Article
Generative Artificial Intelligence in Academic Research: Opportunities, Ethical Dilemmas, and Governance Pathways
Generative artificial intelligence has moved into the academic world faster than academic norms, and so far the institutional reaction has been split into two moves: first, stating that such systems cannot be authors; and second, trying to detect the presence of such systems. This paper suggests that both responses are symptomatic responses and do not tackle the underlying structural problem. In scientific practice, authorship includes four elements which, historically, have gone together: intellectual contribution, accountability for correctness, credit, and provenance of warrant. Generative systems separate them, and the ban on machine authorship only addresses the accountability aspect, not the other parts of the problem, such as contribution, credit and provenance. We formalise this as the attribution decoupling problem and place it in the context of the warrant chain, a series of traceable human judgements that make a scientific claim answerable. We then argue that a system's governance should be calibrated for its use, on top of its use, and present two dimensions of a task-risk gradient: verifiability of the task's output by the researcher and closeness of the task to the scientific claim. The current uniform disclosure rules are seen as being both too broad for low risk uses and too narrow for high risk ones. We also claim that the use of detection-based enforcement is a dead end as it is impossible to have reliable detection equipment, and its errors are systematically distributed among researchers who write in a second language, making it an instrument of discrimination. Another approach is a governance based on process, declaration, provenance and verification, which takes place at four institutional levels. There are no invented statistics reported, seven propositions are provided and an analytical protocol.