Ethical Controversy Over Researchers — and Soon, IRBs — Using AI Tools

By Stacey Kusterbeck

Institutional review boards (IRBs) are starting to see study protocols that at least in part were written by artificial intelligence (AI) tools. “Our research coordinators follow a format when submitting something that we expect them to follow, and there’s pretty good consistency. We’re getting to a point now where we say, ‘This doesn’t sound right or accurate. Who wrote this?’” says Thomas D. Harter, PhD, director of the Department of Bioethics and Humanities and chair of the IRB at Emplify Health-Gundersen.

Some study protocols did not follow the expected format based on the institution’s operating procedures, so IRB members concluded that the protocols were written by AI. “Someone wrote it, but they were pulling from AI — and you can’t do that. If you are writing a paper with sound methodology, there has to be accountability and a human signing off on it,” says Harter. If detected, the IRB should ask the research team if AI was used and, if so, either to rewrite it or add a disclosure statement as part of the protocol indicating that AI was used and where, adds Harter.

Some researchers use AI to create scientific methodology for studies or to come up with a list of references. The AI tool might recommend a research question, which then becomes the basis for the study protocol. That does not necessarily make the study protocol unethical — it is possible that the research question is valid and appropriate. “But we have to rethink how we are engaging with AI in how we are learning about research. As we know, AI is not infallible. Sometimes it makes up stuff that doesn’t actually exist or cites papers that are what we call phantom research. As IRBs, we want to address that,” says Harter.

Since the issue with AI is relatively new, IRBs are handling it inconsistently. Some already have policies on AI use in place, others are developing policies currently, and still others have not yet considered this issue. Emplify Health-Gunderson’s IRB recently amended the application used by researchers to submit their research protocols to the IRB to include questions about whether AI was used in writing the protocol.

“There has to be trust,” says Harter. The IRB has to assume, as a starting point, that the principal investigator (PI) is a capable researcher, is knowledgeable about what they are doing, and has used an ethical, appropriate process to create the study protocol. The PI has to assume that the IRB knows enough about the study protocol’s contents to evaluate it. “The problem is, the PI says, ‘I’m not overly worried about the ethics or merit of this, because the IRB’s job is to catch those errors.’ And the IRB says, ‘We’re relying on the PI to be sure the study is scientifically sound.’ They are both relying on each other,” says Harter.

There are ethical concerns that inappropriate use of AI tools could result in research that is based on incomplete or inaccurate information, or it could subject research participants to unnecessary harms. Another concern is compromised integrity of medical research processes that could negatively affect or harm the general public regarding long-term health outcomes, says Harter.

Researchers’ current use of AI includes analyzing data for cohort eligibility and rewriting consent forms in plain language. “I suspect many researchers also use AI to prompt new research ideas or finetune their study designs, even if it’s not openly discussed. Right now, we don’t have a clear picture of how widely tools like ChatGPT are being used in research. It’s happening, but potentially unevenly across fields and demographics. Until disclosure is required, institutions won’t really know how much AI is shaping the research-design process,” says Challace Pahlevan-Ibrekic, MBE, CIP, director of regulatory affairs at The Feinstein Institutes for Medical Research.

According to Pahlevan-Ibrekic, the issue is not whether AI was used, but how and to what extent it influenced the research design process. Research proposals that have been “substantially developed” by AI are not considered to be original ideas of applicants, according to a 2025 guidance from the National Institutes of Health.¹

“Since there is currently no federal standard governing AI use in research development, institutions must define their own thresholds for acceptable AI use,” says Pahlevan-Ibrekic. Asking researchers to disclose whether and how AI tools were used in protocol development and to take full responsibility for verifying all AI-generated outputs to ensure they are valid, fair, unbiased, reliable, complete, and align with their institution’s professional and ethical standards, policies, and regulatory requirements, is a good start, says Pahlevan-Ibrekic.

“It is not useful, or accurate, for IRBs to treat AI as categorically problematic. Instead, IRBs can help ensure AI use aligns with the institution’s standards for integrity, transparency, and human accountability,” offers Pahlevan-Ibrekic.

Some research ethics experts view AI tools as potential solutions to some longstanding problems facing IRBs, such as inefficient, inconsistent processes. “Some people have suggested using AI to help speed up reviews and standardize evaluations, but this raises big ethical questions we don’t yet have good answers to. We wanted to take a closer look at those issues,” says Brian D. Earp, PhD, an associate professor of biomedical ethics at the Yong Loo Lin School of Medicine at National University of Singapore.

Delays in securing IRB approval can cascade into broader research setbacks and missed opportunities, such as postponement of study advertisement, enrollment, and data collection. “Delays can place considerable pressure on investigators to rush their studies and produce results, especially when projects are tied to grant timelines. Ultimately, such delays can inadvertently hinder the production of socially valuable knowledge and slow scientific progress,” says Joel Jiehao Seah, PhD candidate at the National University of Singapore's Centre for Biomedical Ethics. “Many IRBs operate with limited manpower relative to the growing volume of research conducted by their institutions because of underfunding. In the U.S., small IRB offices are not uncommon. Some are, quite literally, an office of one,” observes Seah.

Earp, Seah, and colleagues propose that IRB-specific large language models could increase the quality and efficiency of research oversight.² AI could help speed things up by flagging missing information or obvious errors, allowing human reviewers to focus on tougher ethical questions. “But there are big concerns, too: AI systems can embed hidden biases, may not be transparent, and could tempt committees to rely too heavily on automation. There’s also the problem of accountability. If the AI misses something important, who is ultimately responsible?” asks Earp.

IRBs must ensure that the AI tools support human judgment, rather than replace it. “That means transparency about how tools work, ongoing checks of their performance, and strong protections for the confidentiality of research protocols,” says Earp. Institutions will need clear guidance on how these tools should be tested and monitored. “Any system would need to be tailored to local circumstances, not just rolled out as a one-size-fits-all solution,” adds Earp.

Ethicists can play a major role by helping to write guidance for how AI tools should be used by IRBs. “Part of this should involve clear training of committees on the limits of AI and the risks of misuse,” says Earp.

Although IRBs currently are not employing generative artificial intelligence (GenAI) tools to support their reviews and ethical oversight, some are becoming more open to the idea, reports Seah. The Association for Accreditation of Human Research Protection Programs, an accreditation body for IRBs, recently held a webinar on using AI to support IRB reviews. “I anticipate that for-profit commercial IRBs may be the first to integrate GenAI in such ways, as they have the means to build and deploy these systems rapidly,” Seah predicts.

When AI is used to support clinical decision-making, users (physicians and other healthcare professionals) generally possess the knowledge and skills gained through medical education to critically evaluate the AI tool’s recommendations. “By contrast, in the context of IRBs using AI to support ethical judgments, IRB members may lack the formal ethical or moral expertise needed to accurately assess and question the AI’s outputs,” says Seah. IRB members typically include clinicians, scientists, lawyers, religious leaders, or lay representatives. Many IRBs do not include research ethicists or moral philosophers among their members, observes Seah.

“Most existing ethical guidance, frameworks, and institutional policies focus on GenAI in education, research, medicine, or administration. To date, none specifically address its deployment in research ethics oversight and ethical review,” says Seah.

For IRBs, AI tools can streamline workflows that do not require substantive ethical judgments. That includes administrative tasks in low-risk, high-reward areas, such as intake review classification or to verify submission completeness, says Pahlevan-Ibrekic. ”AI tools can also serve as an IRB reviewer assistant to confirm that all applicable ethical frameworks, regulations, and policies have been considered,” says Pahlevan-Ibrekic. This allows IRB members to evaluate deeper ethical considerations that require contextual understanding and nuanced regulatory application.

“That said, AI is prone to errors and could miss ethical oversights in novel or complex protocols, and is not reliable for more comprehensive regulatory or ethical reviews,” cautions Pahlevan-Ibrekic. For pre-review processes that assess completeness of the submission as well as the quality of the information provided, AI alone is insufficient. “Nuanced judgment remains essential; AI should augment, not replace, human regulatory and ethical review,” underscores Pahlevan-Ibrekic. Institutions must translate guidance into policies aligned with their mission and values. Those can vary across research centers and hospital systems. For instance, hospital systems may prioritize patient data protection, whereas academic institutions may view transparency and research integrity as paramount.

“Ethicists should partner with educators to embed ethical AI literacy into curricula and research integrity frameworks, ensuring ethical judgment remains human-centered and AI-assisted, not AI-dominated,” concludes Pahlevan-Ibrekic.

Stacey Kusterbeck is an award-winning contributing author for Relias. She has more than 20 years of medical journalism experience and greatly enjoys keeping on top of constant changes in the healthcare field.

References

1. National Institutes of Health. Supporting fairness and originality in NIH research applications. Released July 17, 2025. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html

2. Porsdam Mann S, Seah JJ, Latham S, et al. Chat-IRB? How application-specific language models can enhance research ethics review. J Med Ethics. 2025; Aug 19. doi: 10.1136/jme-2025-110845. [Online ahead of print].