Foundational principles

Any model for consent must consider the ethical foundation of voluntary, informed consent and the benchmarks established for the protection of human subjects [Reference Miracle11–Reference Bierer13]:

• • Disclosure – Potential subjects must be given all relevant information about the research. This should be in clear, understandable terms and include the study’s purpose, procedures, risks, potential benefits, alternative options, and the right to withdraw at any time [Reference Wisgalla and Hasford9].

• • Comprehension – Providing information is insufficient; participants must also understand it. Investigators have an obligation to facilitate and verify comprehension before accepting consent [Reference Wisgalla and Hasford9].

• • Voluntariness – Regulations require that consent be entirely voluntary, made freely without coercion or undue influence, and that individuals are informed of their unconditional right to refuse or withdraw at any time without penalty [14].

• • Capacity – To be competent to consent, a potential participant must be positioned to understand the information, appreciate its implications, consider options, and communicate a decision [Reference Kadam7]. When a person lacks capacity, consent must be obtained from a legally authorized representative and often with special safeguards in place.

• • Documentation – In the US, regulations require a signed, written informed consent document, unless an institutional review board (IRB) has waived the requirement for documentation [15].

Requirement for oversight of the consent process

While an investigator is accountable and responsible for ensuring informed consent satisfies foundational principles, there are no laws, regulations, or ethics guidelines requiring that informed consent must be conducted by a human. Regulations mandate that information be presented in an understandable way, that the participant’s agreement is truly voluntary, and that consent is appropriately documented [14], but they do not prescribe how this occurs. Guidance by the FDA and the Office for Human Research Protections permits electronic or multimedia consent provided required elements are included and the process supports participant understanding and voluntariness [16,17]. Meaningful informed consent depends not only on disclosure and documentation, but also on participant understanding of the information provided [Reference Wisgalla and Hasford9].

Electronic consent has been increasingly adopted to help lower barriers to achieving a valid informed consent [Reference Vogt-Yerem, Heck and Gjorgjieva18,Reference Doerr, Moore and Barone19]. Many eConsent platforms improve accessibility through electronic document delivery, multimedia educational materials, and remote consent workflows. However, these approaches frequently remain digital representations of traditional consent forms and often provide limited opportunities for personalized, interactive dialog. In contrast, an interactive computer system (artificial intelligence (AI)-agent) is designed to conduct a dynamic consent conversation that adapts to participant responses in real time, supports iterative clarification and teach-back, and provides auditable documentation of participant understanding. An interactive AI-based system could therefore provide a more engaging and participant-centered consent experience [Reference Allen, Schaefer, Porsdam Mann, Earp and Wilkinson20].

AI-driven consent

We have designed a technology driven by AI that represents a novel approach to the informed consent process. The AI agent conducts the informed consent conversation, presenting study information in plain language tailored to participant characteristics, assessing comprehension through teach-back, providing clarification as needed, and documenting voluntary agreement. In this way, the AI agent, performing selected functions commonly undertaken by a clinical research coordinator (CRC), is designed to meet the traditional benchmarks for disclosure, understanding, and voluntariness. This technology enables human CRCs to focus on addressing complex questions and building relationships with participants.

This AI-driven approach offers unique advantages compared to traditional consent. Compared with traditional consent, the system delivers standardized information, tailors explanations to participant literacy and context, supports unlimited participant questions, and is available asynchronously. The chatbot automatically records the consent dialog, creating a detailed record that can be reviewed by investigators or IRBs to assess how study information was communicated and discussed. These transcripts may also support monitoring, quality improvement, and oversight activities by providing a traceable record of consent interactions.

To satisfy the legal requirement of a signed informed consent form, the transcript is accompanied by formal documentation of consent, which participants can sign and retain for their records. Nevertheless, institutional and regulatory acceptance of AI-mediated consent processes may vary across IRBs and research settings, particularly as governance frameworks for clinical AI applications continue to evolve.

Validation of approach

In collaboration with our ethics and IRB teams, we have developed and orchestrated a set of three AI agents, as illustrated in Figure 1. These agents leverage commercial large language models (LLMs, i.e., Open AI’s GPT variants), while restricting its operational context to IRB approved information related to the target study. Specifically, we designed and implemented:

Figure 1.

Illustration of the workflow used to generate synthetic consent sessions using three AI agents, namely CRC, participant and evaluator AI agents. The CRC AI agent has access to the IRB approved informed consent form, IRB approved trial protocol, and participant eligibility. The participant AI agent only has access to the participant profile and eligibility. The evaluator AI agent has real-time access to the transcript of the CRC and participant interactions for the purpose of producing the evaluation metrics (Figure 2) and informing the CRC AI agent of areas in need of attention. At the conclusion of the consent process, participants are provided with IRB-approved documentation of their consent. This typically includes a copy of the signed informed consent form in PDF format. In addition, participants receive a plain-language summary of the consent conversation and key study information, ensuring that essential details are accessible and understandable. Importantly, the CRC AI agent chatbot remains available to participants throughout the duration of the trial, allowing them to revisit information or ask additional questions at any time. This ongoing accessibility supports participant understanding, autonomy, and engagement beyond the initial consent interaction. Colors distinguish system components: blue indicates approved study-source inputs, green for the CRC and Evaluator AI agent, orange the simulated participant AI agent, and gray session outputs. Arrows indicate the direction of information or prompt flow. The upper Evaluation Summary represents the evaluator’s session-level assessment used to guide the CRC AI agent, whereas the lower Evaluation Summary contains the final participant-performance metrics generated after the session.

Figure 1. Long description

Long description is in the word document in text

• • A CRC AI agent to conduct the informed consent process. The CRC AI agent had access to an IRB-approved informed consent form, relevant portions of the IRB-approved target study protocol, participant eligibility information, and individual participant characteristics. The CRC AI agent was restricted from accessing external information sources during the consent interaction, thereby limiting responses to protocol-specific and IRB-approved content. Additionally, the CRC agent was equipped with operational guardrails intended to mitigate hallucinations, coercive language, and inappropriate conversational behavior. Potential coercive or undue influence behaviors were operationally defined as the use of persuasive, directive, or emotionally manipulative language encouraging enrollment despite participant hesitation or uncertainty. Safeguards included strict contextual grounding to IRB-approved study materials, constrained prompting that prohibited generation of unsupported information, refusal behaviors when requests exceeded the approved knowledge scope, refusal to provide advice of any kind, escalation to a human CRC when comprehension or voluntariness is in doubt, and audit logging of all interactions to support oversight.

• • A participant AI agent to simulate prospective participant interactions during the consenting process, having access to the participant characteristics (e.g., demographics), eligibility information, and a randomly generated behavioral profile used for the consent simulation. The purpose of this agent is to enable rapid and scalable synthetic testing of the system given a new target study, thus reducing the need for extensive testing and evaluation before deployment.

• • An evaluator AI agent, operating independently from the CRC AI agent through separate prompts and contextual inputs, was used to assess participant understanding through recall fundamental aspects of informed consent. We developed metrics for fundamental aspects of consent (Figure 2) that were quantified by the evaluator AI agent on a scale of 0 to 1. The scoring framework was designed to operationalize participant comprehension using a structured teach-back approach, in which higher scores reflect increasing evidence of participant understanding through explanation in the participant’s own words [Reference Kripalani, Bengtzen, Henderson and Jacobson21]. Specifically, we used a metric scale of [0, 0.25, 0.5, 0.75, 1] to assess participant responses: a score of 0 indicates the concept of interest was not mentioned; 0.25 that it was briefly mentioned but not explained without a response from the participant; 0.5 that it was adequately explained by the CRC AI agent without response from the participant; 0.75 reflects instances where the concept is adequately explained by the CRC AI agent and elicited either acknowledgment or requests for clarifications by the participant; and a score of 1 is given when the participant provides a satisfactory explanation of the concept in their own words, during the teach-back prompt initiated by the CRC AI agent. The scoring framework was developed for this initial feasibility study and has not yet been formally validated against expert human raters or established comprehension assessment instruments.

Figure 2.

Metrics of consent (Left) and performance characteristics (Right) of the in silico experiment with the following studies: PROMISE (NCT01174550), COMPARE-VOE (NCT03933397) and PROMISE-Lung (NCT06399302). Therapeutic misconception was assessed by whether participants demonstrated understanding that the purpose of the study was to generate generalizable knowledge rather than provide individualized clinical care. The “compensation and treatment” metric assessed participant understanding of the availability of compensation and medical care in the event of a research-related injury, as described in the informed consent document. ICF indicates informed consent form. Results summarize a total of 500 simulated consent sessions. Scores were assigned as follows: 0, concept not mentioned; 0.25, concept mentioned but not adequately explained; 0.5, concept presented adequately but without participant response; 0.75, participant acknowledgment or request for clarification; and 1, satisfactory explanation in the participant’s own words during teach-back.

Figure 2. Long description

Long description is in the word document in text

Additional details of the models, settings and prompts are provided in the Supplementary Materials.

In silico testing was conducted using the protocol and informed consent form from three distinct clinical studies representing different research contexts: two large comparative effectiveness trials, PROMISE (NCT01174550) [Reference Douglas, Hoffmann and Patel22] and COMPARE-VOE (NCT03933397) [Reference Knisely, Barnhart and Ibemere23], and one large observational registry, PROMISE-Lung (NCT06399302) [Reference Neely, Mongodin and Whitson24]. For the in silico testing, participant AI agents were created based on the trial eligibility criteria, to simulate distinct virtual participants of varying clinical and demographic characteristics (Figure 3). Simulated participants were also assigned varying demographic, clinical, and behavioral profiles to model heterogeneous consent interactions and broader population variation. We simulated 500 consent interactions (PROMISE: 179, COMPARE-VOE: 165 and PROMISE-Lung: 156) between the CRC and participant AI agents. During iterative development, selected simulated interactions were qualitatively reviewed by members of the study team to assess whether the CRC AI agent remained faithful to the IRB-approved material, stayed within the approved information scope, avoided directive or coercive language, and appropriately initiated escalation under predefined conditions. These qualitative reviews informed prompt refinement but were not conducted as a prespecified, systematic evaluation of factual accuracy, completeness, or source fidelity, nor were they intended to formally validate the CRC AI agent’s performance.

Figure 3.

Illustration of the application front end used for the in silico testing. The application screen capture shows the clinical (3a) and behavioral (3b) profile of the participant. Figure 3c shows the chatbot window, where the CRC AI initiated the consenting session by greeting the participant. Figure 3d shows the consenting form that the participant sees at the end of the process if he/she chooses to consent. All displayed participant characteristics, behavioral profiles, and dialogue are synthetic.

Figure 3. Long description

Long description is in the word document in text

The current interface starts with a concise summary by the CRC AI agent of a study’s key information, in accordance with FDA guidance. The participant AI agent is given the opportunity to ask clarifying questions before walking through the rest of the consent process (Figure 3c). Responses by the CRC AI agent are personalized based on context and the wording used by the participant. As illustrated in Figure 1, the evaluator AI agent monitors the conversation and adjusts the prompts for the CRC AI agent accordingly to ensure each evaluation metric is assessed. The participant may leave the conversation at any time. If an acceptable score on each evaluation metric required for consent is achieved, the evaluator AI agent instructs the CRC AI agent to wind down the session and the participant AI agent has an option to proceed to consent (Figure 3d). If the scores or teach-back is unsatisfactory, the CRC AI agent is prompted to revisit and clarify the concept with the participant until a satisfactory teach-back. If repeated teach-back attempts remained unsuccessful, if participant responses suggested uncertainty regarding voluntariness or comprehension, or if questions exceeded the predefined scope of approved study information, the system was designed to escalate the session to a human CRC for additional support and review.

The intended deployment model is human supervised rather than fully autonomous. The CRC AI agent facilitates delivery of study information and assessment of participant understanding, while ultimate responsibility for the informed consent process remains with the principal investigator and study team. Participants may request assistance from a human study team member at any time, and interactions are escalated immediately when predefined concerns regarding comprehension, voluntariness, or participant preference arise. In the absence of escalation, consent interactions are documented for subsequent review by the principal investigator and study team before any study-related procedures occur, and confirmation that informed consent requirements have been satisfied remains the responsibility of the study team.

Results from the 500 simulated consent interactions characterize performance on predefined teach-back and consent-coverage metrics within the simulated environment (Figure 2). Across the ten fundamental metrics, 4,056 of 5,000 metric-level observations (81.1%) received the maximum score of 1, indicating the concept was addressed and the simulated participant provided a satisfactory teach-back response. The proportion receiving a score of 1 was 86.5% for PROMISE-Lung, 79.4% for COMPARE-VOE, and 77.1% for PROMISE. An additional 7.7% of observations received a score of 0.75, 6.5% scored 0.5, and 0.7% scored 0.25. In the remaining 3.9% of cases, the fundamental concept was not addressed before the session ended.

At the session level, 96 of 500 simulations (19.2%) ended before all ten fundamental consent elements received a score of 1, and none of these simulated participants reached consent. Among the 404 simulations in which all ten fundamental elements received a score of 1, 284 resulted in consent, corresponding to a consent rate of 70.3%.

Among individual fundamental metrics, the lowest average scores were observed for compensation and medical care available in the event of a research-related injury (1: 81%, 0.75: 0%, 0.5: 2%, 0.25: 1%, 0: 16%), point of contact (1: 81%, 0.75: 0%, 0.5: 2%, 0.25: 2%, 0: 15%), and ability to withdraw from the study (1: 81%, 0.75: 8%, 0.5: 6%, 0.25: 1%, 0: 5%). Secondary metrics, which are used solely for monitoring and quality improvement, are not presented here.

This study has several important limitations. First, evaluation was conducted entirely through in silico simulations across three clinical study contexts rather than in human participants, and therefore the findings should be interpreted as preliminary evidence of feasibility and system behavior rather than real-world consent effectiveness. Although the evaluator AI agent operated independently from the CRC AI agent using separate prompts, restricted contextual inputs, and predefined scoring criteria, both agents were derived from the same underlying LLM ecosystem, introducing potential risks of circular evaluation, automation bias, and shared model bias. In addition, factual accuracy, completeness, and fidelity of CRC AI responses to the informed consent form and protocol were not independently evaluated. Finally, the operational safeguards were incorporated into the system design and examined qualitatively during iterative development, but their effectiveness was not formally evaluated through systematic adversarial stress testing or red-team testing.

Under close human oversight, the consent chatbot is now entering pilot testing with real participants to evaluate usability, participant comprehension, acceptability, and operational feasibility in clinical research settings. Future evaluation will also assess system performance across diverse participant populations and research environments. Participant trust and acceptability will also require careful evaluation, as some individuals may prefer direct interaction with human research staff, while others may benefit from the accessibility, personalization, and flexible pacing enabled by AI-assisted consent systems. At every step, careful attention to the foundational principles of informed consent provides a clear framework for continued development and evaluation.