected as a "targeted population" likely to be actively experimenting with emerging technologies due to the dual selection criteria of their awards (scientific research quality and innovative education plans).

Data Collection: Conducted in Fall 2025, the survey achieved a 20% response rate (116 responses), with 90 responses retained for analysis after applying a 50% completion threshold.
Instrument Design: The survey captured AI usage patterns across research, teaching, and other professional tasks; sources of tools; institutional initiatives; attitudinal statements (Likert-scale); and demographic data. Items were informed by DoI and UTAUT constructs.
Quantitative Analysis: The authors performed an Exploratory Factor Analysis (EFA) on attitudinal and contextual variables to identify latent structures. A mixed correlation matrix (tetrachoric, polychoric, and polyserial) was used to handle binary and ordinal data. Logistic regression and correlation analyses were subsequently used to validate the identified factor against self-reported usage.
Qualitative Analysis: Open-ended survey responses were analyzed using an iterative, emergent coding procedure. Initial codes were consolidated into four higher-order constructs: Practice, Epistemic, Affective, and Structural. Inter-rater reliability was assessed using Krippendorff's $\alpha$ .
Integration: Quantitative factor scores were correlated with the consolidated qualitative codes to examine how reasoning and experiences align with the identified latent construct.

Key Contributions and Results
The primary contribution of the study is the identification of a coherent latent construct termed "AI pedagogical orientation."

Factor Identification: EFA revealed a unidimensional 9-item scale explaining 55% of the variance in attitudinal data. This factor strongly predicts self-reported AI use across research, teaching, and other professional activities (Odds Ratios ranging from 2.8 to 5.0).
Nature of the Construct: The "AI pedagogical orientation" does not merely measure positive or negative sentiment. Instead, it reflects differing views on the role AI should play in disciplinary thinking, learning, and expertise development.
- High Orientation: Associated with viewing AI as a tool to support disciplinary thinking, delegate ancillary tasks, and increase efficiency. These respondents integrate AI into workflows while managing risks through critical evaluation.
- Low Orientation: Associated with viewing AI as a potential threat to deep learning and the development of expertise. These respondents often limit use or reject LLMs in favor of traditional methods, prioritizing effort and deliberate engagement over efficiency.
Weakness of Traditional Predictors: Contrary to standard adoption models, institutional initiatives, demographic variables (career stage, institution type), and information sources showed comparatively weak or non-significant associations with AI use. While access to tools is a component of the orientation, the conceptualization of the tool's role is the dominant predictor.
Qualitative Corroboration: Qualitative analysis confirmed that the orientation is defined by epistemic boundaries. High-orientation respondents framed AI as an "interactive cognitive partner," whereas low-orientation respondents framed it as a source of "superficial engagement" or "overconfident incorrectness." The distinction lies not in the recognition of risks (both groups acknowledge them) but in whether those risks are viewed as reasons to avoid the tool or as constraints to be managed.

Significance and Claims
The authors claim that existing technology-adoption models may be incomplete for technologies that interact directly with knowledge production and disciplinary reasoning. The study suggests that in the context of AI, adoption is driven less by external factors (access, policy, training) and more by internal, epistemic interpretations of how the technology fits into the practice of "doing science" and teaching.

The paper modestly concludes that while factors like perceived usefulness and institutional support remain relevant, they do not fully explain adoption variation within this pedagogically engaged population. The findings imply that efforts to support AI adoption must move beyond providing access and technical training to helping faculty articulate how AI might support (or undermine) disciplinary thinking. The authors note that their sample is specific (Cottrell awardees) and represents a snapshot in time; thus, the relationships observed may evolve as institutional structures mature and adoption becomes more widespread. Future work is suggested to examine how these orientations develop over time and how they transfer to student learning environments.

1. The "Tool vs. Partner" Divide

2. What Didn't Matter (The Surprising Part)

3. The "Orientation" Concept

4. The "Middle Ground" Reality

The Bottom Line

More like this