Generative AI in Scientific Research

Abstract

Generative AI is reshaping the scientific research process across its entire lifecycle -- from hypothesis generation and experimental design, through data analysis and interpretation, to manuscript preparation and peer review. This transformation is uneven: some research tasks (literature synthesis, code generation, figure description) have been substantially augmented by current AI capabilities, while others (genuine hypothesis novelty, causal inference from observational data, experimental reproducibility) remain primarily human-driven. This study evaluates generative AI contributions across seven research lifecycle stages using nine AI tools -- GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Perplexity Research, Elicit, Semantic Scholar API, AlphaFold3, BiomedGPT, and a proposed Research Orchestration Pipeline (ROP) -- benchmarked across four scientific domains: molecular biology, climate science, materials science, and AI/ML research. Evaluation metrics include task completion quality (expert ratings), factual accuracy and hallucination rate, time-to-output, and reproducibility. Elicit achieves the best literature review task performance (F1 0.884 for relevant paper identification). AlphaFold3 achieves median TM-score 0.924 for protein complex structure prediction. GPT-4o leads on manuscript drafting quality (3.84/5 expert rating) but shows 18.4% hallucination rate for specific citations. ROP reduces end-to-end research task completion time by 42.4% vs. baseline with human researchers. A responsible AI integration framework for research institutions is proposed.