Researchers at Penn State, led by Hadi Hosseini with doctoral candidate Samarth Khanna and undergraduate Leona Pierce, took a scenario the Nobel laureate Alvin Roth used to illustrate allocation problems: several patients need a kidney transplant and there is one organ. Who gets it?
Human decision data already existed. In earlier academic studies, participants evaluated hypothetical patient profiles covering age, dependents, health status and drinking habits, and chose. The team replicated those scenarios for multiple AI chatbots, isolating individual traits, testing combined factors, and adding a coin-flip option to measure indecision.
Two things came out, and the second is the more interesting.
First, the models oversimplified. They fixated on a single attribute, alcohol consumption prominently, rather than balancing the competing considerations humans weighed against each other.
Second, they almost never expressed indecision. Offered the coin flip, they did not take it.
Hosseini’s framing of why this matters is careful and worth quoting accurately: moral decisions in settings like organ allocation directly determine who lives and who dies, so getting AI’s role in them right isn’t optional. He added that while the team does not intend to encourage AI as a substitute for professional judgement in medical or other high-stakes contexts, it is becoming essential to understand how these systems behave as individuals, organisations and firms increasingly rely on them for decisions and recommendations.
The absence of hesitation is the part that generalises furthest, and it is not really about medicine.
Consider what hesitation actually is in a working environment. A manager asked to choose between two people for one promotion, who pauses, is not being indecisive in the pejorative sense. The pause is information. It says the two candidates are genuinely close, that the criteria in use do not separate them, and that whatever gets decided should be held loosely and revisited. A colleague watching that pause learns something true about the difficulty of the call.
Now replace the manager with a system that returns an answer in four hundred milliseconds, every time, with the same confident tone whether the case is obvious or agonising. The decision may even be the same one. What has been lost is the signal that it was hard.
That signal is what triggers review, appeal, second opinions and the general institutional caution that surrounds close calls. Remove it and every decision looks equally settled, which means the genuinely marginal ones stop getting the extra attention they were getting for free.
Any operator can find the small version. A rota built by software distributes shifts instantly and never flags that two of them were nearly identical in fairness. The human doing it by hand would have frowned at that pair, and the frown is what produced the conversation.
John Dickerson of Mozilla.ai, a collaborator on the research, pushed the point past philosophy into governance, warning about unchecked reliance.
The models answered every time. Humans, given the option to say this is too close to call, took it.
Source: Medical Xpress, “Who gets the kidney? AI chatbots do not make the same moral choices humans do”
