Scientific Publishers Set Clear Rules for AI Writing Tools

The new boundaries for AI in research papers

Scientific publishers are drawing firm lines around artificial intelligence as language models become commonplace writing aids. The American Society for Biochemistry and Molecular Biology has implemented specific policies governing how researchers may—and may not—use AI tools when preparing manuscripts for its journals.

The core restriction: AI cannot be listed as an author. ASBMB editorial policies require that authors remain "accountable for all aspects of the work" and capable of resolving questions about accuracy or integrity. Because AI systems simply generate output based on programming patterns, they cannot accept responsibility for errors or defend research findings.

Researchers may use AI to refine grammar or improve stylistic clarity in existing text, particularly valuable for scientists writing in a second language. But the technology may not draft any portion of a manuscript from scratch.

Mandatory disclosure requirements

Authors who employ AI assistance must include a dedicated disclosure section titled "Declaration of generative AI and AI-assisted technologies in the writing process." This section must name the specific tool used and state that authors have reviewed the output and accept full responsibility for the content. The disclosure cannot appear in acknowledgments or author lists—it requires separate, explicit documentation aligned with Elsevier publication standards.

For images, AI tools generally cannot create, manipulate, or refine figures unless the AI-generated images themselves constitute experimental results. A study using deep learning to enhance medical imaging, for example, may include those AI-produced images as data, provided authors describe the generation process in reproducible detail.

Why it matters

The proliferation of AI-generated errors threatens the foundation of scientific publishing. An estimated one in 277 biomedical papers now contains fabricated references—complete with convincing but nonexistent DOIs that slip past peer review. Because citations function as research data, fabricating them may constitute research misconduct. Meanwhile, "tortured phrases" produced when AI text runs through plagiarism-avoidance software yield absurdities like "bosom peril" for breast cancer or "polymerase chain response" for polymerase chain reaction. These distortions erode clarity and credibility in technical communication.

The hallucination problem

Language models generate text by predicting probable next words based on training data patterns. When a query falls outside that training scope, the model still produces an answer—it fills gaps with plausible-sounding fabrications rather than acknowledging knowledge limits. This "hallucination" behavior represents a fundamental feature unlikely to disappear from current AI architectures.

Publishers cannot reliably detect AI-generated text through automated means, placing the burden of responsible use squarely on researchers. As AI tools continue reshaping scientific writing workflows, maintaining research integrity requires both clear institutional guidelines and careful human oversight at every stage.

These details were first reported by the American Society for Biochemistry and Molecular Biology in ASBMB Today.