American Journal of Health-System Pharmacy · 6 Nov 2026
Can Generative AI Crush It? A Systematic Comparison of ChatGPT Plus (GPT-4o) and Google Gemini Pro vs. Human Expert Analysis for Oral Dosage Form Crushability Analysis
Vivek S Dave, Christina Tzikas, Fang Zhao, Anne Schweighardt
Citations by database
Each database indexes a different set of journals, preprints and books, so they are shown side by side rather than blended.
- OpenCitations0
- Crossref0
Citations per year
The current year is still counting.
- 2026 so far0
Abstract
Abstract Purpose To systematically evaluate ChatGPT Plus (GPT-4o) and Google Gemini Pro as decision-support instruments for the crushability analysis of immediate-release (IR) oral solid dosage forms, benchmarked against a human expert reference using a validated four-question analytical checklist. Methods Sixteen IR oral solid dosage forms approved by the Food and Drug Administration (FDA) in 2024 were independently assessed for crushability (YES/NO) by a human expert team and each AI tool using an identical structured prompt. Evaluator reasoning was coded across four analytical domains. Inter-rater agreement across all three evaluators was quantified using Fleiss kappa (κ) with a 95% confidence interval (CI) and two-tailed z-test; kappa values were interpreted using the Landis and Koch scale. All computations were performed in Microsoft Excel. Results Human experts classified 11 of 16 products (68.8%) as crushable and 5 (31.3%) as not crushable. Eight products achieved unanimous three-way agreement. The Fleiss κ was 0.329 (Fair; 95% CI, 0.046-0.612; z = 2.28; p = 0.023), indicating statistically significant agreement above chance. Both tools substantially over-identified pharmacokinetic concerns relative to human experts (Gemini, 11/16; ChatGPT, 8/16; Human, 3/16), the primary driver of disagreement. Gemini's identification of QTc interval prolongation risk for mavorixafor prompted expert revision of the crushability recommendation to not crushable. ChatGPT's false positive recommendation for lazertinib, failing to detect a reactive acrylamide degradation pathway, represents the highest-risk error in the cohort. Conclusion Generative AI demonstrated statistically significant but fair overall agreement with human expert analysis. Systematic scope deviations, including out-of-scope pharmacokinetic reasoning and hallucination-adjacent misclassification, suppressed concordance and produced clinically consequential errors. These findings support AI as a structured first-pass screening tool, contingent on pharmacist-led verification, institutional governance, and heightened scrutiny for hazardous drugs and novel formulation-technology products.
What the data shows
- No database records a citation yet, as is usual for a paper this new.
Graph
This paper in the middle, its authors and their most-cited papers below, the papers it cites to the left and the papers citing it to the right. The timeline lays every paper out by year.
Select a node for its details. Drag to pan, scroll to zoom, drag a node to move it.
Start with
Cited by 0 papers
Who cites this paper, from open citation data, with Crossref's record of each; the 100 newest when there are more.
References (0)
What it cites, from open citation data, with Crossref's record of each.