Overview

Lesson 5.3: Maintaining Research Integrity with AI

This lesson teaches researchers how to maintain research integrity standards while using AI, including proper documentation of AI use, ensuring compliance with institutional, journal, IRB, and funder requirements, and addressing the genuine ethical questions that arise when AI is involved in research. You will learn what documentation is required for reproducibility, how to disclose AI use in the contexts most researchers encounter, how to identify and mitigate AI bias in research outputs, and how to navigate the ambiguous compliance landscape as institutional standards continue to evolve. The lesson provides a comprehensive documentation template, a disclosure statement framework for each major compliance context, and a bias identification checklist for research using AI-generated content.

Title

Lesson 5.3: Maintaining Research Integrity with AI

Purpose

This lesson teaches researchers how to maintain research integrity standards while using AI, including proper documentation of AI use, ensuring compliance with institutional and journal requirements, and considering ethical implications of AI in research. You will learn what documentation is needed, how to disclose AI use appropriately, and how to assess whether AI use aligns with research ethics.

By the end of this lesson you will be able to: (1) create comprehensive AI use documentation that meets reproducibility standards; (2) draft compliant disclosure statements for journal submissions, IRB protocols, and grant applications; (3) identify potential AI bias risks in your specific research context; (4) apply a bias mitigation protocol to AI-generated research content; and (5) navigate ambiguous compliance situations through the correct institutional channels.

Core Concepts

Research integrity in the AI era rests on the same foundations it always has, transparency, reproducibility, and accountability, but AI use introduces specific new requirements for documentation and disclosure that are not yet universally standardized. Understanding the principles behind these requirements, not just the rules, is what allows researchers to apply them appropriately in novel situations.

The Documentation Imperative

Good research documentation enables two things: verification (can others confirm that what you say you did is what you actually did?) and reproducibility (could others follow your documented process and arrive at the same result?). AI use introduces two new documentation requirements. First, AI use must be documented with enough specificity to characterize AI's role: which tool and version, which task, what inputs or prompts were provided, what outputs were generated. Second, human oversight must be documented: what verification was performed, what was accepted unchanged, what was modified, and what was rejected.

Why document inputs and prompts? Because the same prompt to two different AI tools, or the same prompt to the same tool at different times, may produce different outputs. Documenting the prompt enables others to understand what was asked and to test whether the output was reasonable given the input. This is a new form of methodological transparency with no direct precedent in pre-AI research practice.

Compliance Landscape

Four types of institutional requirements apply to AI use in research, and they operate independently:

University/institutional policies govern AI use in academic work produced under institutional affiliation. Most institutions have updated policies between 2023 and 2026. When policy is silent, consultation with the research integrity office is appropriate before proceeding with novel AI applications.

Journal policies now include explicit AI disclosure requirements in most major journals. Nature journals require a specific AI disclosure statement. ICMJE guidelines (applied by most biomedical journals) prohibit AI authorship and require disclosure in methods. Most journals also prohibit use of AI as a peer reviewer. Always check the specific guidelines of any target journal before submission.

IRB and ethics review requirements are triggered when AI is involved in human subjects research, particularly when AI tools process participant data. Whether using an AI tool to analyze qualitative interview transcripts requires a protocol amendment depends on your specific IRB, how data are deidentified before being provided to the AI tool, and the terms of service of the AI platform. When in doubt, consult your IRB before proceeding.

Funding agency requirements are evolving. NIH and NSF accept AI use with disclosure as of 2026. Some DOD and DARPA programs apply stricter requirements. Some funders require that AI-assisted outputs are human-verified and that no AI-generated content appears as a primary scientific finding without independent confirmation.

Researcher Responsibility Principle

Regardless of how AI is used, the researcher remains fully responsible for all research content. AI is a tool, not a co-investigator, co-author, or responsible party. If AI generates an incorrect citation, it is the researcher's error. If AI-generated code produces wrong results and the researcher includes those results in a paper, it is the researcher's responsibility to correct them. This principle is non-negotiable and is explicitly stated in the authorship guidelines of all major journals and the policies of all major funding agencies.

AI Bias in Research

AI training data reflects patterns in the texts and datasets the model was trained on. Those patterns can introduce systematic bias into AI-generated research content in ways that are difficult to detect. Common AI bias risks in research include: recommendations that overrepresent English-language, Western, and high-income-country research while underrepresenting work from other contexts; suggestions that reflect historical biases present in the literature (e.g., in medical research, historically underrepresented patient populations); and code that applies assumptions appropriate for common data types but inappropriate for the researcher's specific population or context.

Recognizing that these biases may be present is the first step. Mitigation requires: explicit checking for representational bias (does the literature AI recommends systematically underrepresent certain populations?), review by colleagues with relevant expertise in underrepresented populations or contexts, and transparent acknowledgment in methods that AI was used and that outputs were reviewed for bias.

Practical Applications

The following protocols apply the core concepts to the four main compliance contexts researchers encounter.

Protocol 1: Documenting AI Use for a Journal Manuscript

Maintain a running AI use log throughout manuscript preparation, not retrospectively at submission. The log format: date | tool and version | task description | prompt summary (paraphrase, not necessarily verbatim) | output summary | verification performed | what was accepted/modified/rejected.

At submission, check the target journal's AI disclosure requirements in their author guidelines. Draft a disclosure statement from the log: 'The following AI tools were used in the preparation of this manuscript: [Tool 1, version] was used for [task 1, e.g., improving the clarity of the writing in the Methods and Discussion sections]. [Tool 2, version] was used for [task 2, e.g., generating initial code for the statistical analyses in Section 3]. All content produced with AI assistance was reviewed by all authors for accuracy, completeness, and fidelity to the original data and research findings.'

Place the disclosure statement where the journal guidelines specify (typically Acknowledgments or a dedicated AI Use Statement section).

Protocol 2: IRB Protocol for AI-Assisted Qualitative Analysis

Before using an AI tool to assist with analyzing interview transcripts or other human subjects data:

Step 1: Check whether your current IRB protocol permits this use. If the protocol does not mention AI tools, a protocol amendment may be required.

Step 2: Assess data sensitivity. If transcripts contain identifiable information, they must be deidentified before being provided to an AI tool. Confirm that the deidentification procedure meets your IRB's standard.

Step 3: Review the AI platform's terms of service to confirm that data you provide is not used for model training without consent. Enterprise versions of major AI tools typically provide this assurance; consumer versions may not.

Step 4: If the above steps raise concerns, consult your IRB directly before proceeding. Document the consultation and any guidance received.

Step 5: In the IRB protocol (amended or new), document: which AI tool, what data types will be processed by AI, deidentification procedures, and what human oversight will verify AI outputs.

Protocol 3: Addressing AI Bias in Literature-Based Research

When using AI to assist with a literature review or synthesis:

Step 1: Examine the recommended literature for representational balance. Are the suggested papers predominantly from English-language, US/European, or high-income-country contexts? If so, supplement with a targeted search of databases covering underrepresented contexts.

Step 2: Have a colleague with relevant expertise review any AI-generated synthesis for bias in framing, omission of relevant perspectives, or assumptions that reflect particular cultural or methodological contexts.

Step 3: In the methods section, acknowledge AI use in literature synthesis and note that representational balance was explicitly checked: 'AI tools were used to assist with literature identification and initial summarization. The resulting literature set was reviewed for representational balance and supplemented to include [specific population, language, or regional literature] as needed.'

Protocol 4: Disclosure for Grant Applications

Before using AI in any aspect of grant preparation, check the current guidance from the specific funding agency. For NIH, check the NIH Office of Extramural Research. For NSF, check current NOSIs and FAQs.

For the application itself: AI assistance in drafting narrative sections is generally acceptable with disclosure. All scientific claims, budget justifications, and regulatory statements must be independently verified. A disclosure statement in the narrative: 'AI writing assistance was used to improve the clarity of narrative sections of this application. All scientific claims, proposed procedures, budget rationale, and compliance statements were independently verified by the investigators and represent the investigators' own expertise and planning.'

Document which sections had AI assistance and what verification was performed, and retain this documentation in case of audit.

Key Takeaways

Comprehensive documentation of AI use enables verification and supports the reproducibility that is foundational to research integrity. Document tool, version, task, prompt inputs, outputs, and verification performed, not just that AI was used, but how it was used and how human oversight occurred.

Proactive disclosure demonstrates integrity and addresses potential concerns before they become problems. Waiting until a reviewer or editor asks about AI use creates an adversarial dynamic; volunteering disclosure in the manuscript positions you as a researcher who understands and respects emerging norms.

The compliance landscape for AI in research is multi-layered and evolving: institutional policies, journal policies, IRB requirements, and funder guidelines each apply independently and may differ. When standards are ambiguous, the correct move is consultation, not assumption.

Researchers bear full responsibility for all research content regardless of AI involvement. AI is a tool; you are the responsible party. This principle is non-negotiable and is universally adopted across journals, institutions, and funding agencies.

AI bias is a real methodological concern, not a hypothetical one. AI-generated literature recommendations, summaries, and analysis code can reflect systematic patterns from training data that introduce bias into research outputs in ways that are difficult to detect without deliberate checking.

Ethical practice extends beyond compliance. The questions 'Does this AI use align with my research values?' and 'Am I maintaining genuine intellectual ownership of this work?' are not answered by any compliance checklist. They require ongoing reflection as AI capabilities and norms continue to evolve.

AI Use Documentation Template

Use this template to maintain a running AI use log throughout any research project. The log serves both as internal documentation for reproducibility and as the source from which disclosure statements are drafted.

Project AI Use Log

Project title: _______________
Log started: _______________
Last updated: _______________

Entry format:

Date: _______________
Tool and version: _______________
Task category: [Literature search / Summarization / Writing assistance / Code generation / Data analysis / Other: ___]
Task description (1-2 sentences): _______________
Prompt summary (paraphrase): _______________
Output summary (1-2 sentences describing what AI produced): _______________
Verification performed: _______________
Content accepted unchanged: _______________
Content modified (describe changes): _______________
Content rejected (describe why): _______________
Verification date and verifier: _______________

Disclosure Statement Draft (from log)

For each distinct AI use logged, draft one sentence following this template:
[Tool name and version] was used to [task description]. [Describe verification performed and by whom]. [Note any modifications made to AI-generated content.]

Compile all sentences into a disclosure statement. Review the target journal's or funder's required format and adapt accordingly before submission.

Retention Policy
Retain the complete AI use log as part of the study's research documentation for at least as long as your institution's research record retention policy requires (typically 5-10 years post-publication). Store alongside raw data, analysis code, and other primary study documents.