Man and Machine—May the Lord Save: VI. On Empathy — Halakhic Rulings from an Experiential Distance (Column 699)

I’ve just finished a series of columns that dealt with AI and with the relationship between a human being and his thinking and the “thinking” of an LLM (Large Language Model). In a comment to the last column, Michael posted a beautiful clip from the wonderful film Good Will Hunting, in which the psychologist describes to Will the relationship between his vast knowledge and his lack of understanding regarding that knowledge (just as we saw in recent columns with the Chinese Room example and regarding LLM machines):

I wrote to him that this clip aptly describes my point about AI machines, for as I explained throughout the series, they may “think” but they do not understand (in fact, they don’t even think; they merely channel electrical currents, while thinking, meaning, and understanding all reside solely in the user). This video essentially illustrates the distinction that Wittgenstein discussed (and apparently denied; see Column 694, the first in the series) between understanding/meaning and use. That distinction was demonstrated, among other things, by the Chinese Room, and can also be seen in Mary’s Room (see, for example, Column 591) and in Russell’s example of “the yellow color.” In my reply to Michael I also mentioned my essay on halakhic rulings in extreme situations, which I’ve referenced more than once (see, for example, Columns 446, 452, 686, and more). For some time I’ve been considering dedicating a column to rulings in extreme conditions—situations far removed from the decisor’s world—and I now see that this connects to this series and to the discussion of an AI posek. So I decided to add this column as (apparently) the concluding one in the series.

An AI Posek

In Column 590 I discussed whether AI software can issue halakhic rulings. Since then, my friend Prof. Moshe Koppel sent me a link to the Rabbi Dicta model, which includes an automatic posek trained by him (at the company Dicta that he founded). The results are quite impressive (and all in fine rabbinic Aramaic, just like in his pashkevilim [wall posters]). One concern raised there (in the name of Rabbi Cherlow) was that such a model would rule without sensitivity to the person standing before it, relying only on sources—i.e., a dry ruling (not necessarily “second-order”; see below). I wish here to sharpen that concern, and in the end qualify it in light of what we’ve seen in the series thus far.

On Experiential Distance

It’s commonly thought that when someone comes to a posek with a question, the posek must try to understand the person in order to tailor the correct halakhic answer to him. I have often objected to this view (following my friend Nadav Shnerb, in his essay “The Jewish Ark of Lies”). A halakhic ruling is the dry product of halakhic reasoning, and the tailoring to the individual should be done only by that person himself, not by the posek. The posek should present the various options and their costs, and the decision is to be made by the person (the questioner). See my essay “On Leniency and Stringency,” and in greater detail in the third book of the trilogy, Walking Among the Standing.

But that description is somewhat extreme. Understanding the questioner’s situation is important even just to present each of the halakhic options. Even when there is only one halakhic option, the decision regarding it is influenced by the circumstances in which it is applied—circumstances composed of objective factors and of the inner world of the person in them. In my essay on halakhic rulings in extreme situations, I argued that when there is experiential distance between posek and questioner, the posek cannot rule for him even if he wished to, and even the mere presentation of options is problematic. I brought several examples; below I’ll touch on some of them. First, I’ll offer a few non-halakhic examples to sharpen the claim.

Experiential Distance: Non-Halakhic Examples

1. Some challenged Abraham our forefather, who heard God command him to bind Isaac: why didn’t he suspect deception (Descartes’ deceiving demon)? After all, the command contradicted the moral norms God expects of us and the promise he received—“for through Isaac shall your seed be called.” He had every reason to suspect some error. This question ignores an important point: Abraham directly experienced the divine command. He was a prophet. We are not, and therefore it’s hard for us to grasp the situation in which God appears to someone and commands him something, and we toy with the possibility that it was a hallucination. It’s akin to a blind person challenging a sighted person: why trust what you see—perhaps it’s all an illusion? The sighted person has no way to explain the trust he places in his vision, but clearly such a challenge won’t shake him. Presumably the same applies to a prophet who experiences revelation.

This means we must be very careful when judging a situation not familiar to us firsthand. Lacking a direct encounter, we can be very confident and yet err in understanding the situation. The basis of this distinction is the difference between an external, intellectual understanding of the situation and a firsthand experiential encounter with it. We can sharpen this with the Mary’s Room example mentioned earlier.

2. Mary is a brilliant physicist who knows all of optics (the theory of light and vision) through and through, yet she has lived and worked her whole life in a black-and-white room. When she steps outside, she suddenly meets a world of colors. Did she learn anything new? Each color is but light of a particular wavelength, so apparently she learned nothing new. Yet clearly she encounters something new: she now understands what red is. Beyond the wavelength that produces it and the behavior of light at that wavelength, she now meets what it does in the real world. This nicely illustrates the difference between theoretical knowledge and firsthand encounter.
3. We saw another example in the previous series: Turing proposed his test for when a machine is a person. We found that after meeting computers that pass the Turing Test, it’s quite clear they aren’t human. Thinking theoretically about the situation is not the same as meeting the machine-person that passes it.
4. Consider how people abroad view our situation here in Israel. They’re sure we’re constantly shooting and fighting everywhere, and that our streets are battlefields. Sometimes those at Israel’s center see the settlements likewise. But those living inside the situation know that everyone there lives life like anyone else in the world. Physical and experiential distance generates misunderstanding—and can of course lead to bad decision-making about the situations in question.
5. Here’s another example that can sharpen this distinction. Imagine you’re driving and there’s a bird standing on the road ahead. Sometimes it waits until you’re half a meter away and only then flies off. Why not fly earlier? Why wait and take such a risk instead of acting with a greater safety margin? The answer is that due to the bird’s quickness, half a meter for it is like ten meters for a person. For the bird there’s no significant risk; it acts very much like we do. Our misjudgment of its behavior stems from experiential distance, since we’re viewing it in terms of our own parameters (our reflexes). That is a failure to understand the situation.
6. Another amusing example of a far-removed situation we can scarcely imagine appears in a Louis C.K. stand-up bit, hilariously depicting a tiger suddenly showing up at a commuter train station. In a hunter’s world that’s normal; in ours it’s unimaginable.
7. Another example: I’ve often been asked about yearning for redemption and rebuilding the Temple. Can we truly yearn for a Temple in which priests wade ankle-deep in blood and animals are slaughtered en masse? Indeed, it’s hard to wish for the Temple’s renewal. Yet we must consider that as long as we haven’t experienced a Temple and sacrificial worship, we have no way to know how that would affect our religious lives. Perhaps one senses the Divine Presence there and a completely different spiritual state of connection to God; then we’d understand that it’s right and enriching. So long as we’ve never experienced it directly, it’s not right to judge it.
8. In a recent Q&A I was asked precisely about this, and there I also cited my essay on the proper attitude toward secular Jews today, where I argued that we cannot form an opinion on the matter based on biblical or rabbinic sources that spoke to a completely different reality and didn’t know modern secularism. We live in a wholly different state, and Chazal couldn’t formulate a position about our situation today. The questioner there cited the aggadah about Rav Ashi, to whom King Menasheh appears in a dream and tells him that had Rav Ashi lived in his time, he would have lifted the hem of his robe and run to worship idolatry. True, that’s a halakhic example—and more such examples will follow in the next section.

Experiential Distance: Halakhic Ramifications

In the essay mentioned above on halakhic rulings, I dealt with questions that arose in the ghetto or the camps during the Holocaust. I argued that to rule on such questions the posek must understand the situation in which people lived and its meanings for them. Without that, he could reach patently mistaken conclusions. The experiential distance between us and such an extreme state doesn’t allow us to make halakhic decisions about it. The example I focused on in an earlier essay was monetary law in the Kovno Ghetto. Rabbi Gibraltar, who lived there, held that monetary law did not apply in the ghetto. People had no ownership of property; thus there were no concepts of theft, loan and repayment, damages, and so on. I won’t go into details and caveats here (see there). I cited a critical article by a contemporary halakhist and explained why he cannot judge such a situation. As an illustration, I offered a hypothetical: two people are in an elevator whose cable has snapped; it’s plummeting and will crash in thirty seconds. One asks the other for a pen to write to his family; the other refuses. May the first forcefully take it? On the face of it—certainly not. That’s theft, with no halakhic justification. Yet one could say that in so extreme a state, monetary law and property rights lapse, and one cannot speak of ownership under such conditions. So long as we’ve not lived through such a state, we cannot judge or understand it firsthand, and therefore cannot decide about it. The same holds for “life” in the ghetto.

In chapter 3 of that essay I also looked at the same extreme situation in the Holocaust (also in the Kovno Ghetto) from the reverse angle. I brought a responsum of Rabbi Oshri in his MiMa’amakim, which addresses halakhic responses to hard situations in the Holocaust (he too operated in the Kovno Ghetto). He discusses whether one may work in the camp kitchen and cook on Shabbat if the alternative is forced labor with real mortal danger, and whether that person may eat from what was cooked on Shabbat (the issue of ma’aseh Shabbat). Rabbi Oshri conducts a halakhic analysis from primary sources and concludes it is permitted. At the end he adds what I would have written at the outset: this is a case of pikuach nefesh, so there’s nothing to deliberate; clearly one may do whatever improves the chance to survive. I wondered there why he even bothers with the (to us) unnecessary analysis and offered two answers. One was that for those who had lived in the ghetto for years, this was normal life. To us it seems like total systemic collapse and an obvious case of pikuach nefesh, thus no deliberation at all. But for them, this was ordinary life; therefore it wasn’t self-evident that it was pikuach nefesh that permits everything (recall the earlier example of how people abroad view life here). This is the same phenomenon from the other side: an inability to decide when there is experiential distance—only this time leading to stringency rather than leniency.

Another, less extreme example I treated there: a modern, young community deliberating women’s roles in synagogue (Mourner’s Kaddish, teaching, receiving an aliyah, etc.). The community turns to its rabbi for a halakhic opinion. He is young and turns to his own teacher from yeshiva, an older, conservative figure living in a different world. The elder rabbi cannot grasp what these practices mean to members of that community whose world is far from his. In such a case the elder can advise his student, but the decision must be the student’s, not the elder’s: “When the words of the rabbi and the words of the student conflict, we heed the student.” The younger rabbi is less learned, but he knows the situation firsthand; only he can decide about it.

Consider another case. I’ve written more than once that, in my view, one can form a minyan via Zoom: people in Australia, Belgium, Morocco, Israel, etc., on the same Zoom call can combine to a quorum. This of course doesn’t meet the accepted halakhic requirements (under one roof, etc.), but I claim that in our era a site on the internet is literally a “site.” Those people truly are in the same place; it’s just that the place is virtual. Ask an older posek unfamiliar with the situation and he’ll reject it out of hand. But a posek who has experienced it can understand that, from the participants’ perspective, they indeed are in one place; therefore the halakhic conditions for combining to a minyan are met. Again, experiential distance can hinder a posek from making the right decision about a situation far from his world.

A final example along similar lines concerns the proper regard for “place” today. Biblically/Talmudically there’s an injunction of “lo titgodedu” against two courts of practice in one city. Does anyone know a city without two synagogues? People hardly notice that, on the face of it, each of us everywhere is violating a Torah prohibition (see the question I received in the Q&A). My explanation is that due to our world’s dynamism, “space” today is not geographic. The proper translation for our time is “two synagogues in one community,” not “in one city.” Likewise with the natural shift from the Talmudic norm, where the fixed custom was that of the place (minhag ha-makom), to our state where the fixed custom is that of the group—by ethnic origin rather than locale. Again, my claim is that the cause is our world’s dynamism. In the Talmudic world, a person was generally born and died in one place, like his parents and grandparents; hence “place” was the fixed frame of custom. Today a person wakes up in Tel Aviv and by evening is in Australia or Zimbabwe. Place has lost its meaning, and we build substitutes for it. In some contexts a virtual place replaces the geographic; in others, the eidah (ethnic group) or community replaces the place. In all these cases, a rabbi unfamiliar with the modern, dynamic situation won’t accept these shifts, since on their face they blatantly contradict halakhic sources. But one who lives the situation understands that this is the proper application of halakha to new circumstances. Again: experiential distance prevents correct decisions about situations far from us.

Thus far I discussed experiential distance and its implications. This is important for understanding the proper way to rule halakhically in general, but now I want to touch on the implications of this picture for an AI posek.

Back to an AI Posek: A Look at the Nature of Empathy

Our way of understanding another’s situation is through empathy. Empathy is the ability to enter the other’s state and live it. This is of course a human capacity, and all of us (aside from psychopaths) have it. But artificial intelligence does not, since this is a mental dimension of ours, and as noted in the series, an AI machine has no mental components. If halakhic ruling requires empathy and grasping the other’s situation, doesn’t that disqualify AI from ruling?

First, note a basic distinction. To rule halakhically, one doesn’t truly need to feel empathy toward the other. What’s needed is a firsthand understanding of his state. If I deeply understand his state even if I feel nothing (that is, I intellectually grasp his feelings and distress but have no empathic feelings), there’s no impediment to ruling for him. I’ll account for his state coolly and reach the correct halakhic result. If so, seemingly AI can do this as well. If it imitates us, it can factor in all elements of the situation and compute the right result without feeling anything. If it reaches the right result, the method doesn’t matter—there’s no bar to its ruling.

Perhaps we can distinguish two kinds of factors about the questioner that the posek must consider. There are the special objective circumstances in which he lives and acts; these can be described to AI, and it will “understand” them as it “understands” anything else and factor them into its ruling. But there are the questioner’s subjective feelings, and here it’s doubtful how well AI can grasp them and account for them. Is a verbal description of those feelings sufficient to yield the right outcome? I don’t know. Perhaps yes—after all, we too as humans are nourished by the person’s report of his feelings; the rest is our processing and response. But there’s also body language and our understanding that comes from entering the other’s situation, and I’m not sure AI can do this fully like a living person. Still, if we train the machine on many cases of such descriptions along with poskim’s responses, it will learn the human processing the posek does to what he hears. It probably won’t pick up the questioner’s body language—yet, in principle, even that could be captured by AI. If we train it on inputs that include not only speech but also video of the questioner while speaking, and of course feedback—the posek’s answers responsive to the speech and the video—I don’t rule out AI’s ability to factor those elements in as well.

The picture I’ve drawn here qualifies the concerns raised above, namely that AI cannot rule on delicate questions because it lacks empathy. Perhaps that’s not so: even if it lacks empathy in the emotional, human sense, it may be able to mechanically emulate human empathy.

A Scientific Note

A side observation arising from this is an interesting insight—and perhaps even an empirical way to test it. I suggested that our empathy might include something beyond verbal messages and body language, i.e., beyond what passes through the senses. Perhaps there are messages transmitted in another way, reflecting an ability to sense the other directly. If so, then it’s likely a machine cannot be trained on that; we have no way to feed it those inputs, and thus cannot train it to respond to them. That would be an inherent disadvantage of the machine relative to a flesh-and-blood posek. I don’t know if this is the case. Can we test it scientifically?

Perhaps yes—by looking from the other direction. Suppose we could train an AI that receives all the sensory inputs a posek receives: the questioner’s words, video of him while speaking, voice tone, and any other relevant sensory data—and of course the rulings given to each such questioner. Now take the trained machine and pose to it the same questions addressed to human poskim, and compare its answers to theirs. If there is a significant gap—i.e., we failed to train the machine to emulate human empathy—then that very gap would indicate that our empathy includes something beyond verbal messages and sensory inputs. Here AI could reveal something fascinating about us. This is itself a riveting scientific question, and I find it hard to think of another empirical way to test it.