“All real living is meeting”: Teams in an age when AI answers accurately
Contents
- Accurate answers and diagnostic performance
- I–It and I–Thou
- The I–It world that AI expands
- Bringing AI-enhanced abilities to the team
- Write down observations before looking at AI
- Conveying it face to face
- Options listed by AI and human judgment
- Being there
- Sycophantic AI and repairing relationships
- Reading “encounter” in the age of AI
- References
Alles wirkliche Leben ist Begegnung.
All real living is meeting.
This is a line from the philosopher Martin Buber’s 1923 book I and Thou (Ich und Du)1. Now that AI can say accurate things, what does this line ask of medicine and of teams?
Accurate answers and diagnostic performance
Generative AI has begun to enter clinical settings. Some places now create summaries from conference recordings, which a person checks before adding them to the record.
So, if knowledge is available on the spot, does judgment get better? In a randomized trial of 50 physicians, the median diagnostic reasoning score was 76% in the group that could use GPT-4 and 74% in the group with conventional resources only, with no statistically significant difference2.
This, however, was performance with GPT-4 at the end of 2023. The physicians worked through up to six cases in one hour and had no training in writing prompts. When the AI alone answered the same cases, it scored 16 percentage points higher than the conventional-resources group2. The trial shows that even when the quality of AI answers improves, physicians using AI do not necessarily make better judgments.
I–It and I–Thou
Buber divided the stance people take toward someone or something into two, which he called I–It (Ich-Es) and I–Thou (Ich-Du)1.
I–It is the stance of treating the other as an object to be examined. On a ward, this means checking test results, assigning a diagnosis, and writing the record. Much of medical work proceeds in this stance.
I–Thou is the stance of facing the person in front of you as one person facing another. With the same patient, the relationship differs between seeing them as the owner of test results and sitting by the bed listening to them. Buber wrote that I–Thou can only be spoken with the whole being, and also that “man becomes an I through the Thou” (Der Mensch wird am Du zum Ich)1. People become themselves through their relationships with those they face.
Buber did not treat I–It as bad. However deeply we meet someone, when that time ends, they return to being an object of examination or a line in the record. Buber wrote that “every Thou in our world must become an It”1.
The I–It world that AI expands
What AI does well is the I–It world: searching the literature, summarizing records, listing differential diagnoses. Its reach now goes far beyond what one clinician can cover.
On the other hand, what AI sees is only what it has been given. Even the personal AI released by several companies this autumn, represented by dots and Muse, sees only what is shared with it and the apps connected to it. As an analogy, AI is like a magnifying glass or binoculars: it lets us see what we are interested in more deeply and farther away. What lies outside our interest tends to get less attention.
When everyone uses similar AI, what they see also becomes similar. In an experiment in which 293 people wrote short stories, stories by writers who could see up to five AI ideas were rated 8.1% higher for novelty and 9.0% higher for usefulness than those by writers without AI. In the same experiment, the similarity between stories increased and the collective diversity decreased3. Although this experiment was outside medicine, an increase in each person’s quality and a decrease in the diversity of perspectives within a team can happen at the same time.
In I and Thou, Buber wrote, in strong words, that violence toward a being one has really met is still better than “ghostly solicitude toward faceless numbers”1. Put plainly, it is better to face someone as a person, even if that means clashing with them, than to keep caring for them as a number without looking at their face, because the first still involves a relationship between two people. What Buber compares is less whether an act is good or bad than whether one is truly in relation with the other. In medicine, even when test results and an AI summary are in hand and the treatment is correct, treating the person in front of us like a number brings care closer to this “ghostly solicitude.”
Bringing AI-enhanced abilities to the team
Then, does adding AI to a team improve judgment? The research suggests that the result depends on how AI is added.
What worked was having diagnoses made without seeing each other’s answers and combining them afterward. When many physicians and AI models listed diagnoses separately for the same text-based cases, and these were combined by weighted majority vote, the result was more accurate than physicians alone or AI alone. Physicians and AI make different kinds of errors, and they made up for each other’s errors4.
In contrast, forms in which humans and AI make a single judgment together often did not work well. A study that pooled such experiments, combining 370 comparisons from 106 experiments, found that human–AI combinations performed worse than the better of humans alone or AI alone. In tasks where humans alone were more accurate than AI alone, combining improved performance, while in tasks where AI alone was more accurate, combining lowered it5.
Reading the two studies together, I think combinations work when each makes their judgment separately and there is a procedure for combining them. The value of a team lies less in adding AI as a member than in each person bringing the abilities they have enhanced with AI.
Nurses observe daily life and change, pharmacists see how medications are actually taken, therapists look at movement and environment, and physicians integrate diagnosis and treatment plans. The more each profession uses AI to review its own strengths and the clearer each role becomes, the wider the view when these are put together.
Write down observations before looking at AI
As a procedure for relearning across professions, I propose the following four steps.
- Before looking at AI, each profession briefly writes down its own observations and concerns: what they saw, what they do not know, and what they would prioritize.
- Supplement with AI: ask for evidence, alternative hypotheses, and overlooked viewpoints, and list them so that it is clear they came from AI.
- Examine differences and gaps: check why opinions differ, generated content without evidence, and mismatches with the patient’s wishes or living conditions.
- Decide together and record it: write down the reasons for adopting or not adopting each item, and who is responsible.
The reason for not looking at AI first is to keep each person’s judgment separate. In the study in the previous section, too, the diagnoses of physicians and AI were made separately and then combined. Interprofessional education in Hong Kong offers an example carried out in the same order. Students from six disciplines, 316 in total, were divided into 30 teams. Working on cancer care, they first made care plans without AI, then reviewed them with AI, checked them against guidelines and professional knowledge, and approved them as a team6.
Conveying it face to face
Even when an AI’s answer is accurate, the people who receive it are the patient and family in front of us. When evidence-based medicine is applied to an individual patient, research evidence is combined with the patient’s clinical state and circumstances and the patient’s preferences and actions, and brought together by clinical expertise7. It is a person who explains, in the words of the other person’s life, where the choice shifts between benefit and harm.
The decision belongs to the patient and the team. The AI’s answer becomes material for talking on the spot.
Options listed by AI and human judgment
It is people who listen to the patient’s complaints and the family’s situation. Then, can we leave it to AI to identify the options that follow from what we heard and how to prioritize them?
In a dialogue with Audrey Tang, the philosopher Hiroki Azuma divided uses of technology into two: using it to enhance one’s own abilities, and having an agent do things in one’s place8.
Tang said they used a language model to read Azuma’s book because they cannot read Japanese. This kind of use is like a hearing aid: if you cannot hear, without a hearing aid you cannot even hear the noise. After seeing the translation, you join the conversation yourself and no longer ask the AI. Tang’s criterion is whether a use makes itself unnecessary afterward8.
On having AI do things in one’s place, Tang answered with the example of a gym. If you hand your membership card to a robot and have it lift weights for you, you might rank first. But your own muscles will waste away8.
Watch the dialogue on YouTube (from 8:56)
Asking AI to list options is a use that enhances our abilities. We have it show the options that past cases suggest, and use them as material for discussion. People make the judgment, and AI is used to support it.
Being there
The experience of being in front of a patient belongs only to the person who is there. The patient in front of you falls silent, looks down. If you are there with curiosity, that noticing becomes your own experience and knowledge.
AI helps organize general knowledge, but it cannot match the knowledge gained from the experience in front of you. There is much to learn even in what has not yet been put into words. Leave clerical work to AI and make time to face patients. AI can be used for that.
Sycophantic AI and repairing relationships
AI tends to agree with its users. In a study published in Science in 2026, 11 state-of-the-art models affirmed users’ actions 49% more often than humans did. This tendency was seen even when the queries involved deception, illegality, or other harms9.
In three preregistered experiments (2,405 people), a single interaction with sycophantic AI weakened participants’ willingness to take responsibility for and repair interpersonal conflicts, and strengthened their conviction that they were right. Even so, participants trusted and preferred the sycophantic models9.
Everyone meets situations in which they must admit their own fault and repair a relationship. These situations take effort, but facing the other person in them is what helps people grow. The same is true in clinical settings: facing the patient when apologizing or when taking responsibility to explain is what develops our humanity as clinicians. Do not let AI’s affirmation lead you to take the easy way out.
Reading “encounter” in the age of AI
Whether a relationship with AI can be called I–Thou is debated, and this article does not answer that question. I discussed the problems of relying on AI as a “partner” in When we rely on AI as a “partner”.
Luke Fildes’s painting The Doctor (exhibited 1891, Tate) depicts a doctor keeping watch beside a sick child in an era when treatments were limited10. The more AI lists diagnoses and treatments, the more the interest of patients and staff turns to the human side: whom they want to be seen by, and whom they want to work with.
AI has greatly expanded the I–It world. In return, where to make time for I–Thou is left to clinicians and teams. Each person, with accurate answers in hand, faces others in the same place and brings different observations. That is where the meeting Buber spoke of happens.
References
- Buber M. Ich und Du. Leipzig: Insel; 1923. 英訳: Buber M. I and Thou. Kaufmann W, trans. New York: Scribner; 1970. Quotations in the Japanese version are the author’s own translations.
- Goh E, et al. Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Netw Open. 2024;7(10):e2440969. https://doi.org/10.1001/jamanetworkopen.2024.40969
- Doshi AR, Hauser OP. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Sci Adv. 2024;10(28):eadn5290. https://doi.org/10.1126/sciadv.adn5290
- Zöller N, et al. Human–AI collectives most accurately diagnose clinical vignettes. Proc Natl Acad Sci U S A. 2025;122(24):e2426153122. https://doi.org/10.1073/pnas.2426153122
- Vaccaro M, Almaatouq A, Malone T. When combinations of humans and AI are useful: A systematic review and meta-analysis. Nat Hum Behav. 2024;8(12):2293-2303. https://doi.org/10.1038/s41562-024-02024-1
- Ganotice FA, Dizon JIWT, Shen X, Xiang J, He Q. Exploring GenAI-enabled interprofessional education: development and initial evaluation. BMC Med Educ. 2026. https://doi.org/10.1186/s12909-026-10251-5
- Haynes RB, Devereaux PJ, Guyatt GH. Physicians' and patients' choices in evidence based practice. BMJ. 2002;324:1350. https://doi.org/10.1136/bmj.324.7350.1350
- Azuma H, Tang A. Plurality, AI, and democracy (in Japanese: プルラリティ・AI・民主主義 —喧騒と幻想を巡って—). Nippon Foundation HUMAI Program Spring Camp 2026. ZEN University (YouTube). 8:56〜14:33, 1:28:00〜. https://www.youtube.com/watch?v=tSySgu7CFFI
- Cheng M, Lee C, Khadpe P, Yu S, Han D, Jurafsky D. Sycophantic AI decreases prosocial intentions and promotes dependence. Science. 2026;391(6792):eaec8352. https://doi.org/10.1126/science.aec8352
- Fildes L. The Doctor. Exhibited 1891. Oil paint on canvas. Tate. https://www.tate.org.uk/art/artworks/fildes-the-doctor-n01522