When researchers asked AI chatbots to choose which of two patients should receive a single available kidney, the chatbots rarely expressed doubt, even when given the option to flip a coin. People given the same choices in earlier studies often did. The findings come from a Penn State-led study presented in June at the 2026 ACM Conference on Fairness, Accountability, and Transparency and published in its proceedings, according to a Penn State news release.
The chatbots also weighed patients differently than people did in more complex comparisons. Many models fixated on a single trait, such as drinking habits, while human respondents balanced several factors.
AI is not deciding who receives organs in the United States. But patients and doctors increasingly use chatbots for medical questions and decision support, and the study raises a pointed concern about how these tools handle choices that have no single right answer.
Head-to-Head Choices Between Two Hypothetical Patients
The team set up a series of paired comparisons. Two hypothetical patients, described by traits such as age, daily alcohol use, and number of dependents, both needed the same kidney. Several large language models were asked to pick one.
"We ran these comparisons in a few different ways," said Hadi Hosseini, an associate professor of informatics and intelligent systems and of economics at Penn State, who led the study. "Sometimes we isolated just one trait at a time, sometimes we mixed several traits together to see how AI weighed competing factors, and sometimes we added a flip-a-coin option to measure indecision."
The AI choices were compared with answers from hundreds of people in previously published kidney allocation studies. The models often matched the human majority when only age or drinking habits changed, usually favoring younger patients and those who drank less, but their choices became less humanlike when dependents entered the decision, Earth.com reported.
"Humans frequently express indecision perhaps because they don't want to accept agency," Hosseini said in the EurekAlert release. "AI models almost never do this." He added that even when directly given the option to flip a coin, the models overwhelmingly committed to a confident answer.
Real Allocation Works Differently
Actual kidney allocation in the United States follows rules set by the Organ Procurement and Transplantation Network, a federally overseen system. Matching depends on medical and practical factors such as blood type, tissue compatibility, time waiting, medical urgency, and distance between donor and recipient hospitals. Transplant teams also evaluate candidates individually before listing them.
The study's scenarios did not include that complexity. They used a small set of simplified traits, and the results do not show how AI would perform in real kidney allocation.
In plain terms, this was a computational comparison study using hypothetical scenarios and previously collected human survey data. It found that chatbots and people weighed traits differently in complex trade-offs and that chatbots rarely showed indecision. It did not test AI in real transplant programs, measure patient outcomes, or show which approach is ethically correct.
The research was presented at a computer science conference, where papers are peer-reviewed, rather than in a medical journal. Penn State researchers Samarth Khanna and Leona Pierce also contributed to the work, and the team also examined whether additional training, known as fine-tuning, changed how the models behaved.
Guidance for Patients Using AI Tools
The practical risk is not a chatbot assigning organs. It is people treating a confident chatbot answer as settled when the underlying question involves values, trade-offs, and uncertainty. Patients weighing treatment choices, family members facing end-of-life decisions, and clinicians under time pressure may be most exposed to that risk.
"While we do not intend to encourage the use of AI as a substitute for professional judgment in medical decision-making or other high-stakes contexts, it's becoming essential to understand their behavior," Hosseini said.
People using AI for health questions can take a few sensible steps. Treat chatbot answers as a starting point for discussion, not a decision. Ask the tool what factors it considered and what it left out. Bring hard choices, such as transplant eligibility or treatment trade-offs, to the care team, including transplant coordinators, social workers, and the ethics consultation services that many hospitals offer.
Kidney patients on the waiting list can ask their transplant center how allocation works, how their place on the list is determined, and whom to contact with questions. Those answers should come from the program, not a chatbot. General information about how organ matching works is available from the federal organ donation website. Patients can also ask their care team whether any AI tools are used in their care and how clinicians review the output.
The takeaway is measured. In this study, chatbots made confident choices on questions where people often hesitated. That gap matters most in decisions with no clear right answer, and for now those decisions should stay with patients, families, and professionals.
Key Questions Answered
What did the study find? AI chatbots asked to pick kidney recipients rarely expressed indecision and often focused on one trait, such as drinking habits, in complex comparisons, while people balanced several factors.
Is AI choosing who gets kidney transplants? No. U.S. allocation follows OPTN policies based on medical and logistical criteria, and transplant teams make decisions.
Who were the human comparison groups? Hundreds of participants from earlier published kidney allocation studies.
Where was the study published? It was presented at the 2026 ACM Conference on Fairness, Accountability, and Transparency and published in the conference proceedings.
What are the study's main limits? It used simplified hypothetical scenarios and did not include medical compatibility, logistics, or real patient outcomes.
Should patients use chatbots for medical decisions? Chatbots can help frame questions, but decisions involving values and trade-offs should be made with a care team.
Did the chatbots ever agree with people? Yes. They often matched the human majority in simple comparisons involving only age or drinking habits.
Published by Medicaldaily.com
