Abstract
Generative AI tools are increasingly embedded in the scientific publication pipeline, from hypothesis synthesis to abstract drafting to peer review assistance. This integration creates liability vacuums that existing scholarly publishing norms were not designed to address. When a paper containing AI-generated methods text leads to a retraction, or when a fabricated citation sourced from a large language model propagates into downstream research, current legal and editorial frameworks offer no clear mechanism for assigning responsibility. This piece maps those gaps and evaluates emerging proposals to close them.
How the Pipeline Has Changed
Researchers at institutions from MIT to ETH Zurich now routinely use tools such as Elicit, Semantic Scholar’s AI layer, and direct API access to GPT-4-class models to accelerate literature synthesis, generate structured summaries of retrieved papers, and draft manuscript sections. A 2024 survey by the Committee on Publication Ethics (COPE) found that 38 percent of corresponding authors at major journals reported using generative AI for some stage of manuscript preparation, while fewer than 15 percent of journals had disclosure requirements specific enough to capture that use.
The core liability problem is that authorship in academic publishing carries legal weight in some jurisdictions, particularly for clinical research. If AI-generated content contributes to a flawed study that influences clinical practice, the responsible party is unclear. The journal? The authors who used the tool? The tool provider? Existing software liability frameworks, which generally shield providers under terms-of-service disclaimers, were not written with scientific publishing in mind.
The Citation Fabrication Vector
A specific and already-documented harm is the propagation of hallucinated citations. A 2024 preprint from Stanford’s RegLab identified 1,247 papers indexed in PubMed Central containing references to articles that did not exist as described, with DOIs either absent or pointing to unrelated work. Of those, at least 94 had already been cited by other published papers, beginning a laundering process where a fabricated source acquires apparent legitimacy through secondary citations.
Current retraction infrastructure, managed by organizations such as Retraction Watch and CrossRef’s Retraction Notice Metadata, operates reactively and relies on post-publication human review. It has no mechanism to flag AI-probable fabrications proactively. CrossRef’s Event Data API tracks citation velocity but not citation integrity.
Closing the Accountability Gap
Three interventions deserve attention. First, mandatory structured disclosure of AI tool use at the section level, modeled on the author contribution taxonomy (CRediT), would make AI involvement auditable without banning legitimate assistance. Second, publishers should require cryptographic provenance records for citations, using infrastructure already developed for reference verification by projects including the Initiative for Open Citations (I4OC). Third, jurisdictions with research integrity statutes should clarify that AI-generated factual errors in funded research do not automatically transfer liability to the tool provider when the human author had reasonable means to verify the claim.
None of these proposals eliminate the broader epistemological challenge of AI-assisted science. They create accountability surfaces where none currently exist.