Cureus. 2026 Jun;18(6):
e111207
Background Artificial intelligence (AI), particularly large language models (LLMs) such as ChatGPT, is increasingly being used for statistical interpretation and research support. While traditional statistical software such as IBM SPSS remains the gold standard for transparent and reproducible analyses, concerns persist regarding the accuracy, consistency, and reproducibility of LLM-generated statistical outputs. Given the continuous evolution of LLMs, replication studies are needed to evaluate whether newer models demonstrate improved statistical reliability and alignment with established analytic standards. Methodology In total, 14 statistical procedures were applied to real datasets that previously generated peer‑reviewed, published scientific articles. The analyses encompassed descriptive statistics, Pearson product-moment correlation coefficient (Pearson r), multiple correlation using Pearson r, Spearman's rho, simple linear regression, one-sample and paired t‑tests, two independent‑sample t‑tests, multiple linear regression, one-way analysis of variance (ANOVA), repeated‑measures ANOVA, two‑way (factorial) ANOVA, and multivariate analysis of variance (MANOVA). Datasets were collected within a systematically defined timeframe (2012-2023), ensuring temporal consistency and representativeness. All analyses were executed by copying and pasting prompts into GPT-5.5, accompanied by the corresponding SPSS 31.0 variable names copied and pasted directly from the original SPSS datasets. Results Results demonstrated concordance between SPSS 31.0 and GPT-5.5 across most descriptive and inferential statistical procedures, although notable discrepancies did occur. Identical or near-identical findings were observed for descriptive statistics, Pearson correlations, one-sample t-tests, paired t-tests, two independent-sample t-tests, simple and multiple linear regression, and one-way ANOVA. Correlation coefficients, effect sizes, confidence intervals, and significance values were mostly consistent across platforms. Discrepancies emerged in selected nonparametric analyses, factorial ANOVA, MANOVA, and repeated-measures ANOVA procedures. Despite these computational differences, the overall substantive interpretations and statistical conclusions remained largely consistent between SPSS 31.0 and GPT-5.5, but not perfectly. These findings suggest GPT-5.5 demonstrates some statistical consistency with traditional statistical software for many common analytic procedures, while more advanced multivariate analyses may still require independent verification using established statistical platforms. Conclusions GPT-5.5 demonstrated considerable agreement with SPSS 31.0 across many commonly used statistical procedures, including descriptive statistics, correlations, regression analyses, t-tests, and one-way ANOVA. However, more notable discrepancies emerged in several analyses, particularly more complex procedures such as factorial ANOVA, MANOVA, and repeated-measures ANOVA. These findings suggest that GPT-5.5 may serve as a useful adjunct for statistical interpretation and exploratory research, but AI-generated statistical outputs should be interpreted with caution. Validated statistical software and appropriate methodological oversight remain essential for confirmatory analyses and research involving clinical decision-making.
Keywords: ai-assisted statistics; alignment; chatgpt; ibm spss; large language models