Artificial Intelligence (AI) Policy

Version 1.0
Drafted: May 1, 2026
Last Updated: Sept. 19, 2026
Drafted by: Kevin Horecka
Edited by: JSMCAH AI Policy Working Group and JSMCAH Editorial Board

AI Disclaimer: This document was prepared in Google Docs without the use of AI technologies beyond spelling and grammar correction. Generative AI was subsequently used during editorial review of the entire JSMCAH site to identify consistency issues and assist with minor revisions.

Policy for authors
Policy for Reviewers and Editors
Recommended Modes of Use of AI Systems
Common Pitfalls in the Use of AI Systems
Questions and Concerns about the AI Policy

This document describes the Artificial Intelligence (AI) usage policy for JSMCAH for reviewers, editors and authors. It is intended to be permissive on the use of modern AI tools while maintaining the scientific integrity and authenticity of published works.

The scientific process, by its nature, demands that we not simply accept the truth or falsehood of any one (or more) result. Replication, rigorous statistics, experimental design, and other core scientific tenets and techniques attempt to verify the reliability of results through the scientific process. In the space of scientific publishing, trust in and the integrity of the process has always been a core requirement, enforced by blinding, data sharing requirements, and other forms of documentation. Historically, any breaches of this trust have, whether by malice or ignorance, been treated very harshly. This trust relationship spans editors, reviewers, authors, and readers in a journal setting. We believe that the use of AI does not inherently degrade these trust relationships provided it is used appropriately. Cultural factors, technological advancements, and methodological evolution will mean that what is considered permissible in this space is a moving target. As a result, rather than attempting to specifically adjudicate all possible permissible and impermissible uses of AI in research and publications, we focus instead on the core trust relationship. For convenience, we also provide some examples, current as of September 2026, of where appropriate navigation of this trust may tread.

To aid in this policy definition, we adopt the following core beliefs:

Authentic knowledge is knowledge which has been challenged, tested, and connected to other knowledge in a way that allows it to be of persistent value
Humans construct authentic knowledge using a variety of tools and publish it for the purposes of advancing scientific understanding of the world
AI systems can aid humans in this process, but they cannot be the entities which provide justification for the authentic knowledge
Trust between author and reader is anchored in the peer review process establishing that the published knowledge is authentic in nature and, to the best of the ability of the parties involved, it has been vetted to be a factually presented representation of the underlying scientific concepts, methods, results, and interpretations
AI policy should attempt to curate a positive trust relationship between authors, reviewers, and readers via the use of AI systems in ways which enrich but do not replace the need for the authors or reviewers to communicate authentic knowledge

Some social tools which can be leveraged to help aid in this trust relationship are:

Disclosure of use - everyone should be clear what content was created by, contributed to, or untouched by AI systems
Transparency of methods - everyone should be clear on the precise ways in which AI created or contributed to disclosed content
Recommendations for practices - everyone should be clear on reasonable uses of AI systems
Pitfalls for practices - everyone should be clear on unreasonable uses of AI systems

The description of an AI system may differ across different groups, so for the purpose of clarity, we provide the following definitions and examples:

Generalized Description of an AI system - An AI system is one which makes a decision that otherwise would have required a human to make, especially when that decision is anchored on a set of prior training data. The definition of an AI system is a hotly debated topic, with some being comfortable applying the term to classical algorithms which appear intelligent in sometimes surprising ways while others prefer it only cover data-driven or even generative systems. Rather than fully adjudicating this definition here, we provide a few examples of things which are certainly and are likely not covered by this policy.

Common AI systems covered by this policy include:

Chat or code systems backed by a large language model (e.g., Claude Code, Claude, ChatGPT)
AI search tools like Gemini or ChatGPT (to the extent their outputs are taken as fact rather than useful hints)
Fine-tuned local models which generate text, images, or other artifacts and decisions
Models trained on large corpuses of data for the purpose of performing general purpose linguistic tasks and interacting with the researcher

AI systems which are not covered under this policy include:

Traditional search tools (Google Search without Gemini, Duck Duck Go, Bing, Google Scholar, etc)
Spelling and grammar correction (and generally, common non-generative features in word processors that existed before 2017)
AI accessibility tools such as transcription or screen readers

Policy for Authors

Authors should:

Completely disclose the ways in which AI contributed to their work, at any stage of the process, but with a special focus on any decision making, writing, statistics, or other steps which might influence the outcome or interpretation of the work. If no AI was used, this should be stated. Disclosure must be complete. Incomplete disclosure will be met with the same judgment as no disclosure at all, though good faith efforts will be accounted for in determining the consequences for non-disclosure.
Provide clear, complete documentation which transparently explains the AI use in a way which would allow inspection or interrogation in the same ways in which Methods in a usual academic setting might be scrutinized. This transparency should be progressive with the extent of AI use. A small use of AI may be explained with a small amount of documentation. An extensive use of AI requires a concordantly extensive amount of documentation. If you’re unsure if documentation is sufficient, please contact your editor to have a discussion about it before review begins. When in doubt, the litmus test is “is this documentation sufficient to maintain the trust of all parties involved?” - if the answer is ambiguous, have a discussion about it.
Be held accountable for all content they submit, regardless of disclosure, with clear consequences for trust violations (see below)
Read and understand this document, especially the Recommendations and Pitfalls section of this document, and utilize this information to attempt to be responsible scientific users of AI systems
Communicate proactively with journal editors when unsure about any aspects of this policy

Consequences for Trust Violations by Authors

In case of a trust violation, the editor will determine the level of severity of the violation. In the case of minor violations, a warning will be issued to the corresponding author and an opportunity to correct the issue will be provided.

For more severe violations, e.g., substantial use of AI that is not disclosed or an entire fabricated submission, editors will reject the submission. An appeal can be made to the journal and is subject to approval by the Editor in Chief.

Critically, an issue which might otherwise receive significant leeway from an editor or journal in the context of a best-effort human submission will receive a more severe judgment when misuse of AI is involved. AI presents great temptation to take shortcuts in the scientific and publication process. By using AI systems, authors, reviewers, and editors have a greater burden of quality than those who do not.

Policies for Reviewers and Editors

Reviewers have a unique responsibility to contribute their knowledge during the peer review process. Although this knowledge may be enriched by the use of AI systems, a review which is entirely AI-Generated is of no value to the process. As a result, reviewers are strongly encouraged to avoid the use of AI when performing an initial examination of a work, and if they use AI in follow-up reviews, they are subject to all the same transparency requirements as an author. Reviewers must protect the confidentiality of submitted manuscripts and peer review and must not provide confidential manuscript or review content to an AI system where confidentiality cannot be assured. It is extremely dangerous to the trust relationship if a review provides feedback sourced from an AI even if that feedback is valid and claim it as their own. It is at the discretion of the editor to help reviewers and authors navigate conflicts that may come from any element of a review that is AI related. The process of review is a negotiation and depends on maintaining the authenticity of the interaction. It is ultimately up to editors to examine reviewer feedback and determine its authenticity as well as to hold reviewers accountable to submitting authentic reviews.

The same policy which applies to Authors applies to Reviewers and Editors, however, the Editor has the discretion over the Reviewer, while the Editor in Chief has the discretion over the Editors. In cases of a complaint about an editor, the board will have private, offline discussions excluding that Editor if they are a member to determine what level of punitive measure is appropriate up to and including removal of Editor status and/or removal from the board.

Editorial Use of AI

JSMCAH editors may use AI systems to support editorial work, including preflight checks for completeness, internal consistency, reporting quality, reference concerns, and other issues that may warrant clarification or revision before or during peer review. AI may also be used to help organize editorial observations or draft editorial correspondence.

These tools support, but do not replace, editorial judgment. Editors are responsible for verifying AI-assisted observations before acting on them and for all requests, recommendations, and decisions communicated to authors. AI systems do not make decisions about whether manuscripts are sent for review, revised, accepted, or rejected.

When AI materially contributes to the evaluation of a manuscript or to substantive editorial feedback, its use will be disclosed to the authors. JSMCAH will not provide confidential manuscript or peer-review content to an AI system where confidentiality cannot be assured unless appropriate permission has been obtained.

Consequences for Trust Violations in Reviewers and Editors

Reviewers found in violation of the policy will be banned from further review after 2 warnings.

A Note on Copyright

It is up to the individual user of any software system to ensure they are doing so in accordance with the laws in their region or municipality. The use of some materials in AI systems may constitute intellectual property violations. This policy does not cover any legal requirements related to any of the topics which it discusses. All legal requirements are the responsibilities of the parties involved. None of the recommendations here constitutes legal advice nor should it serve as a basis for the violation of any law.

Recommended Modes of Use of AI Systems

Here, “risk” means the risk of damaging trust between authors, reviewers, editors, and readers, not simply that AI will give a “bad” answer. The main concern is that AI use makes the work harder to trust by making it hard to understand the validity, authenticity, and completeness of the underlying knowledge, introducing material that is difficult to verify, introducing incorrect information into the scientific record, or shifting the burden of checking the work from the authors onto Editors and Reviewers.

JSMCAH does not wish to discourage responsible AI use. In many cases AI can increase the quality of research, and may be particularly beneficial to newer or non-traditional authors without the institutional support found in academia. As mentioned in the introduction, what is deemed culturally acceptable around the use of AI in academic publishing is likely to evolve over time. However, as of 2026, the following categories provide a guide to common uses and misuses.

Note: This is not a comprehensive list. It is meant as a helpful guide.

Scientific Question Generation and Study Design

Trust Risk: Moderate to Extremely High

Reason: Allowed and potentially helpful with appropriate author ownership and disclosure. Encouraged when human-led and transparently documented, and human authenticity of knowledge is maintained.

Explanatory Notes: AI systems may be used to help researchers operationalize a research question from a broad observation and translate practice-based concerns into testable hypotheses. The trust risk is when AI is used as a substitute for human inquiry and curiosity. Authors must be able to explain and defend the final hypothesis or research question. Moreover, AI systems are explicitly trained to provide compelling answers to humans. As a result, AI systems may lead to a condition sometimes termed "AI Psychosis," where a user of an AI system, through long and intense interactions with that system, may develop extreme confidence in a set of beliefs which are not grounded in authentic knowledge.

AI can help authors understand study design options and determine whether studies are feasible via tools such as sample size analysis. This can significantly improve study quality, particularly if it is done before a study is performed. The trust risk is when it is used post-hoc to make a study look more rigorous, prospective, etc than it actually was.

When in doubt, consider if you could completely explain and defend the concepts in person without the AI system present. If not, you may be treading into a misuse of AI systems.

Common Failure Modes: AI often misses the larger context around the experiment, environment, or other critical abstract concepts which might impact study design. This can lead to bad design decisions which appear excellent at first glance and are only caught when it is too late to recover from the error. Picking a reasonable scope for a study is also difficult for an AI system and can lead to waste or overuse of resources which is irresponsible in that community.

Drafting and Editing

Trust Risk: Low to High

Reason: Unreviewed and unedited AI-assisted or AI-generated text is a serious trust violation because it introduces the risk of serious errors and propagation of inauthentic knowledge. The trust risk increases when AI generates content that the authors have not independently developed or verified.

Explanatory Notes: AI can help improve clarity, grammar, organization, and readability. This may be particularly helpful for newer, non-traditional or non-native English speaking authors. Different sections of a paper have different tolerances when it comes to AI first drafts. Methods and Statistical sections tend to be better targets for complete AI drafting (i.e. no human written elements or limited human outlining). AIs can be helpful for assisting with initial outlines and structures for the Introduction and Discussion sections, but are unreliable for providing an authentic draft because they cannot reproduce the authors’ thought processes, judgment and reasoning. Summaries can benefit from AI drafting, but this risks misrepresenting the content of the work in these key areas. The Abstract is the most likely section to be read, and the Conclusions crystallize the importance of the findings. Careful attention therefore needs to be paid by the authors to ensure the authenticity of these high-impact sections.

Language choice is a key element of the trust relationship that is central to the guiding principles of this policy. Language that is clearly AI-generated casts the trustworthiness of the rest of the work into question. For this reason, authors must avoid “AI Speak” in their writing. Common AI phrasing such as repeated hard stop sentences, currently AI-preferred words, or excessively vague or hedged sentences should be avoided. This is a moving target as AI systems evolve, and authors must stay updated and ensure that the final edited version is clearly written or appropriately edited by a human.

Common Failure Modes: AI Speak, overly hedged language, excessive detail, and poorly supported statements are all common trust-degrading failure modes. These failure modes are non-uniform across sections of papers.

Statistical Analysis

Definition: The use of an AI system for the analysis of any data or creation of any statistical functions as well as the interpretation of statistical results.

Trust Risk: Moderate

Reason: AI-assisted statistical analysis is unacceptable if results cannot be verified. The trust risk occurs when analyses are not reproducible. Authors must be able to provide an auditable path from data to reported results.

Explanatory Notes: AI can be helpful to support statistical reasoning or identify possible analytic approaches, explain concepts, generate or refine code, suggest or perform sensitivity analyses and improve reporting. This can significantly improve manuscript quality, particularly when authors don’t otherwise have access to formal statistical support. The risk associated with statistical analysis use is generally tied to the degree of supervision and expertise of the human reviewer of the AI work. AI use in exploratory data analysis is typically appropriate if statistical assumptions or blindedness policies are not being violated, but when it comes to hypothesis testing and generalized analysis of data, it’s critical the authors guide the assumptions and model choices rather than relying on the AI for those designs alone.

AI can be surprisingly good teachers of statistics, and the skepticism of the human in the loop in leveraging AI for statistics is a critical factor for success.

Common Failure Modes: Failures are most likely when AI systems are used uncritically by researchers who lack the ability to question and verify the methods and results. In these cases, assistance from a qualified human is essential. AI systems may misinterpret the meaning of input data, overlook important context, or apply otherwise valid statistical methods inappropriately. Issues associated with P-hacking (i.e. the misuse of data analysis to force a statistically significant result) and verifiability/reliability are more severe when the AI is the sole expert in the statistical process.

Literature Review

Definition: The use of AI in the examination, aggregation, collation, and interpretation of existing published academic and non-academic works.

Trust Risk: High

Reason: The trust risk is high because current AI can generate fabricated citations, incorrect metadata, and inaccurate summaries. Authors must independently verify every cited source, without exception. This responsibility must not be transferred to reviewers. Each cited article must exist, be accurately represented and support the statement for which it is cited (in this regard it is no different from a non-AI assisted citation, but the risk of a fully fabricated citation is much greater).

Explanatory Notes: AI systems can be helpful for suggesting search terms, performing literature searches, summarizing author-provided articles or checking whether stated claims are present (if the text of both is available and may be used lawfully). One mitigation which can assist in making AI literature reviews more trustworthy is enabling the system to download the full article text of all citations and verify the legitimate article context.

Common Failure Modes: Fabricated citations, citations for the wrong work, fabrication of context around a citation, author fabrication, and a litany of other errors may occur. These errors cascade into all other aspects of the scientific process. Worse, the continued publication and citation of these fabrications poisons the scientific literature as a whole.

Citation Management

Definition: Using an AI system to exclusively manage the citations and references within and at the end of a paper without the use of intermediate citation management software such as Zotero, Mendeley, and EndNote.

Trust Risk: High

Reason: AI is not accurate enough to be used as the authoritative source for references because the risk of errors is high. Citation management requires accurate, verifiable records.

Explanatory notes: AI is strongly discouraged as an authoritative citation source but can be helpful for verifying reference format. Authors should use a reference manager to manage citations. AI is helpful for reviewing a reference list generated by a citation manager and flagging possible formatting problems, incomplete metadata, duplicate entries or other formatting errors. Suggested corrections should be checked against an authoritative source.

Common Failure Modes: As with Literature Review, chronic literature poisoning is the major failure mode. This is an insidious failure mode and lack of detection compounds this failure over time. We are all responsible for the health of the academic literature as a whole.

Mock Review

Definition: The use of an AI system for examining existing written work for the purpose of providing a simulated set of reviewer feedback.

Trust Risk: Low to Moderate

Reason: This use of AI can often be one of the most helpful. AI systems can perform pre-submission quality checks which, although they vary in their effectiveness and depth, can serve to help an author consider aspects of their writing style, grammar, word choice, explanatory clarity, or reasoning from a new lens. The risk is that uncritical acceptance of the mock review can be detrimental to the ultimate quality of the manuscript.

Explanatory notes: AI systems can be used to conduct a mock review before submission. This can help authors identify flaws before submission and peer review, improving the likelihood of acceptance and reducing preventable burdens on editors and reviewers.

Common Failure Modes: AI system recommendations may make assumptions about what is or isn’t valid or practical in data or an experimental setting. These reviews should be treated with a higher degree of skepticism than a typical peer reviewer in areas of nuance or in larger, sweeping change recommendations. Extending these reviews into edits adopts the risk profile in the Drafting and Editing section above, including risks associated with fabricated knowledge or information, style misalignment, over-hedging language, unsupported or undersupported content, and excessively detailed explanations.

Common Pitfalls in the Use of AI Systems

The following section lists AI policy violations for the use of AI.

Unreviewed AI Text and “AI Speak” - Text quality and style is governed by cultural norms, and any errors in text or stylistic choices which violate norms or compromise the integrity of the academic work are considered a violation. It is the responsibility of Authors to review all text and be responsible for its content.
Fabricated Citations, Statements of Fact, or Data - AI systems can still hallucinate. They can do so in nuanced ways which can lead to a variety of fabricated references to other facts or works in the world. Perhaps it should not require saying, but synthetic or simulated data must never be presented as empirical observations. Fabricated data is just as fraudulent when generated by AI as by hand by a human. Any of these fabrications are considered a severe violation.
Non-Auditable or Diagnosable Statistics - Any statistical methodology which cannot be clearly audited, reproduced, and understood by Reviewers or Readers is considered a violation. This is true of non-AI statistical analyses as well, but it is worsened by the rapid pace at which an AI system might generate an extensive analysis.
AI-Driven P-Hacking - As with (3), because AI systems are capable of rapidly trying a variety of methods, this makes P-Hacking (i.e. varying the statistical methods until a significant value is achieved) easier than ever to do by accident. It’s critical that all statistical tests be documented and published, not just those that turned up significant results. AI-Driven P-Hacking is a severe violation.
AI-Generated Peer Review Submitted as Human Expertise - Although this may be thought to fall under both (1) and (2) in this section, the Reviewer role specifically requires that any AI-Generated content be called out as such. Any attempt to pass off AI-Generated review as Human Expertise is a severe violation.

Questions and Concerns about the AI Policy

Please reach out to aipolicy@jsmcah.org if you have concerns about this policy, requests for additional information, or requests for edits to the policy.