The first data used was a set of 51 published and unpublished journal articles written by the author, as used in previous studies (Thelwall, 2025). Each article has an associated quality score from the author, using the Research Excellence Framework (REF) scale of 1* (recognised nationally), 2* (recognised internationally), 3* (internationally excellent) and 4* (world leading) for originality, significance, and rigour. Articles were given fractional scores (1.5*, 2.5* or 3.5*) when they fell between two thresholds. These scores are private.

The data used for the last two research questions was the collection of up to 200 journal articles from the highest and lowest scoring departments in each of the 34 UoAs in REF2021, as used in a previous study for ChatGPT (Thelwall & Yaghi, 2024). Although each article is individually scored in the REF by two experts, with scores norm-referenced within and between UoAs, only aggregate departmental scores are published. Each article was therefore associated with the departmental mean REF score for its UoA. Assuming that there are only weak departmental biases in the system, this substitution of actual article quality scores with a departmental proxy would have a damping effect on any correlations with potential quality indicators, reducing their magnitude but not changing their sign.

Each article was uploaded to Gemini 1.5 through its API. The main versions at the time of the experiment (October 2024) were: 1.5 Pro (gemini-1.5-pro-002), the “best performing” multimodal model; 1.5 Flash (gemini-1.5-flash-002) a cheaper variant “balancing performance and cost”; and 1.5 Flash-8B (gemini-1.5-flash-8b-001), the “fastest and most cost-efficient” (ai.google.dev/gemini-api/docs/models/gemini). Gemini 1.5 Flash was chosen because the more expensive alternative may not be practical for research evaluations. It is equivalent to ChatGPT 4o-mini, the cut down version of ChatGPT 4o used in most previous similar studies.

The system instructions used were as previously used for ChatGPT (Thelwall, 2025). They are the REF2021 scoring guidelines for assessors in Main Panel D (arts and humanities but including information science) for the first dataset and the appropriate guidelines for the second dataset (UoAs 1-6: Main Panel A, health and life sciences; UoAs 7-12: Main Panel B, physical sciences and engineering; UoAs 13-24: Main Panel C, social sciences; UoAs 25-34: Main Panel D, mainly arts and humanities). Article scores were extracted from Gemini reports with an extended version of the software designed to extract scores from ChatGPT reports for the same task (Webometric Analyst, AI menu; ChatGPT: Extract scores from reports; github.com/MikeThelwall/Webometric_Analyst).

As in the previous studies, the main statistical test was correlation between the Gemini scores and the article quality scores (first dataset) or departmental quality proxy for article quality (second dataset). Correlation is the most appropriate test because the results can easily be scaled (Thelwall, 2025) if Gemini tends to score high or low. Spearman correlations were used since the data are fundamentally ranks and for comparability with previous papers.

The maximum Spearman correlation with author scores for Gemini 1.5 Flash is 0.645 for PDF and 0.549 for truncated text (no references or tables), after 30 iterations (Figure 1). The correlations are lower than previously found for ChatGPT in all cases except Full text which was not checked in ChatGPT, and PDF, for which the ChatGPT 4 correlation was 0.509 after 15 iterations (Thelwall, 2025) (the Gemini 1.5 Flash correlation was 0.613 after 13 iterations). Thus, Google Gemini may be relatively better than ChatGPT at analyzing long documents, including full text. The differences are not statistically significant, however. For example, a bootstrapped 95% confidence interval for the PDF Spearman correlation is (0.38, 0.81). This should be interpreted as the likely range of values for the correlation for a similar but different set of articles. Comparing ChatGPT 4o-mini with Gemini 1.5 Flash, the optimal strategy found so far is using Gemini 1.5 Flash with PDFs, at least in terms of the highest correlation with article quality scores. Of course, this is only a tentative conclusion, given the small sample size and small correlation differences involved.

For the REF data, Gemini 1.5 Flash correlations with departmental mean REF2021 scores tend to be similar in magnitude to the correlations for ChatGPT 40-mini. Nevertheless, the overall mean of the correlations is marginally higher for ChatGPT 4o-mini (rho=0.409) than for Gemini 1.5 Flash (rho=0.399) (Figure 2). In both cases, the correlations are relatively weak for most arts and humanities but strong in all health, life and physical sciences and some social sciences. Clinical Medicine has an anomalously low Gemini correlation, but it is positive, unlike the corresponding ChatGPT correlation.