Where AI Belongs in IRB Review, and Where It Doesn’t

November 16, 2026

Advarra’s position on AI in IRB review: what the board will always decide, where AI can fit in IRB operations, and the questions any research organization should be able to ask its IRB.

Most conversations about AI in research oversight start with what the technology can do. We want to start with what it won’t do at Advarra.

An institutional review board (IRB) has two parts that are easy to blur together. The IRB itself is the committee: the members who perform the ethical review of a study and determine if the research meets the regulatory requirements for approval. IRB operations are the staff and systems that support the submission from start to finish and support the Board with meeting scheduling, drafting minutes, retaining records, and other required administration.

At Advarra’s IRB, no AI system makes a determination about key ethical issues, including, but not limited to, risk, benefit, consent adequacy, and participant protection. A board member or committee applying ethical training, scientific expertise, and human judgment does. AI supports IRB operations by streamlining administrative tasks and identifying potential issues within submissions. All AI outputs undergo human review and verification before they are provided to the Board for ethical review.

We’ve provided independent IRB oversight for more than 40 years, and strict firewalls between our IRB and our business and commercial operations protect that independence. Advarra also develops AI technology, which is exactly why those firewalls matter here.

We’re stating this now because the questions have already arrived. Protocols that describe AI in their design or conduct are becoming more common. Human research protection programs (HRPPs) are asking how their own staff and reviewers should use these tools, and Public Responsibility in Medicine and Research (PRIM&R), with grant support from Advarra, has launched an initiative on AI ethics and governance. The sponsors, sites, and participants who rely on an IRB should know where its boundaries are before AI becomes part of its daily work.

Speed belongs in this conversation, so we’ll address it directly. Our IRB is designed to work quickly while keeping participant protection central to every review. We process submissions ethically, as well as efficiently, and AI, used in the right places, can help improve that efficiency while remaining subject to regulatory or ethical requirements.  

We’ve both been in IRB work long enough to remember when consent meant a paper form and a wet signature, and when submissions arrived in the mail. We were there for the move to digital submissions and electronic consent. Each of those changes altered how sponsors, sites, and participants worked with the IRB and each other, and each one shifted how IRB operations ran day to day. The standard of review stayed where it was. We expect AI to follow the same path, provided the boundaries are set first.

The standard of review doesn’t change

The ethical principles articulated in the Belmont Report (respect for persons, beneficence, and justice) have stood the test of time. They were designed with enough flexibility that later generations could apply them in the context of their own time and their own society. The federal regulations for human subjects protections were built the same way.

Take the first criterion for IRB approval under 21 Code of Federal Regulations (CFR) 56.111 and 45 CFR 46.111: Risks to subjects are minimized by using procedures consistent with sound research design. What counts as sound research design has changed a great deal since those regulations were written, and AI tools are now part of it. A protocol may use AI for numerous purposes, including, but not limited to, identifying eligible patients, supporting enrollment decisions, building a synthetic control arm, analyzing endpoints, and detecting safety events.

The criteria still apply to every one of those uses. Reviewers still ask whether risks are minimized and reasonable in relation to anticipated benefits, whether subject selection is equitable, whether the plan for monitoring data is adequate, and whether privacy and confidentiality are protected, among other criteria. For an AI component, that means asking whether it works as designed without introducing new risk, and whether the data it collects or analyzes can be trusted and protected. If it introduces a reasonably foreseeable risk, the consent form has to describe that risk (21 CFR 50.25(a)(2); 45 CFR 46.116(b)(2)).

Until future regulation or guidance establishes different requirements, IRBs retain the flexibility to evaluate AI under the existing regulatory framework, just as they have adapted their review to other advances in research methods. While AI introduces new considerations, IRBs can continue to structure their reviews within their established model.

We hold our own use of AI to that same standard.

Where we draw the line: The board

These are institutional commitments for Advarra’s IRB:

  • AI will not be used in place of a board member applying ethical training, scientific expertise, and judgment to research.
  • AI will not replace the committee for determinations about safety or ethics.
  • AI will not be asked to make value-based decisions.

The reason is straightforward. Generative AI models produce outputs based on patterns learned during training. In our view, they don’t understand ethics or context the way a trained IRB professional does, and an ethical determination depends on both.

The ways AI can fail illustrate this point. A model can hallucinate, describing something in a protocol that isn’t there with the same confidence it describes what actually exists in the protocol. It can make ordinary errors, such as misreading a table or working from the wrong version of a document. It can reflect biases in its training data, which may underrepresent the very populations that equitable selection is meant to protect. It can produce an incomplete analysis, leaving out the one detail that changes a risk assessment.

A less immediate risk is the gradual erosion of reviewer expertise. When reviewers increasingly rely on AI to interpret submissions, they may lose proficiency in the close reading and critical analysis that rigorous review requires. This decline may be difficult to detect in any single review, becoming apparent only over time as review quality diminishes and rebuilding those skills becomes more challenging. Clinicians are discussing this concern in clinical practice, and it also applies to research.

We’ve made a similar point about AI translation of consent forms: The tools produce useful drafts, but human translators remain essential for judging context, cultural appropriateness, medical terminology, and the intent behind the original consent language. Ethical review is a higher bar still. Relying on AI to review an application in place of a reviewer would be an abdication of professional responsibility. Humans must always decide whether it is ethical to experiment on another human being, and there will always be humans in the room making that decision.

Where AI belongs: IRB operations

IRB operations staff perform detailed submission work that is critical for IRB review but can be highly time-consuming. As AI comes into IRB operations, these are a few places we think it fits:

  • Data extraction. A tool can pull information from protocols and supporting documents into submission fields, so submitters and staff spend less time transferring information that already exists.
  • Completeness checks at intake. A tool can check a submission for missing documents or mismatched information before it reaches the board, with a staff member verifying the result. A gap found at intake costs less time than one found during review.
  • Drafting for human review. AI can prepare first drafts of minutes and approval letters, which record decisions the Board has already made. It can also draft consent form language, which the board then reviews and approves on the same terms as consent language from any other source.

Each requires the same safeguard: A person must verify that the AI tool’s output accurately reflects the relevant source materials and, where applicable, the Board’s decisions.

To decide whether any proposed AI use belongs in IRB operations, we propose three requirements:

  • It supports IRB administrative work while preserving human judgment.
  • Its output is verified by a qualified person who takes responsibility and accountability for its use.
  • It preserves the Board’s access to information and independent decision-making.

Any proposed AI use must meet all three criteria to be appropriate for our IRB operations.

What to ask any IRB, including us

You can put the same five questions to any IRB you work with, whether you’re a sponsor, a site, or an HRPP. Here are our answers:

  • How and where is AI used? AI does not make ethical determinations. For IRB operations, the requirements above define the framework for its use.
  • Are any determinations made or recommended by an AI system? No. Board members make every determination.
  • Who checks each AI output, and how? A staff member who is responsible and accountable for the output checks the AI output against the submitted materials before it moves forward into the review process.
  • Is there a process for training and governance of AI use by IRB staff? Staff should receive role-specific training and AI use should be governed by defined policies and standard operating procedures (SOPs), human oversight, and clear accountability.
  • How does the IRB review AI components embedded within the research it oversees? The IRB reviews them against the same Belmont principles and approval criteria as every other part of the study, including whether the consent form discloses any reasonably foreseeable risks that the use of AI introduces.

Final thoughts

AI has the potential to accelerate clinical research and advance human health. Realizing that potential requires thoughtful implementation, clear operational safeguards, and human accountability to uphold the IRB’s regulatory obligations and responsibility to protect research participants.

From paper submissions to digital platforms and electronic consent, IRBs have adapted to new technologies while upholding their ethical responsibilities. Those responsibilities will continue to guide us as AI evolves.

Learning more about responsible AI in clinical research

For a broader look at putting responsible AI into practice across clinical research, read Advarra’s industry report, Operationalizing Responsible AI in Clinical Research, and learn more about PRIM&R’s AI Ethics and Governance Initiative.

Advarra supports the initiative with a grant. PRIM&R develops and leads its programming.

Thomas Closson

Thomas Closson

Senior Vice President, Global Review Services

View all posts
James Riddle

James Riddle, MCSE, CIP, CPIA, CRQM

SVP, Global Review Services

With 25+ years’ experience providing support to the clinical research community, James helps sponsors, CROs, and research sites advance clinical research with a mission to improve human health.

View all posts

Advarra to your inbox

Be the first to know about
new content, products, and
services from Advarra. Sign up for our newsletter and stay in the loop!

Login
Scroll to Top