SAGE Splits AI Disclosure Into Two Tiers, and Most Authors Only Need the First

SAGE’s policy is organised around a single distinction that does most of the work: did the tool merely improve wording you already wrote, or did it generate content that shapes what the paper actually claims?

Tier one, no disclosure. “Assistive tools that suggest improvements to your own work” — for language, grammar or structure — need no statement. The author stays responsible for accuracy and rigorous scholarship either way.

Tier two, disclosure required. Generative AI that produces content “directly affecting methodology, analysis, results, or conclusions” must be disclosed upon submission. SAGE names specific uses that fall here: literature assistance, translation, AI-generated software or code, data visualisation help, reference compilation, and generating illustrations or infographics.

Sources and verification. Authors are told to cite original sources rather than the AI tool as a primary reference, and to check every AI output against those original sources, because large language models “can produce false content or nonexistent citations”.

Images. Citing a GenAI tool is appropriate where it created an image or visualisation presented as representative rather than as original research data; presenting an AI-generated image as a unique or novel research image is specifically prohibited.

Peer review and editorial decisions. Reviewers may use GenAI to help write and format a report but remain responsible for its content, and must not generate the report itself with AI, including by uploading or pasting manuscript text into a tool. Misuse can mean a reviewer is not invited to review again, and their review may be excluded from the decision. Editors “must not use ChatGPT or other GenAI to generate decision letters” or to summarise unpublished research.

A rule built around what the tool did, rather than whether a tool was used, scales much better than a flat declare-everything rule once AI assistance is built into ordinary writing software. Requiring a statement for every spellcheck would make the disclosure meaningless through sheer volume; drawing the line at content that shapes the paper’s actual claims keeps the disclosure pointed at what a reader needs to know.

The source-citation rule — cite the original, not the AI tool — answers a specific failure mode directly: a model can summarise or paraphrase a source accurately most of the time and invent one occasionally, and an author who cites the model rather than checking the underlying source has no way to catch the invented case before a reader does.

Naming consequences for reviewer misuse, rather than only describing the rule, is what makes a confidentiality policy function in practice: a reviewer who understands a misused AI tool can cost them future review invitations has a concrete reason to check before pasting text into one.

How it reaches a specific journal. SAGE publishes the policy at group level and applies it across its journals, including the many society journals it publishes on behalf of other organisations; a society may add further requirements on top.

For an author moving between SAGE and another publisher on different projects, the practical point to hold onto is the tier structure itself: other publishers in this set ask for disclosure of AI use generally, where SAGE asks you to first classify what the tool did.