Epistemic responsibility: toward a community standard for human-AI collaborations

BEGINNING OF THE ENGLISH ADAPTATION · BASED ON THE SITE’S UNOFFICIAL, AI-ASSISTED SPANISH TRANSLATION, NOT THE ORIGINAL ENGLISH TEXT

P(doom), the probability that artificial intelligence will destroy civilization, has recently emerged as a subject of public debate (https://en.wikipedia.org/wiki/P(doom); Thomas, 2024). The supposed likely causes of an AI apocalypse vary and are often colored by science-fiction tropes, but they express and reveal underlying concerns about AI’s entry into domains previously reserved for human beings (Kokotajlo et al., 2025).

Concerns about AI generally arise from two sources: the ethical and the epistemic. AI developers strive to keep their systems “aligned” with human interests and values, an ethical concern (Christian, 2020). They also struggle to control AI “hallucinations,” which might be better described as confabulations, an epistemic challenge (Sun et al., 2024). If large language models (LLMs) were compulsive liars, they could be dismissed as sources of knowledge, regardless of their appearance of sincerity and conviction.

But in several recent cases, AI systems have genuinely contributed to human knowledge, drawing on the rapid processing of large volumes of data and pattern detection through neural networks. This expanding field of AI application includes biomedical research, particularly drug discovery (Gao et al., 2024; Jiang and Zhao, 2025), astronomy and cosmology (Huang et al., 2024; Spindler, 2022), and data analysis in the social sciences (Balla et al., 2025).

The latest versions of “agentic” and “reasoning” systems have offered researchers, and in practice everyone, opportunities to make useful use of AI’s distinctive capabilities across numerous applications (Mollick, 2025). This has heightened concerns about the deception involved in presenting AI output as one’s own original work, alongside the threat of AI-generated disinformation. These concerns haunt every level of education and the work of scholars, artists, scientists, and other professionals.

Although deception and unwarranted credulity are irresponsible, a simple ban on the use of AI in education and the professions would preempt potentially beneficial applications. In any case, it is increasingly clear that such a blanket ban will not work (Mello, 2023). What is needed instead is a positive code defining the appropriate use of AI in contexts that seek to discover, preserve, extend, and communicate truth.

These “knowledge contexts” constitute the domain of epistemic responsibility. The aim of this brief opinion essay is to outline principles that researchers can apply and that can serve as a benchmark for evaluating good practice in academic and professional writing.

“Epistemic responsibility” is a broad subject, connected with philosophical discussions of the ethics of belief formation and the intellectual virtues that promote the production of knowledge (Battaly, 2008; Meylan, 2013). This essay’s concern is narrower. It does not ask whether AI should be used or avoided in the pursuit of knowledge. As we have seen, AI is already widely used in every intellectual domain and in most workplaces.

Nor do these guidelines specify how AI should be used, or how the claims it produces should be understood and evaluated. Rather, they seek to secure a precondition for addressing those broader questions. The question here is how to acknowledge and make explicit AI-generated content. The following guidelines aim to delineate the human and the artificial in publications; once the two are clearly distinguished, broader questions concerning evidence and justification may also receive new illumination.

This, however, is a subject for future discussion.

READING GUIDE · EDITORIAL HEADINGTransparency and replicability

The aims of the proposed protocols for AI collaborations are transparency and replicability. The two are mutually reinforcing. Achieving one usually also makes the other easier. These are familiar aims in academic publishing and scientific research.

Providing a standard, visible protocol for AI use will help reassure readers and consumers about authors’ good practice, and will give authors a clear, visible standard of conduct requiring full disclosure of any AI-generated material in published research. These standards need not be ratified by professional or publishing organizations, although they could be. Rather, the hope is that they will spread until they become a straightforward community expectation.

Their widespread use will allow AI to be used effectively, but with maximum transparency.

The proposed standards are listed below, together with a hypothetical example (Box 1).

READING GUIDE · EDITORIAL HEADINGFour standards for making AI visible

Standard 1: visibility. The inclusion of AI content must be immediately apparent to all readers, even at first glance. The AI source should be indicated in the title heading, in as much detail as possible. As the hypothetical example shows (Box 1), this declaration does not imply AI co-authorship, but appears on a separate line. This includes identifying the sections of the article that draw on AI output and delimiting the text written by it. The article’s abstract should also include this information.

If there is no AI content, that should also be stated.

Standard 2: replicability. AI applied to research inevitably involves directing its behavior to serve the researcher. This is the engineering of instructions (prompt engineering), and the prompt used is an essential tool for understanding the result, its implications, and its limitations. The prompts actually used must be set out explicitly and in full in any work that includes AI-generated content. The specific prompt or prompts that produced that content are also essential for replication.

Ideally, the stages in the evolution of the prompts should be documented and presented as supplementary material, if they are not included in full in the main text.

Standard 3: content checking. Because LLMs readily confabulate, no reference or quotation can simply be taken on trust. Every factual claim in AI content therefore needs a person to check it. At a minimum, the fact-checker must confirm that the bibliographic information supplied by the AI is correct. They must also confirm that the claims made by the AI are actually supported by the cited sources.

Finally, they must confirm that no AI-generated text duplicates or closely paraphrases text from other sources. All of this is essential and must therefore be explicitly confirmed, also at the beginning of the article. An author can perform this role, but a research assistant can also contribute. In either case, the person must be identified along with their contact details, should questions arise.

In this way, authors and fact-checkers are identified as explicitly and transparently responsible for overseeing AI-generated content.

Standard 4: intratextual clarity. All AI-generated content within a research report or any other publication must be distinguished from human-generated content through distinct stylistic markers, such as block quotations or alternative typefaces.

These are relatively simple guidelines, readily adaptable to different contexts. Research in AI and related fields should certainly be governed by these epistemic safeguards. However, these expectations can apply to any writing that presents evidence and arguments in support of a conclusion. They should be automatic and become second nature in academic writing. Likewise, they should be taught as part of expository writing, on an equal footing with prohibitions against plagiarism and other forms of academic dishonesty.

In this way, epistemic responsibility regarding AI can become part of the school curriculum and apply explicitly to both students and teachers.

The guidelines also apply by direct analogy to creative contexts. Artistic works may not have the immediate aim of producing knowledge, but their originality and authorship are equally threatened by undisclosed AI use. Artists’ statements routinely accompany artworks in every medium. Those statements should explicitly comply with the standards described above.

READING GUIDE · EDITORIAL HEADINGAuthorship, publication, and trust

These guidelines differ from those of COPE, the Committee on Publication Ethics, which broadly addresses good practice at every stage of publication (COPE Council, 2023). COPE emphatically rejects human–AI co-authorship and therefore recommends that AI applications or programs not be included on title pages. Instead, AI use should be detailed in methods sections. However, this invites a degree of uncertainty by burying AI’s crucial influence inside the article.

By contrast, the guidelines recommended here make AI content immediately apparent to everyone. As the hypothetical example demonstrates, this initial visibility does not imply that AI is a co-author. The recommended AI notice constitutes a distinct, new piece of information that should appear by default at the beginning of any publication.

Similarly, the International Committee of Medical Journal Editors (ICMJE) agrees with COPE’s prohibition on attributing co-authorship to AI and advises authors to use it cautiously when writing and editing (American Medical Writers Association, 2023). This also leaves some latitude in acknowledging AI-generated content and allows ambiguity to remain for the reader. By contrast, these guidelines expressly require that content to be acknowledged and its presence in any publication to be clearly delimited.

In addition, requiring publication of the relevant prompts facilitates replication and further exploration of AI’s potential contributions. Finally, the guidelines require explicit acknowledgment that AI content has been fact-checked and emphasize the fact-checker’s role, thereby making both author and checker responsible for ensuring that content’s reliability.

As these guidelines become community expectations, they will help ensure prudent AI use. This, in turn, could reduce concerns that information sources are covertly infected with confabulated AI content. In an ideal future, AI-generated content would be as reliable as human research. But in the real world, AI confabulation will continue to threaten our understanding of the world with its flood of accidental or deliberate falsehoods.

Against this trend, the proposed criteria for making AI use explicit can clarify and consolidate human responsibility and authority in the production of knowledge.

Declarations from the original article

Author contributions

DL: Writing the original draft; reviewing and editing the text.

Funding

The author or authors declare that no financial support was received for the research or publication of this article.

Conflict of interest

The author declares that the research was conducted in the absence of commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author or authors declare that generative AI was not used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of its authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors, or the reviewers. Any product that may be evaluated in this article, or any claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. American Medical Writers Association (2023). ICMJE Authorship Guidelines and Acknowledging Non-author Contributions. Available at: blog.amwa.org.
2. Balla J., Huang S., Dugan O., Dangovski R., Soljačić M. (2025). AI-assisted discovery of quantitative and formal models in social science. Humanit. Soc. Sci. Commun. 12, 1–12. 10.1057/s41599-025-04405-x.
3. Battaly H. (2008). Virtue epistemology. Philos. Compass 3, 639–663. 10.1111/j.1747-9991.2008.00146.x.
4. Christian B. (2020). The Alignment Problem: Machine Learning and Human Values (1st edition). New York, NY: W. W. Norton and Company.
5. COPE Council (2023). Authorship and AI tools. Eastleigh: COPE, Committee on Publication Ethics.
6. Gao S., Fang A., Huang Y., Giunchiglia V., Noori A., Schwarz J. R., et al. (2024). Empowering biomedical discovery with AI agents. Cell 187, 6125–6151. 10.1016/j.cell.2024.09.022.
7. Huang K., Hu T., Cai J., Pan X., Hou Y., Xu L., et al. (2024). Artificial intelligence in astronomical optical telescopes: present status and future perspectives. Universe 10:210. 10.3390/universe10050210.
8. Jiang W., Zhao Z. (2025). Trends in research on AI-aided drug discovery from 2009 to 2023: a 15-year bibliometric analysis. Intell. Pharm. 3, 71–83. 10.1016/j.ipha.2024.09.001.
9. Kokotajlo D., Alexander S., Larsen T., Barnes E., Dean R. (2025). AI 2027. Available at: https://ai-2027.com/.
10. Mello J. P. (2023). Experts Say Workplace AI Bans Won't Work. TechNewsWorld. Available at: TechNewsWorld.
11. Meylan A. (2013). Foundations of an Ethics of Belief. Piscataway, NJ: De Gruyter. 10.1515/9783110327816.
12. Mollick E. (2025). On Jagged AGI: O3, Gemini 2.5, and everything after. Available at: One Useful Thing.
13. Spindler A. (2022). How artificial intelligence is changing astronomy. Astronomy Magazine. Available at: Astronomy Magazine.
14. Sun Y., Sheng D., Zhou Z., Wu Y. (2024). AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content. Humanit. Soc. Sci. Commun. 11, 1–14. 10.1057/s41599-024-03811-x.
15. Thomas S. (2024). Are We Ready for P(doom)? The Spectator. Available at: The Spectator.

END OF THE ENGLISH ADAPTATION