What is meant by AI citation and disclosure?
AI citation covers citing AI-generated content as a source and disclosing AI tool use during manuscript creation.
The first practice, treating AI output as a citable source, is widely discouraged. The International Committee of Medical Journal Editors states that AI-generated material should not be referenced as the primary source. The reason is straightforward: large language models produce text that can sound authoritative but may be incorrect, incomplete, or biased. A citation implies the reader can verify the claim by consulting the source. With AI, there is no stable, retrievable document to consult, and the output can change between queries.
The second practice, disclosure of AI use, is what most current guidance focuses on. Disclosure means describing in your manuscript or cover letter which AI tools you used, how you used them, and to what extent. This is not the same as listing the tool as an author. It is a transparency measure, similar to declaring a conflict of interest or acknowledging a funding source.
Examples help clarify the distinction. If you use ChatGPT to generate a first draft of a paragraph, you must disclose that use. If you use an AI image generator to create a figure, you must disclose it. If you use an AI tool to check grammar or improve phrasing, many publishers now consider that basic copy editing and do not require disclosure, though practices vary. The key question is whether the AI contributed substantively to the content or structure of the work.
What do major standards and journals require about citing or disclosing AI?
The International Committee of Medical Journal Editors published recommendations in 2023 that have become a reference point across disciplines. The ICMJE requires authors to disclose at submission whether they used AI-assisted technologies such as large language models, chatbots, or image creators in the production of submitted work. Authors who use such technology must describe, in both the cover letter and the submitted work, how they used it.
The ICMJE is explicit: chatbots such as ChatGPT should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work, and these responsibilities are required for authorship. Authors must be able to assert that there is no plagiarism in their paper, including in text and images produced by AI. Humans must ensure there is appropriate attribution of all quoted material, including full citations.
Publishers and editorial bodies are moving from detection to disclosure. The Scholarly Kitchen reports that publishers such as Wiley are updating their AI guidance and preparing more detailed author guidance following recommendations from the Committee on Publication Ethics and the STM Association. Wiley updated the AI use section of its general guidance for journal authors in March 2025, and has conducted extensive author interviews as the basis for more detailed guidance to be released later in 2025.
The shift is significant. Early in 2023, some journals experimented with AI detection tools to identify undisclosed use. By 2025, the consensus had moved toward requiring authors to self-report. Detection tools proved unreliable, and the focus shifted to establishing clear disclosure norms and enforcement mechanisms. The Committee on Publication Ethics updated its retraction guidelines in August 2025 to explicitly include undisclosed involvement of artificial intelligence as grounds for retraction, alongside paper mills, identity theft, and fraud.
What remains unsettled is the format of disclosure. The ICMJE and most publishers require disclosure but do not prescribe a single, standard citation format for LLM output. Practices vary across publishers and are still emerging. Some journals ask for disclosure in the methods section, others in the acknowledgements, others in a dedicated statement. The common thread is that the disclosure must name the tool, describe how it was used, and indicate the extent of use.
How can provenance and watermarking support AI citation?
Provenance refers to whether a piece of content was generated by AI or a human. The National Institute of Standards and Technology's AI Risk Management Framework recommends maintaining provenance of training data and supporting attribution of AI decisions to subsets of training data to improve transparency and accountability. NIST's guidance is aimed at AI system developers, but the principle applies to content as well.
One proposed solution for signalling whether content was AI-generated is provenance recorded via a non-fungible watermark or cryptographic signature that is permanently associated with the piece of content. The idea is that a reader or platform could check the watermark to verify the origin of a piece of text or an image. This approach has been discussed in policy circles, including by the Federation of American Scientists, as a way to reduce the information integrity risks posed by synthetic content.
The technical feasibility and privacy trade-offs of provenance systems remain unresolved. NIST and the Federation of American Scientists advocate provenance and cryptographic watermarking, but sources do not resolve implementation trade-offs such as privacy, interoperability, resistance to tampering, or a timeline for standard adoption. Watermarking works well for images, where metadata can be embedded in the file. For text, the challenge is harder. Text is easily copied, reformatted, and stripped of metadata. A watermark that survives these transformations would need to be embedded in the language itself, which raises questions about how robust such a system could be.
Provenance is not a substitute for disclosure. Even if a watermark could reliably indicate that a piece of text was AI-generated, it would not tell the reader how the AI was used, what prompts were given, or how much human review and editing took place. Provenance and watermarking are technical complements to editorial disclosure, not replacements for it.
Why does AI citation matter for credibility, responsibility and legal risk?
AI can produce authoritative-sounding output that is nevertheless incorrect, incomplete, or biased, so humans must carefully review and edit AI-generated content before submission. This is not a theoretical risk. Large language models are trained on vast corpora of text, including material that is outdated, inaccurate, or biased. They have no mechanism to verify facts, and they can confidently assert claims that are false.
When AI-generated content is treated as a primary source, the risks multiply. A reader who encounters a citation to an AI-generated text cannot verify the claim by consulting the source, because the source is not stable or retrievable. The output of a large language model depends on the prompt, the model version, and the random seed. A different query, even with the same prompt, can produce different output. This makes AI-generated content unsuitable as a primary source in the way that a published article, a dataset, or a legal document is.
Accountability is the second reason disclosure matters. If AI-produced material is incorrect or infringes copyright, who is responsible? The ICMJE's answer is clear: the human authors are responsible. Authors must be able to assert that there is no plagiarism in their paper, including in text and images produced by AI. This means authors must verify that AI-generated content does not reproduce copyrighted material without attribution, and that it does not misrepresent facts.
Legal risk is a third consideration. Copyright law in most jurisdictions does not recognise AI systems as authors. If an AI generates text or an image, the copyright status of that output is uncertain. Some jurisdictions may treat it as a work without an author, which could place it in the public domain. Others may assign copyright to the person who prompted the AI, or to the company that owns the AI system. Until these questions are settled, authors who use AI-generated content without disclosure are taking on legal risk.
How should authors practically cite or disclose AI in manuscripts and publications?
The ICMJE guidance provides a practical template. At submission, disclose in the cover letter whether you used AI-assisted technologies. In the manuscript itself, describe how you used the technology in the appropriate section. For most research articles, this will be the methods section. If the manuscript does not have a methods section, use the acknowledgements or a dedicated statement.
The disclosure should include three elements: the tool name and version, a description of how you used it, and the extent of use. For example: "We used ChatGPT (GPT-4, version of 15 March 2025) to generate an initial draft of the discussion section. We reviewed and edited the output to ensure accuracy and alignment with our findings."
If you used AI to improve language, grammar, or tone, and the AI did not contribute substantively to the content, some publishers consider this basic copy editing and do not require disclosure. However, practices vary, so check the journal's specific guidance. When in doubt, disclose.
For images, the disclosure should be in the figure caption or the methods section. For example: "Figure 2 was generated using DALL-E 2 (OpenAI, 8 March 2025) with the prompt 'pointillist painting of a sheep in a sunny field of blue flowers'. We reviewed the output to ensure it accurately represented the concept."
Do not list the AI tool as an author or co-author. Do not cite AI-generated text as a primary source. If you quote AI-generated text in your manuscript, include the text in an appendix and reference the appendix, not the AI tool, in your citation.
| Scenario | Disclosure required? | Where to disclose | Example |
|---|---|---|---|
| AI generated first draft of a section | Yes | Methods or acknowledgements | "We used ChatGPT (GPT-4, 15 March 2025) to generate an initial draft of the discussion. We reviewed and edited the output." |
| AI improved grammar and phrasing | Varies by publisher | Check journal guidance | "We used Grammarly to check grammar and improve phrasing." (if required) |
| AI generated an image or figure | Yes | Figure caption or methods | "Figure 2 was generated using DALL-E 2 (OpenAI, 8 March 2025) with the prompt 'pointillist painting of a sheep in a sunny field'." |
| AI translated text from another language | Yes | Methods or acknowledgements | "We used DeepL (version 3.2, 10 April 2025) to translate the abstract from English to German. We reviewed the translation for accuracy." |
What open questions remain about standards, enforcement and technical feasibility?
Standards are still silent or inconsistent on several points. The ICMJE and most publishers require disclosure but do not prescribe a single, standard citation format for LLM output. Some journals ask for disclosure in the methods section, others in the acknowledgements, others in a dedicated statement. The lack of a standard format makes it harder for authors to know what is expected and harder for readers to find the disclosure.
Enforcement mechanisms are underdeveloped. The Committee on Publication Ethics updated its retraction guidelines in August 2025 to include undisclosed involvement of artificial intelligence as grounds for retraction, but there is limited consensus or detail on how reliable detection tools are in practice. Publishers are relying on author self-reporting, which depends on authors understanding the rules and choosing to comply. The shift from detection to disclosure assumes good faith, but it is not clear what happens when authors do not disclose.
Provenance implementation remains a technical challenge. NIST and the Federation of American Scientists advocate provenance and cryptographic watermarking, but sources do not resolve implementation trade-offs such as privacy, interoperability, resistance to tampering, or a timeline for standard adoption. Watermarking works well for images, where metadata can be embedded in the file. For text, the challenge is harder. A watermark that survives copying, reformatting, and stripping of metadata would need to be embedded in the language itself, which raises questions about how robust such a system could be.
What publishers, tool-makers and standards bodies need to agree on next is a common disclosure format, a set of enforcement mechanisms that balance trust and verification, and a technical standard for provenance that is practical to implement and resistant to tampering. Until these questions are settled, authors should follow the guidance of the journal they are submitting to, disclose AI use clearly and completely, and take responsibility for the accuracy and integrity of all content in their manuscript.