Are we asking the fox to guard the hen house? In academia, AI has slowly found its way into our writing and publication processes, helping researchers generate and refine content while also being used by publishers to screen manuscripts and identify potential errors. Does it introduce errors? Or is helping identify errors? The answer is not so simple, because in many cases AI is helping identify errors that AI itself has introduced to the paper. It’s like filling a leaky bucket. No wonder we seem to be going around in circles, especially when it comes to research integrity and trust in research. It is this AI paradox that has left us confused.

How can we then trust AI-assisted papers and feel confident that they are free of AI-introduced errors? And should we automatically consider these AI-assisted papers less trustworthy than papers that were written entirely by humans? As Avi Staiman says in this post published in The Scholarly Kitchen,”Damned if You Do, Damned if You Don’t: Multilingual Researchers’ Use of AI in Academic Writing”:

“If the English is rough, the author may be judged as less credible (especially when they could have fixed it up with AI!). If the English seems suspiciously polished or stereotypically “AI,” the author may also be judged as less credible”.

But it is quite possible that the human-written paper may have fabricated references and unsupported claims as well. If we ask ourselves honestly on a scale of 1 to 10, which paper would inspire more trust? The human-written paper? Or the AI-assisted one? Would we place greater trust in the human-written paper. And would that judgment be based on the quality of the paper or on how it was produced?

The absence of AI doesn’t guarantee trustworthy research. And neither does disclosing AI use. AI transparency is not unnecessary but it may not necessarily be an indicator of the quality of research or of research integrity. AI disclosures only tell us whether AI was used and how it was used. It does not prove that technology was used responsibly or that the claims were verified and the paper is free from errors.

Let’s consider this. An author may have been fully compliant with a journal’s AI disclosure policy but may still submit a paper with unsupported claims and fabricated references. Compliance with disclosure policies does not equate to research integrity. This is the transparency trap where we may mistake disclosure for proof of research integrity and quality.

It cannot be denied that AI creates additional risks. Different uses of AI carry different levels of risk. Using AI to improve grammar is not the same as using it to generate interpretations,  analyze data, or write huge sections of the paper. It is important to take into consideration the nature and extent of how AI was used, since it can significantly influence the kind of errors introduced and the verification needed. The question therefore goes beyond whether AI was used, but what it was used for, what risks it may introduce, and how those risks can be addressed. Disclosure can help with only a part of this.

The same scrutiny should also apply to human-written papers. The absence of AI use does not guarantee that a paper does not have serious issues or unsupported claims and fabricated references. AI disclosure for AI-assisted papers is therefore only a small part of the accountability process. The problem is not disclosure itself, but the weight we assign to it. Disclosure is an important starting point for accountability,

The bottom line is that disclosure and detection alone cannot resolve integrity issues. They only give us a false sense of trust. In fact, in August 2026, the JAMA network updated its guidelines and acknowledged that “disclosure policies may not be sufficient to address concerns about AI-generated content that is incorrect or even fabricated (ie, hallucinated)”

The focus should instead be on developing a framework of accountability and validation. The author needs to take accountability for a paper, whether AI-assisted or not. What we need is a meaningful human validation layer where humans add value and validate the authenticity of the research. But what do we mean by human validation? Does human validation mean that the reviewer should independently verify every claim, reference, dataset, or detail in a paper? Peer review was never designed to provide this kind of blanket assurance, and expecting reviewers to catch every error is unrealistic.

Additionally, with AI becoming so sophisticated, human validation cannot serve as a blanket guarantee that a paper is free from errors. Human expertise should complement technical checks. While appropriate tools can help flag potential errors and inconsistencies, human judgment is essential to evaluate the evidence available, whether there are any gaps in reasoning, whether the argument is weak, and whether it makes a meaningful contribution to existing knowledge. This recently published paper on “When disclosure rules are not enough: authorial warrant and judgement in AI-supported scholarship” summarizes:

“The issue is therefore not only where a model was used, but where human judgment and calibration remained and whether the researcher retained the capacities required for responsible authorship.”

Ultimately the use of AI or the absence of it should not become a basis for bias in research integrity assessments. The guiding principle should be the soundness of the research, the accountability of the authors and the value of the research. At the end of the day, the presence of AI does not establish misconduct just as the absence of AI does not establish integrity.

AI disclosure statement: ChatGPT was used to assist with editing, refining, and structuring this article. The author reviewed the suggested revisions and retains full responsibility for the article’s arguments, interpretations, and final content.