The history of peer review in medicine³ demonstrates the scientific community's ongoing commitment to accuracy and integrity. From its origins in the seventeenth century, driven by the need to establish and regulate scientific knowledge in emerging societies such as the Royal Society of London and Edinburgh, to its formalization and widespread adoption in the twentieth century, the peer review process has undergone continuous adaptation. Currently, according to the International Committee of Medical Journal Editors (ICMJE), peer review remains the essential "heart" of scientific journals,⁴ serving as a crucial filter and quality control mechanism.

Peer review can be conducted in several formats: open; anonymous to authors; or anonymous to authors and reviewers. Each of these formats has advantages and disadvantages, and there is an increasing trend in the number of journals with open peer review, given that it increases the transparency and integrity of the process.⁵ The preparation of a robust technical report requires a systematic approach that goes beyond a superficial reading. Even before accepting the invitation, reviewers must judiciously assess fundamental pillars,^5,6 including their own technical expertise on the topic, the absence of potential conflicts of interest, and their time availability for in-depth analysis.

Discussions regarding the sustainability and quality of the peer review process are a constant in the scientific community. The increasing number of manuscript submissions has generated considerable pressure on reviewers and editors, raising the question of the need for standardization in order to ensure high-quality reports.

In this context, the emergence of AI tools has introduced the prospect of enhanced efficiency and productivity in the peer review process, allowing time-consuming tasks, such as report preparation and scientific language editing, to be carried out more quickly. Nonetheless, this potential gain comes with new dimensions of complexity, posing ethical issues that require clear regulation. It is important to develop proposals for more clearly defined protocols on the use of AI in peer review, regarding both the quality of critical assessment of manuscripts and the ethical implications related to the use of AI, while reinforcing the notion that final responsibility remains undeniably human.

Before accepting an assignment, reviewers must weigh whether they have available time, whether the manuscript is within their area of expertise, and whether there are any conflicts of interest that would prevent them from fulfilling their task impartially. Subsequently, they must prepare a structured report with the following elements:

The integration of AI tools into the peer review process has generated intense controversy,^7,8 pitting the potential optimization of tasks against the violation of fundamental ethical principles. Some journals have policies that completely prohibit the use of LLMs in the peer review process. Others have taken a more permissive approach. It is worth weighing the arguments on both sides.

1. Absolute confidentiality: The use of generative AI tools, especially those with public access, represents a significant risk of loss of confidentiality, which is central to the peer review process. When a reviewer inserts a manuscript's text, tables, or figures into a LLM to facilitate synthesis or grammatical correction, they are, in practice, feeding a third-party database that will be used in other future situations.⁹ The confidentiality of the review process is violated when AI stores this information for future training. Reviewers have an obligation to protect the information in the manuscripts entrusted to them by not sharing it with their colleagues, collaborators, or LLM platforms.

2. Human responsibility and duty to disclose: Even if AI is used for restricted purposes that do not violate confidentiality, for example, improving the clarity of the review, the responsibility for the content generated rests entirely with the reviewer. Errors, inaccuracies, or biases produced by the tool are, by extension, human responsibility. The underlying issue is that, if a reviewer lacks the expertise to judge the quality of information generated by AI, they should not use it, given that the production of reliable reviews is fundamental to the progress of science.

3. Intrinsic limitations and associated risks: AI does not substitute a specialist's critical judgment. In addition to its significant technical limitations, generative AI introduces additional risk factors that can compromise scientific integrity.

3.1. Inaccuracies and "hallucinations": AI tools can generate unfounded or incorrect criticisms of complex methodologies. "Hallucination" represents a direct threat to patient safety and the credibility of journals. By mimicking non-existent evidence to support clinical decisions or diagnoses, AI not only "hallucinates" but also compromises the ethical-scientific framework of the editorial process.^10,11 Incorrect information can compromise not only the accuracy of a scientific manuscript but also of all future articles that cite that incorrect information.

3.2. Algorithmic bias: Models can perpetuate biases present in their training data, resulting in unfair evaluations of certain topics, methods, or groups of authors. As noted by Checco et al. in 2021, automation in peer review can perpetuate systemic biases present in training datasets, thereby compromising equity in science.¹²

3.3. Lack of critical reasoning: Manuscript assessment requires the interpretation of nuances and an understanding of scientific contexts. These capabilities currently remain beyond the scope of AI. Although AI can assist in detecting plagiarism and verifying compliance with guidelines, it fails to replicate the intellectual synthesis required to validate medical innovation.⁹

We are currently going through a transition period that is reshaping the way peer review is conducted. Seeing that some scientific journals have adopted more permissive approaches, we have conducted a review of these policies in the leading journals in the field of cardiology (Table 1).

As shown in Table 1, there are conflicting positions among cardiology journals. Although many journals are concerned about ethical issues, conducting high-quality peer reviews involves substantial challenges.²⁰ If a responsible and ethical approach to using LLMs for this purpose can be established, we will be able to improve the entire scientific communication process. As a means of addressing this issue, one proposal is to conduct scientific studies on the use of LLMs and their impact on scientific publications.