Treat Your AI Collaborator Just Like a Human

Every generation of scholars has its technological anxieties. When I was starting out in my career, admitting that you had found a source through Google Scholar felt vaguely embarrassing. Specifically in my field of Buddhist Studies, using the online version of the Chinese Buddhist canon ran afoul of established norms. “Real” scholarship, it was thought, hinged on the printed page. Until one day it didn’t. Seemingly overnight, the hesitations evaporated and online resources became just another research tool.

It feels somehow different this time with artificial intelligence. LLMs can so effortlessly produce text that many fear that human authors are in danger of becoming obsolete. But just because AI can produce words, does it necessarily change anything about our notions of academic authorship?

The difference between wordsmithing and authorship has always been clear when it comes to engaging with human collaborators. We have never considered copy editors to be authors, no matter how much they have improved the clarity and readability of our manuscripts. This is because proofreading, grammar correction, and stylistic polishing do not change who is intellectually responsible for the contents of a scholarly paper. Even if an editor makes changes to every sentence in the manuscript, their contribution is rarely even acknowledged in the final publication.

Likewise, if we are using LLMs simply as editorial tools, then this does nothing to compromise authorship either. We are not after all novelists or poets who imbue our every word with creativity, style, and personality. With most academic writing, originality is judged primarily by the quality of the ideas and arguments rather than by the particular turns of phrase used to express them. Does it really matter if the final polishing was done by a human or a computer?

But there has always been a limit. Imagine now that the editor is sitting beside you while you write, inserting their own interpretations, suggesting arguments, and adding sentences throughout the drafting process. In this scenario, editing has crossed over into authoring. The other party is no longer simply polishing the work; they are now participating in its creation.

A human in this second case is always expected to be named as a co-author, a step that both allows them to be credited and to be held accountable for the work. However, the Committee on Publication Ethics, which issues AI guidelines used by many academic publishers, states that LLMs cannot legitimately occupy that role. As helpful as AI may be, it cannot assume responsibility for its contributions. It cannot defend its arguments, answer criticism, or correct mistakes. That is to say, no matter how many words it extrudes onto the page, a non-human collaborator cannot participate in true authorship.

If AI cannot be listed as an author on the grounds that it cannot be intellectually accountable, it follows that AI should never be used in a way that would ordinarily entail being credited as an author. To do so would be as unethical as, say, a professor refusing to share authorship with a student who made major contributions to a manuscript. There is no new ethical principle needed for AI in this scenario—it’s simply old-fashioned academic integrity.

Between the poles of editing and authorship laid out above there are many shades of grey. But, while on first glance, many of these usages may seem to present new ethical challenges, again, comparing the involvement of AI to that of a human is usually clarifying. For example, if a human research assistant helped me to identify relevant sources or code a database, I wouldn’t necessarily name them as an author. But, it would be unethical not to mention them in the acknowledgements, or at least in a footnote.

The same applies to a colleague offering an insight that changes how I think about my project, or a peer reviewer making a comment that provoked me to fundamentally rethink my argument. By the same logic, if an LLM has contributed materially to my paper in any of these ways, then it would be equally unethical not to mention this. Again, there’s nothing unique or special about AI in this case: it’s simply what we’ve always done, both as good manners and for the sake of transparency.

How exactly to acknowledge these shades of grey is also a matter with long precedent. While routine assistance is usually mentioned in passing, the use of a new research methodology warrants more prominence. The methodology section, the introduction, or even the abstract is where we identify the archives we consulted, the interpretive methods we used, and the analytical frameworks we employed. It’s also here that we should mention AI if it helped to shape any of these activities—to identify recurring themes across hundreds of documents, to summarize large collections of material, or to generate avenues for further investigation, for example.

Thinking about AI as methodology also means honest discussion of its limitations. Every research method has its weaknesses. Archival collections are incomplete. Statistical models rely on assumptions. Likewise, AI introduces its own blind spots: the potential for hallucinations, biases inherited from training data, and uncertainty about the human labor embedded in those datasets. If these limitations have the potential to affect our research, the reader needs to know this up front. Not because AI use is suspect, but because good scholarship always has required clearly documenting how our insights were generated and how our conclusions were arrived at.

With proper acknowledgement and transparency, there is in my view little reason to be defensive about using AI. On the contrary, if a novel research methodology made possible a kind of analysis that would otherwise have been unavailable (like when I used CBETA to do corpus-level analysis of translation terms 20 years ago), explaining that process is itself part of the scholarly contribution of the paper. Rather than glossing over our AI use or relegating it to a passing footnote, we should lean into detailing what the technology made possible, where it may have fallen short, and why we nevertheless have judged the resulting conclusions to be worthy of publication. These have always been and will always be the time-tested practices of good scholarship.

It is to be expected that every new technology will unsettle our assumptions about how we do our work for a time. But eventually it becomes ordinary, and our attention returns to the principles that mattered most all along. AI will probably follow the same path. Today, readers do not care whether we searched CBETA or got a source from Google Scholar, and they soon won’t care if we used Claude, ChatGPT, or Gemini as an editor or research tool either. But they will always care whether we have been honest about how our work was done, the limitations of our tools, and whether our conclusions can be trusted. When it comes to what responsible authorship requires of us, AI has changed nothing at all.