GenAI in Pharmacy Research
Investigating why pharmacy students hesitate to use generative AI for academic research, and what it would take to change that.
Context
Pharmacy students at Thomas Jefferson University use generative AI tools in their research workflows. They summarize papers with it, rephrase drafts, and occasionally explore literature. But they do not trust it.
The gap is not awareness or access. Students know these tools exist and have tried them. The hesitation sits deeper: concerns about accuracy, plagiarism risk, data privacy, and whether AI-generated output holds up to academic scrutiny. Faculty share similar concerns, particularly around critical thinking erosion and student dependency.
This project set out to answer three questions:
The goal was not to design a product. It was to build a research foundation that could inform one: a clear map of behaviors, pain points, and opportunities grounded in primary data from the people who would actually use these tools.
Process
Desk Research
The first step was establishing what was already known. The team analyzed 30+ scholarly papers, articles, and reports covering AI in education, AI in healthcare, and AI-assisted research workflows.
This phase surfaced recurring themes: accuracy anxiety, plagiarism detection gaps, inconsistent citation behavior from AI tools, and a general lack of institutional guidance on acceptable use. It also revealed that most existing research focused on AI in classroom instruction, not on AI as a research tool for students conducting their own literature reviews and analysis.
That gap shaped the rest of the study. The question was not whether AI belongs in education broadly, but whether it can be trusted to support the specific, high-stakes work of pharmacy research.
Competitor Scan
I evaluated five tools that sit at the intersection of AI and academic or healthcare research: DiagnaMed’s Dr. GenAI, Novo Nordisk’s Sophia, Grammarly, Research Rabbit, and Tutor.ai.
The scan focused on three dimensions: what research tasks each tool supported, how it handled source attribution and accuracy, and where it fell short for pharmacy-specific use cases.
Key patterns emerged. Most tools treated “research assistance” as summarization or writing support. None offered meaningful integration with trusted databases like PubMed, NIH, or ScienceDirect. None addressed the plagiarism concern directly. And none allowed students to upload non-text assets like PDFs, images, or datasets for analysis, a limitation that came up repeatedly in later interviews.
Survey
I designed and ran a survey to establish baseline data on chatbot usage frequency, preferred research tools, and initial attitudes toward AI in academic work.
The survey confirmed what the desk research suggested: students were experimenting with AI tools, but usage was shallow. Most interactions were limited to rephrasing sentences or generating rough summaries. Few students used AI for substantive research tasks like literature discovery, data analysis, or citation management.
The survey also identified a clear split in attitudes. Students who had tried AI tools were cautiously interested but lacked confidence in output quality. Students who had not tried them cited plagiarism and accuracy as the primary barriers.
Interviews
The team conducted 12 semi-structured interviews with pharmacy students at Thomas Jefferson University. I personally conducted 3 of those interviews. Separately, I led the stakeholder interview with the Department Chair of Pharmacy, with the rest of the team taking notes.
The student interviews followed a consistent protocol designed to surface actual behavior, not hypothetical preferences. Questions focused on current research workflows, specific moments where AI was used or considered and rejected, and the criteria students applied when deciding whether to trust a source.
The stakeholder interview added an institutional lens. Faculty valued innovation but worried about three things: students losing the ability to critically evaluate sources on their own, sensitive research data being exposed to third-party AI platforms, and the lack of clear institutional policies on AI use in academic work.
After all interviews, the team used card sorting to cluster findings into four categories: Needs, Goals, Pain Points, and Key Insights.
Findings
How students actually use AI
Students rely on established, peer-reviewed databases for primary research: PubMed, NIH, ScienceDirect, and Google Scholar. AI tools are used at the margins, primarily for rephrasing and summarization. Students do not use AI for source discovery or literature review because they cannot verify whether AI-surfaced sources are real, current, or peer-reviewed.
The trust issue is not abstract. Students described specific experiences where AI-generated citations turned out to be fabricated or pointed to retracted papers. That was enough to push AI into a supplementary role rather than a primary one.
What is not working
Three structural barriers surfaced consistently across interviews:
First, free-tier AI tools impose significant limitations. Students cannot upload PDFs, images, or datasets, which means the tools cannot engage with the actual materials students work with daily.
Second, data privacy is a real concern, not a hypothetical one. Students working with sensitive health-related research data have no clarity on whether their inputs are stored, used for training, or accessible to third parties.
Third, the plagiarism risk is unresolved. AI tools do not flag when their output overlaps with existing published work. Students are left to run separate plagiarism checks manually, which creates friction and erodes trust.
What would build trust
Students articulated a clear wish list, and it was more specific than “make AI better.” They want AI tools that pull from verified, peer-reviewed databases (PubMed, NIH) rather than general internet sources. They want built-in plagiarism detection so they can trust the output without a separate verification step. They want the ability to upload PDFs, images, and datasets for analysis, not just text input. And they want explicit data privacy guarantees: confirmation that their queries are not stored or used for model training.
The Department Chair reinforced several student-side findings but added a critical institutional dimension. Faculty are not opposed to AI adoption. They are opposed to unguided AI adoption. Without clear policies, training, and tools that preserve critical thinking rather than replace it, faculty see AI as a risk to academic rigor rather than an accelerant.
Pharmacy students hesitate to adopt GenAI chatbots for research because existing tools fail on the three dimensions that matter most in academic work: accuracy verification, data privacy, and plagiarism prevention.
Reflection
What I learned
This project reinforced that the most useful research output is often a well-defined problem, not a solution. The instinct in UX work is to push toward design recommendations and prototypes. But the value here was in documenting the trust gap with enough specificity that a product team could act on it without repeating the discovery work.
I also learned the practical difference between leading an interview and observing one. Conducting the stakeholder interview with the Department Chair required real-time judgment about when to probe deeper and when to let a response sit. That is a different skill from synthesizing transcripts after the fact.
What I’d do differently
The study included only pharmacy students at one university. A stronger study would recruit across multiple institutions and include graduate researchers alongside undergraduates, since their AI usage patterns and risk tolerances likely differ.
The project stopped at problem definition. If I were to continue, the next step would be a concept validation phase: low-fidelity prototypes of a research-specific AI tool tested against the specific trust barriers identified in the findings. The research laid the groundwork for that, but the work itself remains undone.
I would also push for a diary study component alongside the interviews. Interviews capture recalled behavior, but a diary study would surface in-the-moment decisions about when students reach for AI and when they pull back. That real-time data would add a layer the interview data could not.