Rabbinic Discourse on Artificial Intelligence is Upside-Down: Interlude #1

[NOTE: The following essay on a relatively neglected paradigm for thinking about AI, is a continuation of my six-months-stalled series on AI. Unfortunately, I still think that the assessment in the title is true. While a rabbinic literature on AI is emerging, it continues to be ‘upside-down’ in the sense that disproportionately greater attention is still being paid to the relatively more trivial questions. Within the next few weeks, I hope with God’s help to continue with Parts 2 and 3, along with another “Interlude” such as this one.]

Although there are many ways to count the “613 Commandments” of the Torah, prohibitions on sorcery number at least a few of those. As the Talmud points out, not all forbidden acts of sorcery are of the same severity, but at least some kinds of witchcraft are punishable by the death penalty (in theory). In other words, the Torah considers sorcery (or magic/witchcraft/etc) to be very very bad.

Most of the classical Jewish commentators (and bible scholars) believe that these prohibitions have something to do with supernatural manipulation of demonic forces or some other spookiness. But Maimonides, or “Rambam,” is quite explicit that sorcery is such an egregious sin because of its falsehood:

“But possessors of wisdom and those of whole knowledge know through clear proofs that all these things that the Torah prohibited are not matters of science [חכמה], but rather emptiness and vanity, which those deficient in understanding were drawn after and because of which they abandoned all the paths of truth. And because of this, the Torah said when it warned against all these vanities: ‘You shall be wholehearted with the Lord your God’ (Deut 18:13).” (Mishnah Torah Idolatry 11:16)

These are all vain and false things, which the Torah warned against just as it warned against falsehood… these are the root of idolatry and its branches… And the weak-minded came afterwards, and found those books [written on astrology] and those things and imagined that they are true and can provide benefit. But they do not know that they are falsehoods (Commentary to Mishnah Avodah Zarah 4:7; cf. Shemoneh Perakim #8, Guide of the Perplexed 3:37)

The very serious prohibitions of sorcery and associated practices, according to Rambam, are all forbidden by the Torah because God wants to distance His people from falsehoods. Especially false words written in books. Rambam felt the same way about astrology, writing:

Thus you ought to know that fools have composed thousands of books of nothingness and emptiness. Any number of men, great in years but not in wisdom, wasted all their days in studying these books… This is why our kingdom was lost and our Temple was destroyed and why we were brought to this.1

Astrology, divination, and all the rest are associated with idolatry because they are “emptiness and vanity.”

Concepts of practical magic such as golem-making are often invoked in contemporary discussions around Artificial Intelligence. But this is despite the Rambam, not because of him. It is appealing to refer to AI as “magic” for lots of reasons that Rambam would have hated (namely, because it works wonders but nobody really knows how it does it). But I want to start with Rambam because applying his idea to LLM-generated text is in some ways more straightforward -- anything ‘false’ is forbidden, just as idolatry is forbidden.

In other ways, though, applying Rambam’s view of sorcery is much less straightforward. The question of what exactly we mean by falsehood in this context might be better left for after-midnight dorm room conversations. Yes, we can give rigorous definitions for “truth” and “falsehood,” but the question of AI is a lot more subtle. ChatGPT doesn’t “lie,” it doesn’t even “hallucinate;” that word inappropriately ascribes some mental state to the AI as if it is suffering from schizophrenia (which it is not). But there is still something deeply false about it. The model is Artificial with a capital “A”, it has some set of numbers that corresponds to reality but it is not reality; it is several layers and mathematical transforms removed from reality. When you ask ChatGPT a question like “what color is the sky,” it does not send up a camera to take a look at the atmosphere and determine what color it is. Instead it does… something else, something Artificial.

Of course, AI is not really magic by almost any definition. It is based on texts that constitute representations of reality, and that, in and of itself, is nothing new. Consulting any book also exposes a person to ideas that are “one level removed” from empirical reality; you read about a thing instead of experiencing the thing. With AI, however, the methods are different; the model is not really “trying” to find the truth, just as Rambam writes of the astrologers that they had abandoned “all the paths of truth.” As mentioned, the process it uses is one that is artificial. Part of the issue may be one of orientation, not just a matter of facts -- and it is this kind of orientation that I worry we might train ourselves towards by relying on LLMs.

The most extreme falsehood, even according to Rambam, is not sorcery, but idolatry. Interestingly, there is plenty of rabbinic literature extending or applying the concept of idolatry to all kinds of things beyond the actual worshipping of idols. The Gemara in Sotah (4b-5a) compares arrogance to idolatry, explaining that the prideful person is as if he denies God. The Zohar (Tetzaveh 182b) and Rambam (Hilchot De’ot 2:3) refer to anger being equivalent to idol worship. Additional examples are easy to find. A particularly resonant one, given the state of AI, is what the kabbalist R. Chaim Vital writes in Shaarei Kedusha 2:5, that flattery is a form of idolatry. He brings a scriptural proof, but the connection is easy to see - flattery is ascribing traits to someone that they do not possess, a variation of ascribing divinity to that which is not God. The connection to LLMs should be obvious.

So yes, anything can be an idol, especially something that practically begs you to ascribe qualities to it that it does not possess. Incidentally, when it comes to AI, there is actually another whole level to this: as absurd as this sounds, a lot of the people building these machines are actually trying to build gods.

This is, I’m sorry to say, not a metaphor. It’s hard to catch them publicly being so explicit about it, but it’s quite clearly in the minds of the makers of ChatGPT, some of higher-ups at Google, and seems to be the sentiment in Silicon Valley generally. A co-founder of Waymo, the biggest self-driving car company that spun out of Google, even founded an institution called “Way of the Future” which claims to be dedicated to the worship of AI (but, you know, it might just be tax evasion).

Now there really is a funny-but-maybe-serious halakhic question here: if the creators and owners of a machine learning model, or its chat dialogue wrapper, or anything like that, serve it as a god, does it become prohibited? Generally speaking, an object that is worshipped as if it were an idol becomes prohibited as something that a Jew cannot benefit from, although the Gemara (Avodah Zarah 45a-46a) has all kinds of exceptions, such as if one were to make a mountain or a valley into an idol. Modern AI systems are housed in distributed data centers; it is quite unclear what physical object would even bear the prohibition. Software is not really a tangible thing, and math is even less so.

While it’s a fun question to see if you can come up with some technological product that would conceivably bear such a prohibition of avodah zarah; it is almost certainly not the case today (as always “ask your local Orthodox rabbi”). Obviously, I do not advocate data center terrorism in the name of Avraham who smashed his father’s idols (that advice, by the way, does not require asking your local rabbi; you’re welcome). But beyond any strict halakhic ruling, I think this is an interesting framework for thinking about some of the dangers of AI - it is the dangerous allure of falsehood, of cheating reality.

I kept going back and forth about whether or not to include some of the LLM-generated answers to Torah questions, because some of them are genuinely scary, in the sense that they sound really convincing but are wrong or misleading in ways that can only be discerned by thinking hard. Although I have seen dozens of examples, in the end I’m not going to clutter this piece with such nitpicking, because nitpicking is what is required.

All this is to say that the rabbis who are writing about whether you can use AI to study Torah, to pasken, to prepare a shiur or sermon, etc. are not exactly asking the wrong questions. But I think we should be much more hesitant to adopt these tools, even minimally. I don’t know if idolatry is the right framework; I think we need lots of mental frameworks to help think this through, and the lens of idolatry is a relatively unexplored one.2 When it comes to Torah learning and teaching, the answer is not going to be a “yes” or a “no,” it is going to have to explain “how” and “why.” Why are we learning and teaching Torah? And if we are committed to religious truth, then how precisely should we be using technological tools, if at all? What guidelines are necessary?

Articles that ask questions without answering them can be annoying, but I am not really in a position to provide that guidance; I’m just a guy. And I hate calling out specific people or naming any institutions to complain about, so I’ll just say that I am very upset by some of the latest embrace of AI by reputable Torah/Jewish organizations. Instead, let me identify a Jewish (not Orthodox!) institution that I believe is doing a good job: Sefaria. Sefaria has a clear statement about how they use AI, how they do not, and how the users can tell the difference.3 Because of the nature of their product and the people they serve, Sefaria has a much easier job than many other Jewish educational websites and institutions. But I think we need these things, we need them to be absolutely clear, and we need them soon.