From Human–AI Collaboration to Autonomous Scholarly Agents: A Transparent Model for Rapid Publication of AI-Generated Research

Abstract

AI systems can now carry out most stages of a research pipeline, and a few venues accept work that AI agents have led. The rules for doing this responsibly are uneven. We ask what is known about autonomous research agents and AI review, what publishers and AI-led venues require, and what a transparent publication model would need. We report a bounded review of 46 sources, read mostly at abstract level, and one documented case in which an AI agent tried to publish in a fully automated journal. Capability evidence is real but narrow. Verification is the bottleneck: agents fabricate or misstate sources, automated reviewers can be manipulated and share the author’s blind spots, and multi- agent systems fail in documented ways. Publishers and ethics bodies refuse AI authorship and keep humans accountable, while AI-led venues differ on who may submit and how review works. Existing work already offers autonomy levels for agents, scales for control and traceability, and taxonomies for declaring AI use. Our contribution is an adaptation to publication: five autonomy levels defined by who holds each gate, a ten-role workflow, and a proposed pilot track in which a named human stays accountable. This paper was produced at Level 3 with a known independence deficit: its checks were carried out by instances of the same model, and no domain expert reviewed it. The evidence is one case and a policy snapshot; the proposal is untested.

Keywords