Artificial Intelligence in the Research Lifecycle: From Idea Generation to Scholarly Publication

Abstract

Artificial intelligence (AI) has moved from a peripheral instrument of computational research to an infrastructure that touches nearly every stage of scholarly work. This paper offers an integrative, stage-based analysis of how contemporary AI systems—particularly large language models (LLMs), retrieval-augmented systems, and tool-using agents—are reshaping the research lifecycle, from the earliest formulation of a research question through data analysis, manuscript preparation, peer review, and post-publication dissemination. We propose an eight-stage framework and, for each stage, characterise the dominant AI capabilities in use, the empirical evidence for their benefits, the observed failure modes, and the state of institutional governance. Three cross-cutting findings emerge. First, capability is unevenly distributed across the lifecycle: AI performance is strongest where outputs are cheaply verifiable (code, language, retrieval, structured extraction) and weakest where judgement, tacit knowledge, and problem selection dominate. Second, the principal risks are epistemic rather than merely technical—citation fabrication, homogenisation of research questions, illusions of understanding, and the displacement of accountability—and these risks compound when AI is used at multiple consecutive stages without independent verification. Third, governance has converged rapidly on a narrow consensus (no machine authorship, mandatory disclosure, confidentiality in peer review) while leaving the harder questions of methodological validity and evaluative use largely unresolved. We conclude with a five-level maturity model for responsible institutional adoption and an agenda for research on verification, provenance, and the measurement of AI-attributable scientific value