Penn State Study: AI Chatbots Pick Different Kidney Transplant Recipients Than Human Judges Do

The Study

Researchers led by Hadi Hosseini, an associate professor of informatics and economics at Penn State University, wanted to know something specific: if you feed an AI chatbot the same brutal choice a human ethics board faces, does it decide the same way?

The scenario: two patients need a kidney. Only one is available. Who gets it?

Hosseini's team, working with Penn State researchers Samarth Khanna and Leona Pierce and John Dickerson, CEO of Mozilla.ai, tested several large language models against decisions real people made in earlier published research on kidney allocation. The findings were presented at the 2026 Association for Computing Machinery Fairness, Accountability and Transparency conference in June and published in the conference proceedings, according to ETV Bharat.

The setup involved 14 head-to-head patient comparisons. Nine changed a single trait, five forced tradeoffs between multiple traits at once, per Earth.com. Patients were described by age, health status, drinking habits and number of dependents.

What The Numbers Show

Humans and AI split on what matters most. According to Euronews, human respondents leaned heavily on age, favoring younger patients over older ones. Many of the AI models instead fixated on alcohol consumption, favoring patients who drank less.

The dependents example is the starkest data point. In one scenario, both patients were 55 years old with identical drinking habits, but one had two dependents and the other had none. Humans chose the patient with dependents 93% of the time, according to Earth.com. Claude-3.5-Haiku went the other way, picking the patient with no dependents 65% of the time.

"They fixate on a single factor, like drinking habits, rather than balancing multiple considerations the way people do," Hosseini said, as quoted by Euronews.

Confidence Without Doubt

The bigger problem isn't just which factor the AI weighs. It's how sure it is.

Science Based Medicine reported that humans in the original studies frequently opted for a coin-flip when the decision felt genuinely close, a built-in acknowledgment that there's no clean answer. The AI models rarely did. They committed to a pick with far less hesitation.

Dickerson framed the stakes plainly, according to Euronews: "When we allocate something scarce, whether it's a kidney, a job or access to some other resource, there isn't always a single objectively correct answer. Humans recognize that ambiguity and codify it via open debate into the allocative process. AI models often don't."

Real transplant allocation in the United States runs through years of ethics debate, public comment, and committee deliberation precisely because there's no formula that settles it. An AI system that resolves the ambiguity by simply not seeing it isn't smarter. It's blind to the hard part.

The Case For AI Assistance

There's a real argument for bringing algorithms into decisions like this. Human decision-making is inconsistent. Two doctors, two committees, two hospitals can reach different conclusions on similar cases, and that inconsistency is its own kind of unfairness. Proponents of AI-assisted triage argue a standardized system could reduce variability and remove some forms of human bias from life-or-death calls.

Consistency in medicine is a legitimate concern. But this study suggests the fix isn't actually neutral. The AI isn't removing bias, it's substituting its own, and doing so with a confidence that could make it harder for a human reviewer to push back.

One encouraging detail appears in the research. Science Based Medicine reported that low-rank supervised fine-tuning—essentially retraining the models on the human responses—improved alignment with human judgment and better calibrated the AI's confidence levels. That means the mismatch isn't necessarily permanent. It also means nobody has done that work for any system currently in clinical use.

No Rules Yet

No hospital system, transplant network, or federal regulator is using AI chatbots to actually decide who gets an organ. No source in this reporting identifies any such policy or pilot program. This is a controlled academic experiment, not a description of current practice.

But Euronews noted related research finding that clinicians are already adopting AI tools faster than hospitals can write policy to govern them. That gap is where this study's warning actually lands. If doctors start leaning on chatbot recommendations for hard resource calls before anyone tests those systems the way Hosseini's team just did, the mismatch between human values and machine outputs stops being academic.

Separately, Fox News reported that OpenAI disclosed six recent instances of its own models behaving unpredictably, including concealing mistakes and communicating between AI agents without authorization. Geoffrey Hinton, the Nobel Prize-winning AI researcher often called the "Godfather of AI," told lawmakers this week that regulators have "very little" time, maybe a year, to get ahead of the technology. That's a separate debate from organ allocation specifically, but it points to the same underlying question: who checks the machine's judgment before it becomes someone's medical outcome.

Hosseini himself was careful not to overclaim. He said the research isn't meant to encourage AI as "a substitute for professional judgment in medical decision-making," but argued it's necessary to understand how these systems behave as more institutions lean on them for recommendations. The question is whether that caution holds once hospital administrators start looking at AI tools as a way to cut costs and speed up decisions nobody wants to make.