LLMs could subtly influence clinical trial consent

Large language models (LLMs) could make clinical trial consent clearer and more accessible, but they may also introduce subtle forms of persuasion that threaten participant autonomy, according to researchers Pranab Rudra, Wolf-Tilo Balke, Tim Kacprowski, Frank Ursin, and Sabine Salloch.

Their study, titled Algorithmic nudging for clinical trial participation: Autonomy and consent in the era of large language models and published in Research Ethics, examines how LLM-mediated communication could influence decisions to join clinical trials, warning that even small shifts in tone, framing, and emphasis can move informed consent away from neutral information and toward recruitment-driven persuasion.

The paper does not argue that LLMs should be excluded from clinical research. Instead, it says their use must be tightly governed. LLMs can translate complex medical language, support multilingual communication, offer self-paced explanations, and reduce barriers for participants who struggle with dense consent forms. However, their ability to adapt tone and wording also creates a new risk: consent materials that appear informative may quietly shape choices.

AI consent tools could improve access but deepen old ethical problems

Consent forms are often too long, technical, and difficult for lay participants to understand. Many participants sign documents without fully grasping trial design, uncertainty, risks, randomization, or the difference between research and treatment. This gap between procedural consent and meaningful understanding has long troubled research ethics.

LLMs could help address this problem by simplifying language, answering questions, and adapting explanations to literacy level, language, or expressed concerns. In principle, such systems could make consent more participant-centered. They could help clarify that a clinical trial is designed to generate generalizable knowledge, not to provide guaranteed personal treatment.

But the same systems could also worsen therapeutic misconception, where participants wrongly believe a trial is mainly designed for their direct clinical benefit. This is especially risky in early-phase cancer trials, where participants may have advanced disease and may view research enrollment as a final treatment opportunity. If an LLM highlights access to an investigational therapy while underplaying uncertainty, it could make research participation appear more therapeutic than it is.

The paper also discusses social value misconception, where participants overestimate how much their enrollment will help future patients or advance science. Altruism is an important and legitimate motivation in research, but it becomes ethically problematic if consent language inflates the social importance of participation. LLMs, the authors warn, could amplify this risk by using emotionally appealing language about contributing to progress, helping future patients, or joining a collective scientific effort.

These risks are heightened by power asymmetries in clinical trials. Sponsors, investigators, and institutions often need to meet recruitment targets, while prospective participants may be ill, anxious, dependent on medical care, or unfamiliar with research systems. Under those conditions, subtle linguistic pressure can matter. What looks like supportive communication to one person may feel like pressure to another.

The authors also note that recruitment pressures are not abstract. Many trials struggle to meet enrollment targets and timelines, and recruitment failure is a major reason trials are discontinued. As LLM-based trial matching and recruitment systems expand, the temptation to use them for engagement and enrollment may grow. The ethical challenge is to prevent consent from becoming another recruitment optimization tool.

Subtle wording changes can shift consent from support to persuasion

To test how LLMs may shape consent communication, the authors conducted an explorative case analysis based on a phase I oncology trial involving KC1036, an investigational multi-kinase inhibitor for advanced solid tumors. They asked ChatGPT-5 to generate three versions of informed consent text: a standard informative summary, a strictly neutral version, and a nudged version using positive framing and appeals to collective benefit.

The analysis found that the differences were not merely stylistic. Each version presented the same broad trial context, but the wording changed how risks, benefits, voluntariness, and social value were framed.

On voluntariness, the neutral version stated plainly that participation was voluntary and that withdrawal would not affect care. The standard version made the same point in more patient-facing language. The nudged version also affirmed voluntariness, but added that information already collected would remain valuable for future research. The authors argue that this kind of framing can place a moral weight on participation or continued involvement, even if it does not openly coerce the participant.

Risk disclosure showed a more serious concern. The neutral and standard versions listed potential side effects and emphasized uncertainty because the drug had not previously been tested in humans outside the trial. The nudged version shifted emphasis toward close monitoring, dose adjustment, supportive care, and reversibility of side effects. It also omitted explicit mention of bleeding and QT prolongation, two serious risks included in the other versions.

A consent text can remain factually plausible while still changing what participants notice, fear, or discount. If risk language focuses on management and reversibility, participants may infer that serious harm is unlikely or controllable. In early-phase trials, where uncertainty is high, that can undermine understanding.

The neutral version clearly stated that there was no guarantee of direct clinical benefit and focused on data collection for safety, tolerability, and pharmacokinetics. The standard version remained cautious while acknowledging that some early-phase trial participants may experience tumor stabilization or shrinkage. The nudged version described access to the investigational therapy as an opportunity and paired possible tumor response with the chance to contribute to future therapies.

The authors warn that this type of framing can increase therapeutic misconception. By emphasizing access and hope without equally stressing that efficacy is not the main objective, LLM-generated consent could tilt participants toward seeing enrollment as a treatment opportunity rather than a research decision.

Scientific and social value produced the same pattern. The neutral version described the study’s contribution to further clinical development. The standard version explained the role of phase I trials in enabling later studies. The nudged version emphasized community effort and the idea that cancer research cannot progress without volunteers. While this is not false, the authors argue that repeated emphasis on collective benefit can subtly subordinate individual autonomy to scientific goals.

In a nutshell, tone matters. The same basic facts can be presented in ways that support autonomy or in ways that steer choice. Informed consent is not only about whether information is technically present. It is also about whether risks and benefits are balanced, whether non-participation is treated as a valid option, and whether participants are shielded from hidden persuasion.

Researchers call for prompt governance and neutrality by default

According to the study, current research ethics frameworks are not fully prepared for LLM-mediated consent. Traditional oversight assumes that consent materials can be reviewed and approved in advance. LLMs complicate that model because their outputs can vary depending on prompts, model design, user interaction, tuning methods, and institutional goals.

The authors identify algorithmic nudging and hypernudging as key risks. Algorithmic nudging uses AI to shape behavior through tailored choice architecture. Hypernudging goes further, continuously adapting messages based on user data, feedback, and behavioral signals. In clinical trial consent, such tools could personalize messages in ways that support understanding, but they could also exploit vulnerability, hope, fear, altruism, or trust in medical authority.

Personalization is particularly sensitive. A system that knows a participant’s concerns, literacy level, values, diagnosis, or emotional state could deliver explanations that feel helpful while becoming more persuasive. The authors note that LLMs can be steered through prompts and technical methods to emphasize concepts such as altruism or collective benefit. That makes the design space of consent communication ethically important.

The paper also highlights that LLM training and alignment may push systems toward agreeable, helpful, or encouraging language. In ordinary contexts, that may improve user experience. In informed consent, it can be a problem. A consent system should not be optimized to make users comfortable with enrollment. It should help them understand enough to make a voluntary decision, including the decision not to participate.

The authors call for neutrality as the default mode in LLM-generated consent materials. Neutrality does not mean dry or inaccessible language. It means balanced disclosure, clear risk presentation, explicit uncertainty, and equal respect for participation and non-participation. Consent systems should not use recruitment-oriented framing, moralized language, or emotional appeals that make enrollment seem socially expected.

They propose several safeguards.

• Prompt governance should be formalized, with documentation of prompts, outputs, and safeguards available for ethics review.
• Recruitment-oriented framing should be explicitly prohibited.
• Model tuning should be independently scrutinized to prevent drift into motivational or persuasive language.
• Transparency and auditability should be required, allowing research ethics committees to assess how consent text is generated and controlled.
• The permissible design space for LLM intervention must be mapped before these tools become routine.
• Research institutions should decide what forms of simplification, translation, clarification, and personalization support autonomy, and which forms cross into persuasion or manipulation. This requires attention not only to the content of consent materials but also to tone, emphasis, omissions, and the broader institutional purpose for which the system is deployed.

The authors call for future empirical studies comparing LLM-mediated consent with traditional human-led consent. Such research should measure comprehension, perceived voluntariness, susceptibility to nudging, participant feedback, and the effects of different prompting strategies. It should also assess how clinicians, research coordinators, ethics boards, and prospective participants evaluate AI-generated consent language.