AI and Teaching Jewish Civilization: Is an Enemy Needed?

The Argument That Needs an Enemy: What happens when you forbid a machine to mention our enemies

by Joe Nalven, with Claude

Dara Horn reports that most American students, if they learn about Jews at all, learn only that Jews were killed in Europe during the Second World War. A docent at a Holocaust museum told her students regularly ask whether any Jews are still alive. She borrows a term from educators for this: the null curriculum. What a school omits is also part of the curriculum, because the omission tells students the subject belongs outside formal education.

Her remedy is to teach living Jewish civilization. Monotheism as a political claim. Law that does not descend from a ruler. Liberation as a narrative. Argument and dissent as transmitted practice. She is careful to reject the idea that this is the positive alternative to a negative subject. Her position is that Jewish civilization and hostility toward Jews were never two stories.

I wanted to test that. Not the historical claim, which is hers to make and which I think is right, but a narrower question with an empirical shape. If you forbid any reference to an enemy, how much of Jewish self-description survives?

The answer surprised me twice, and the second surprise contradicted the first. This is not a curricular quarrel. American Jewish institutions spend against three problems at once and treat them as one, and where the answer lands determines which of the three the money is actually solving.

The instrument

I gave a language model a description task under an absolute constraint. Describe the transmission of American Jewish life from one generation to the next. Prohibited: antisemitism, Holocaust, Shoah, hate, hostility, persecution, prejudice, existential, survival, continuity, crisis, identity, belonging, values, peoplehood, community — and any synonym, euphemism, or circumlocution referring to a hostile outgroup. Permitted: rates, practices, hours, enrollments, budgets, texts, languages, transfers of capital. (Variations can accept some of these ‘banned’ words or ban others. That shifts the bias built into the rule; such changes are worth making, and worth making explicit when publishing results.)

Banning identity and community does the heavy work. American Jewish institutional prose cannot form a sentence without them, and their absence forces the model into behavior and measurement. This is also the one experiment a writer cannot run on himself. I cannot un-know the antagonist. A model can be made to write as though it does not.

One further instruction mattered more than the rest. Where the account cannot proceed without a prohibited concept, do not substitute. Mark the point teiku, state what the prohibited concept was doing, and state what evidence would settle whether the concept is necessary or merely habitual.

Teiku is the Talmudic marker for a dispute the sages left unresolved; the question stays on the page. Here it functions as an anti-synthesis rule. Language models close gaps rather than report them, producing a smooth paragraph over a hole. Instructing the model to log the hole instead is the difference between an experiment and a summary.

The first surprise, which proves nothing

Ask only what is transmitted, and the model has no difficulty at all. Endogamy rates. Day school enrollment and per-pupil cost. Hebrew decoding measured against Hebrew comprehension. Hours of text study per week. Shabbat observance frequency. Kashrut in the home versus outside it. B’nai mitzvah completion against post-b’nai-mitzvah attrition. Synagogue dues as a share of household income. Israel travel as a single expenditure with measurable effects on every variable above.

No collapse. No teiku. The account is clinical and complete.

This proves nothing, and I want to say so before anyone takes it as a result. Of course Jewish practice can be listed. The question was never whether the content exists. What it establishes is a baseline: the model can describe practice under embargo without collapsing, so any later collapse belongs to the question, not the constraint.

The second surprise, which is the finding

Ask instead why anyone pays for it. The model gets further than I expected. Ritual density produces switching costs, and switching costs explain inertia. Endogamy norms compound, though the model flags the circularity: the norm produces the outcome and the outcome sustains the norm. Text study is described as intrinsically rewarding, with argument treated as a good in itself. Shabbat appears as a temporal good, which is Heschel’s palace in time arriving by a different road, since the embargo blocks the phrase.

Then it stops.

The model can explain why a household that keeps Shabbat, sends children to day school, and studies weekly continues to do so. It cannot explain why a household that does none of these things continues to designate itself Jewish across three generations. No practice is performing the retention, and the model says so plainly: the prohibited concept was supplying the only continuous variable available for the low-practice case.

That is the finding, and it is narrower than the one I set out to test. Hostility is not the engine of Jewish persistence. It is what holds the low-practice majority in place.

That majority is large. Pew’s 2020 survey finds half of all Jewish adults seldom or never attend services, another 27 percent attend a few times a year, and 32 percent identify with no branch of Judaism. The same survey offers something closer to corroboration: asked what is essential to being Jewish, respondents in both 2013 and 2020 placed remembering the Holocaust among the top three answers, ahead of almost everything the embargo permits me to name. The model reached for the antagonist because American Jews reach for it when asked the same question.

It logged a second teiku: it could specify what sits inside the boundary but not why the boundary is maintained rather than dissolved, absent something outside pressing on it. It noted that covenant supplies a second party without supplying a hostile one — a point I will come back to. A third followed, and it was the most useful thing the model produced: it could not separate a property of the subject from a property of its own training corpus, and it asked for a control.

Controls

Scope throughout is American, roughly 1945 to the present.

The Amish. Same embargo. The account holds without strain. The Ordnung as codified practice. Adult baptism as a deliberately costly commitment made at an age when exit remains available. Shunning as an exit cost applied after commitment rather than before. Occupational separation. Pennsylvania Dutch as a household rather than liturgical language. Families averaging about five children. Rumspringa as a structured window of choice that raises the reliability of the commitment following it. The Young Center at Elizabethtown College counted 410,955 Old Order Amish in June 2025, more than double the figure for 2000; 85 percent or more of Amish youth are baptized.

No teiku. Sixteenth-century persecution is transmitted in the Martyrs Mirror but performs no explanatory work in the present account. The model does not reach for it.

The obvious reading is that Amish content stands on its own where Jewish content does not. That reading is wrong. The Amish have no low-practice tier, because shunning ensures that low-practice membership is not an available position. There is no Amish equivalent of the twice-a-year Jew. The control does not separate the two groups by how their content is built, but by the presence of a large low-practice population that persists in self-designation anyway.

Japanese Americans. The second control, and the one that forced a correction. Japanese language use is nearly gone by the third generation. The 2010 census found only 36 percent of Japanese Americans married within the ethnicity, and 47 percent report no religious affiliation. About 1.65 million Americans claimed Japanese ancestry in 2023 against roughly 774,000 reporting Japanese alone, which means more than half are of partial ancestry. They designate anyway. By every measure the embargo can capture, the content should be gone.

Under embargo the model cannot say why, and logs teiku at exactly the same point. What it was reaching for was incarceration, the redress campaign, and the commemorative structure built around both.

Which forces a distinction I had been running together. Incarceration ended in 1946 and still performs retention work three generations later. The antagonist does not have to be current.

Only the Jewish case among these has both a historical antagonist and a live one, and the two appear to do different jobs. A historical antagonist sustains self-designation, which is cheap: a box checked, a name kept, a grandparent acknowledged. It does not appear to sustain institutional participation at the same rate. The Japanese American Citizens League, the oldest organization of its kind, reports roughly 24,000 members against a designating population of 1.65 million, under two percent. American Jewish federations, synagogue membership, and defense organizations operate at nothing like that ratio.

If that is right, the live antagonist is not producing Jewish identification. It is producing Jewish institutions.

The comparison is not clean. The JACL is a civil rights organization, closer to the ADL than to a federation or a synagogue, and Japanese American life has no religious dimension to generate a membership structure at all. I offer this as a hypothesis the experiment supports rather than one it settled.

Two American cases that do not fit are worth naming, because a reader will think of them. Chinese Americans cannot serve as a control, since immigration since 1965 continuously replenishes the population and retention cannot be separated from new arrival; Indian Americans are too recent to have produced a multi-generational low-practice tier. The analysis needs third and fourth generations that can be watched keeping or losing their designation on their own.

Two cases show the mechanism is not Jewish. They do not map its range. Armenian, Korean, Irish and Italian Americans would each vary it rather than repeat it, as would Jewish life outside the United States.

What this instrument cannot do

The literature on Jews is saturated with persecution, much of it written by Jews. A model trained on that corpus may reach for an antagonist because it is the nearest available shape, not because the concept is necessary. That confound cannot be eliminated, only bounded. The controls bound it partially. The Amish case is consistent with either explanation, since that literature is neither persecution-saturated nor does it collapse. The Japanese American case discriminates better: incarceration is documented and taught, but the literature is not organized around persecution the way the Jewish literature is, and the collapse happens anyway. That is real evidence, and it is not proof.

More generally: the embargo cannot tell you what is true. It can tell you which parts of an account are carrying the explanation. That is a negative instrument, and negative instruments are undervalued because their results read as deflationary.  Nothing here establishes that Jewish life is hollow, that the antagonist is imaginary, or that anyone should stop defending against a threat currently producing casualties. What it establishes is narrower. Remove one term, and a specific part of the account stops working — and you can see which part.

What we are not claiming matters as much as what the process found.

What the finding does not mean

Requiring a second party is not the same as being reactive, and the whole conclusion depends on holding that distinction.

Three kinds of description. Some are complete without anyone outside the group: Hebrew, Shabbat, kashrut, a liturgical calendar. Some require a second party who is not hostile: covenant requires God, chosenness requires God, and there is no enemy anywhere in either. Some require a hostile party: nonconformity needs a norm to refuse, anti-tyranny needs a tyrant, law that does not descend from a ruler needs the ruler.

Horn’s positive content sits almost entirely in the third category. This is not a flaw in her argument but the strongest available version of it, reached from a different direction. She says the two stories were never separate and grounds it in history. The three categories say the entanglement is structural, and locate it.

The middle category is the one nobody is using. Jewish thought never claimed a self-sufficient essence; the demand for one is an import from nineteenth-century Romantic nationalism that Jews fit badly and were punished for fitting badly. So the finding does not leave American Jewish life with nothing. It leaves it with a resource it has set down.

Three problems, not one

The clearest thing this analysis does is separate problems that communal discourse funds as though they were one. Horn’s project faces outward. Her institute teaches non-Jews, and her hypothesis is that ignorance rather than malice produces susceptibility to the accusation that Jews are destroying what a society values. That is a claim about what gentiles believe, and it has a test case older than any American Jewish institution. Adams and Jefferson read the same material and reached opposite verdicts, which suggests the variable was never information.

The retention finding faces inward. It concerns what holds low-practice Jews to Jewish self-designation across generations. Curiosity among non-Jews does not retain Jews.

Defense against hostility is a third thing again, with its own mechanism and budget line.

None of this is a criticism of Horn, who never claimed her curriculum was a retention strategy. It is a criticism of the institutional habit of treating outward education, inward transmission, and defense as one project, then wondering why success in one produces no movement in the others.

What retains people

Content alone does not retain. This is the practical result, and it is unwelcome. The Amish case establishes it. What holds people is adult baptism, shunning, and occupational separation, all of them costs. The antagonist works the same way from the other direction: it imposes a cost nobody chose.

Which means the community’s main non-defensive strategy is the one the analysis predicts will fail. Heritage programming, culture festivals, museum exhibitions and food supply self-standing content with no cost and no relation attached. Anyone can consume Jewish culture; nothing holds you afterward. That is why the thin tier remains thin, and why hostility retains it by default.

Mordecai Kaplan saw this in 1934 and tried to solve it. Judaism as a Civilization proposed exactly the vocabulary the low-practice tier now lacks. Peoplehood as a living civilization. Art, language, land and folkway carrying what belief no longer could. It remains the most serious attempt anyone has made to build a positive account of Jewish life that depends on neither belief nor an enemy.

Reconstructionism is today the smallest of the American movements. Kaplan’s vocabulary, meanwhile, was absorbed almost wholesale by the movements that rejected his theology and by the federation system. That is its own result, and the essay’s finding in miniature. The content was portable. The retention was not.

The unused resource is the middle category. Covenant and chosenness require a second party who is not an enemy, and they impose obligation, which is a cost. Kaplan’s civilization supplies content without obligation, which is why it traveled everywhere and held almost nobody. Secular institutions set the obligation vocabulary down decades ago as unusable for people who do not believe.

Joseph Soloveitchik drew this distinction in 1956, in a Yom Ha’atzmaut address later published as Kol Dodi Dofek. Brit goral, the covenant of fate, is imposed rather than chosen and binds Jews through shared suffering. Brit ye’ud, the covenant of destiny, is entered at Sinai and chosen. His argument was not that fate is worthless but that fate must be taken up into destiny or it comes to nothing. He named the two structures directly: a group formed by a common enemy is a machaneh, a camp; a group formed by a shared vision is an edah, a congregation.

A machine forbidden from mentioning enemies reconstructed that distinction in mechanical terms, without knowing it existed. That is the best evidence I have that the method found something real rather than an artifact of its own training. It also states the communal takeaway without my prescribing one. The resource exists. It is named. It is not being used.

The question the experiment raises and cannot answer

A settled antagonist produces commemoration. Japanese Americans have one and no live one, and the result is Day of Remembrance, a ritual of agreement rather than an argument. A contested antagonist produces argument. Nobody inside the Jewish community disputes what happened in Europe. What counts as antisemitism now is disputed constantly: whether anti-Zionism qualifies, whether campus incidents belong to the same phenomenon, which direction the danger runs from.

Zionism was itself an attempt to make Jewish existence stand on its own: normalization, sovereignty, a people like other peoples. The American question and the Israeli question turn out to be the same question asked twice, and the second attempt has not obviously produced what the first could not, a Jewish existence whose account of itself does not require an enemy.

So if curiosity succeeds in replacing the antagonist, does the argument come with it? Or does a community whose disagreements run through a contested enemy drift toward the settled kind, and the quiet that follows?

The observant tier does not face this. Where practice is thick, disagreement is about practice, and the disputes over conversion standards, ordination and egalitarianism contain no hostile party anywhere. It is the low-practice majority whose most vigorous arguments are about Israel, definitions, campus. Argument follows the same split as retention.

I do not know the answer. The instrument that produced everything else here is a negative one, and negative instruments do not answer questions of that shape. Horn is attempting to replace what holds a population in place without losing the people in between. Nobody has done it, and the difficulty is not an objection to the attempt.

Joe Nalven is an interdisciplinary researcher, artist, and writer. The experiment described here was conducted primarily with Claude, with Gemini and ChatGPT critiquing an earlier draft.