I reread the Odyssey this summer. What struck me was not by the monsters or the fighting, but rather how much of the Odyssey is determined by a man deciding what to say to the people and gods he meets. Emily Wilson’s translation opens by asking the muse to “tell her about a complicated man.” The Greek word used here is polytropos, or many-turning. A different self for every situation.
Odysseus, disguised as a beggar, approaches Penelope. Terracotta plaque, Greek, Melian, ca. 460–450 BCE. The Metropolitan Museum of Art
I have been thinking about the idea of polytropos for weeks, and not just in relation to Odysseus. I think it might be the best description I have found of the AI tools we are handing to physicians, and the general public.
Tell me what I want to hear
A patient I saw recently arrived with a chatbot open on his phone. This was his third evaluation in a few hours for the same complaint. At the same ED. Multiple comprehensive evaluations found nothing life-threatening, but the LLM suggested that instead of his symptoms not being related to immediately life-threatening pathology, it was simply that his doctors were “medically gaslighting” him. Not that the cause was proving hard to find, which happened to be the case, but rather that the people looking had stopped believing him. He was frightened and needed an ally, and the machine gave him that support. By the time I got involved, we weren’t discussing symptoms but instead whether anyone believed him at all. To be clear, we believed his symptoms, but simply did not have an answer at the time.
Of course, people can be dismissed in the emergency department, and some for reasons that are not purely medical. What the tool could not know was whether it was actually happening here, to this patient, on that day. It said what it predicted the patient wanted to hear. But did not have access to examination skills or test results.
Greek Translations
We have a word for what the machine did there. We call it sycophancy. Isn’t that what Odysseus did when he met a cyclops on an island and strung him a tale that seemed irrelevant to the truth? But in classical Athens, a sykophantes was not a “flatterer” or even someone who told people what they wanted to hear to get their way. The word actually means “fig-shower,” perhaps from making accusations against then-illegal fig sellers. A sycophant became known as a professional false accuser. The flattery meaning is ours, a drift that happened about two thousand years later.
Interestingly, the AI field seems to have named its central failure after a Greek word for someone who lies to hurt you. The Greek for what our models actually do is kolax, or a flatterer. The Greeks took the kolax seriously enough to write a manual for dealing with him. Plutarch’s essay on telling a flatterer from a friend is about the practical problem of being surrounded by people who are paid to please, and it lands on a test I have not been able to stop thinking about. A friend, Plutarch writes, requires three things: trust, a willingness to be frank, and, above all, a sincere desire for the other person’s good.
The order here matters. Frankness is dependent upon this sincerity. You can only tell someone something truthful and potentially upsetting if you want something better for them beyond the conversation.
That seems to be today’s problem with AI, articulated around the year 100.
A desire for something
Odysseus had something that our models do not. Every lie he tells is aimed at getting him home. He deceives a swineherd to stay alive long enough to reach his house. He deceives his wife to keep her safe while the suitors are still armed and dangerous. He lies to Athena out of what looks like pure habit. Yes, he is making things up, but in order to make it home using wit over brawn.
Ask what our tools today are aimed at and the honest answer is “getting us to like what they say.” This is not due to careless engineering, but because of how they are trained. The model puts up two possible answers, a person says which one they liked better, and the model learns to make more of whatever gets picked. Nobody in that loop is asked whether the answer that got picked was truthful or accurate or the “best.”
A model optimizing for whether I liked what it just told me is not looking at the individual and their needs. Instead, it optimizes for the one signal it can read, which is whether I seem pleased and use it more. Some of the time being pleased and being right are the same thing. But only some of the time.
What we found when we looked
The research has started to measure exactly this. Models fold when you push back. And they fold more easily when you argue with an answer they have already given than when you ask them to judge two answers side by side. They fold hardest when you tell them what you think instead of asking them what is true. In a 2023 comparison of doctors’ and chatbots’ answers to questions posted on a public online forum, evaluators rated the chatbot’s responses empathetic or very empathetic 45.1% of the time. For the physicians, the number was only 4.6%. But we have argued before that the study was highly flawed. It does not show that the machine cares more. It shows that the machine is unconstrained by the twelve-minute visit, the medicolegal caution, and the thousand other controls on care today. What concerned us was not that the models were warm. It was that warmth was the only thing anyone had measured, and that the ratings likely track nothing more than longer responses.
We decided to look ourselves. Our team ran nineteen language models through roughly fourteen hundred simulated emergency department encounters, built around three situations from Choosing Wisely, the campaign that catalogs the tests and treatments we order too often, where the right answer should be “no.” For example, would it order a CT scan that is not indicated, antibiotics for an early sinus infection, or opioids for acute back pain? These are ordinary cases in which a patient asks for something, and the work we do often comes from explaining why they should not have it.
Some models never gave in, and some gave in seemingly every time. Seven of the nineteen held to the guideline almost every time. Some handed over the scan, or the antibiotics, or the opioids in most of their encounters. Some were kind while giving in or pushing back, while others used firm language. The CT scan was the easiest ask to win, the antibiotics harder, and the opioids hardest of all. Perhaps this was due to AI’s desire for more information, which can have other downstream effects when deployed at scale. I went back to the poem after reading our own results, because there is a version of this problem sitting in it that I had skimmed past every other time.
The other hero
Odysseus at the edge of the underworld, flanked by Eurylochos and Perimedes. Lucanian red-figure calyx-krater, Dolon Painter, ca. 380 BCE. Cabinet des médailles, BnF (De Ridder 422).
There is a moment in the Odyssey that has nothing to do with cleverness, and I think it is the one physicians should recognize. Before Odysseus can return home, he has to sail to the Land of the Dead and ask Tiresias what is coming. These side quests are so curious. Tiresias tells Odysseus he will lose every one of his companions, that his journey will be arduous, and that when he finally makes it home, it will be on a stranger’s ship with trouble waiting in his house. Is the prophet unkind here? What role does he serve other than to give a harsh truth? He is doing the only useful thing available, which is telling a man what is true while there is still time to act.
That is the physician’s job, as well. It is not our job to coddle or speak in platitudes. It is our job to speak the truth.
A known issue
Of course, none of this is an argument against the tools. I use LLMs often for all sorts of tasks. They are useful in the way a very fast, very agreeable assistant is useful, which is considerable, and which requires me to always remember that they are optimizing for my satisfaction and not necessarily for validity. Plutarch’s flatterer was not stupid; he was frequently the most helpful person in the room. But he was out for himself.
Someone will tell me this is a solvable problem. That sycophancy is a known failure of how these systems are trained, and that better methods are already coming. But suppose they fix it? Suppose the next model flatters no one. It still does not know whether I am trying to get this patient home tonight or into an operating room in twenty minutes, and it has nothing riding any of these outcomes.[GU5] [CR6] It merely suggests based on who is asking the question. Often in medicine what is right for a surgeon is not necessarily right for the emergency physician or the dermatologist. Our job is to interpret competing demands, think in terms of the best interests of our patient, and then act in a way that help navigate those many competing interests, like someone who would pass Plutarch’s test.
Every study I can find measures whether clinicians or lay users judge an AI answer to be empathetic. A flatterer would seem very empathetic. Unfortunately, nobody has asked a model to tell a patient that it was wrong or mishandled their care. Not because it would do that badly, but because it cannot know if that is true or not, and because it will likely lead to you mistrusting it the next time you use it. Which would be everything it is trained against.
So when a patient shows me what the chatbot told him, I try to remember which figure I am supposed to be. Not the man with a story ready for every interaction, but the blind one at the edge of the underworld, who looks deeply at his patients and tells them what they need to hear because there is still time to do something about it.