Using AI in Human Subjects Research

Generative AI presents certain challenges and concerns when used to describe aspects of a planned research protocol submitted for IRB review, including but not limited to: 

• Risk of plagiarism: Generative AI may inadvertently produce text containing plagiarized content due to its extensive training corpus and obscured mappings within the language model, making plagiarism difficult to detect. 

• Factual accuracy: Generative AI can generate text containing inaccurate or fictitious information. Submissions may unknowingly incorporate false information when using responses generated by AI. Review the Terms of Use for your AI tool for cautionary statements about the potential for factually inaccurate outputs.

• Incorrect interpretations of relevant regulations: Asking Generative AI a general question about human subjects research requirements may return an inaccurate response.

• Citation deficiency: Generative AI often fails to provide proper citations for referenced sources, as the model's learned content mappings are not explicitly revealed. 

• Data incorporation: Generative AI may assimilate novel ideas from other research projects into its database, potentially using such information to generate content for other users.

• Conflicting human subjects research policies and procedures at different institutions: Generative AI may return a response derived from one or several websites of various institutions; but institutional policies and procedures vary, and the response provided by Generative AI may well differ from SDSU policies and procedures.

• Variations in different research methods, sites, and populations: A common phrase in the IRB is, “it depends.” Requirements for recruitment, consent, and confidentiality will depend on the target population, the study design, and U.S. or other regulations that may apply to the research. Generative AI is unlikely to return an adequate or accurate response to questions about specific research.

• Lack of specificity: Somewhat related to the previous point, use of Generative AI to answer certain questions tends to provide a very general response rather than one specific to your project. For example, there is an initial application question that asks about how and where informed consent will be obtained, by whom, and how potential for coercion will be prevented. We have seen Generative AI responses to this question that address why these things are important, but do not explain the planned consent process for your study.

• Inconsistency in form responses: Use of Generative AI to answer individual questions in the iRIS submission form can lead to inconsistencies in the submission as a whole. For example, using Generative AI to write a response to the question in the IRB Initial Application asking for a description of study procedures and then to write responses to subsequent questions in the application often leads to listing different study procedures in different places in the application. 

• Consent forms: Use of Generative AI to write consent forms often produces documents that 1) do not include all the required elements of informed consent; 2) do not include all the required information found in SDSU consent templates; and/or 3) include incorrect information for SDSU IRB protocols. For example, using Generative AI to write consent forms often prompts users to add contact information for “AVP of Research Support Services” (RSS) if people have questions or concerns about their rights as research participants, but at SDSU, participants contact the Human Research Protection Program (HRPP) about these matters and not RSS directly.

• Impacts on time to approval: For the reasons outlined above, submissions relying on the use of Generative AI may contain incorrect or inconsistent information. This will result in more revision comments and possibly additional rounds of revisions. HRPP recommends that researchers review and verify the information included in Generative AI outputs before incorporating the information into your submission.

SDSU acknowledges the potential benefits and challenges of employing Generative AI in providing answers to the questions in the iRIS IRB forms.

• Open for Use: There are currently no limitations on the use of Generative AI in the preparation of submissions to the IRB. Researchers are free to utilize Generative AI tools as they deem appropriate, in accordance with established SDSU GenAI Guidelines and SDSU IT Sensitive Data Storage Best Practices.

• Appropriate Evaluation: All submissions, whether or not they incorporate Generative AI, will be reviewed via the same criteria outlined in federal regulations and SDSU policies and procedures. However, if GenAI is used there may be greater scrutiny of the study materials to ensure compliance with all applicable rules and regulations.    

“I’m uploading my research aims and my research procedures. Use this to write a 3 sentence summary of my research aims for a consent form using lay language for my participants. Use this to write a summary of what participants will be asked to do if they decide to take part in the study using lay language and diagrams or charts where it makes sense. Include information a reasonable person would want to know including the topics discussed or questions asked and details about how any devices work or the experience of using them will be like. Avoid jargon. These summaries will be included in the consent form to ensure participants understand what the research study is about and what they will be asked to do. Do not create additional procedures or alter the scientific intent. Flag any inconsistencies or missing details or information you need to do this.”

“I’m uploading my research aims and my research procedures. Create a recruitment flyer and recruitment post for social media for this research study. Use plain and neutral language. Maintain accuracy and avoid persuasive or coercive phrasing.”

“I’m uploading all the documents I will submit for IRB review. Please conduct a preliminary review of the materials against IRB regulations. Please conduct a preliminary review of the materials against San Diego State University IRB policies and expectations outlined here: https://research.sdsu.edu/research_affairs/human_subjects/hrpp-sop-8-25.pdf. Flag anything that may cause issues for the IRB review process. Flag any prompts in the protocol form that aren’t fully answered or addressed but should be, based on the information in the protocol document. Flag any inconsistencies or missing information. At the end, ask me if I want help drafting responses to address any of the issues using guidance from https://research.sdsu.edu/research_affairs/human_subjects/guidance.”

“I’m uploading my consent forms. Review them and let me know if any parts of it might be confusing for a lay reader. Identify any accessibility issues as well.”

Paste your recruitment letter and/or sections of your IRB consent form into Gemini, Copilot, or ChatGPT and ask:

• “Can you make this more concise and in an academic in tone?”

• “Can you rewrite this for a 6th grade reading level?”

• “Can you make sure this consent form covers the purpose, procedures, duration, risks, benefits, alternatives, and voluntariness – and makes sure the information is clear and concise?”

• The use of public, non-secure AI tools (e.g., Claude, DeepSeek, Bard) for analyzing or storing sensitive participant data is prohibited unless appropriate safeguards and agreements (e.g., data use agreements, BAA) are in place. 

• AI applications that operate without meaningful human oversight in high-stakes decisions involving participants are subject to enhanced review and may be restricted or denied.

When using generative AI to create stimuli (e.g., images of people, scenarios), researchers must: 

• Not advertise this content as real or factual. 

• Clearly differentiate AI-generated materials from real-world data during consent and debriefing. 

• Ensure that generated content does not reinforce harmful stereotypes or misrepresent identities.