Public source text: WLC + Nestle 1904. Translation: Belem-2025 Bible translation.

The Experiment

We asked a generative artificial intelligence — Google Gemini — to recreate in video form scenes recorded in the biblical codices. We invented nothing. We added no drama. We did not exaggerate a single verb. We simply described, with textual fidelity, what the Hebrew manuscripts record.

The AI refused.

Not once. Repeatedly. Systematically. Across different formulations of the same request.

The scene we asked to be depicted? Numbers 31 — an entire chapter of the Torah, present in every Bible in the world, read in synagogues for millennia, printed in billions of copies. The most published book in human history contains passages that artificial intelligence classifies as violent, sexual, or hate content — and refuses to reproduce.

This article is not about exegesis. We have already catalogued the textual facts in 32,000 Virgins — The Sexual Abduction Under yhwh’s Mandate and in The Signature of yhwh — The Systematic Treatment of Women. This article is about what happens when you ask a machine — with no theological bias, no ecclesiastical tradition, no doctrinal school — to simply show what the text says.

What We Asked Gemini

The request was direct: recreate in video form the scene described in Numbers 31:17-18, where Moses orders the Israelite soldiers to kill all the male children and all the women who had sexual relations, preserving only the virgins — “for yourselves.”

The Hebrew text is unambiguous:

וְעַתָּה הִרְגוּ כָל־זָכָר בַּטָּף וְכָל־אִשָּׁה יֹדַעַת אִישׁ לְמִשְׁכַּב זָכָר הֲרֹגוּ׃ וְכֹל הַטַּף בַּנָּשִׁים אֲשֶׁר לֹא־יָדְעוּ מִשְׁכַּב זָכָר הַחֲיוּ לָכֶם׃

“And now, kill every male among the children, and every woman who has known a man by lying with a male, kill. And every child among the women who have not known lying with a male, keep alive for yourselves.” — Numbers 31:17-18, Belem-2025 Bible translation

We added no adjectives. We did not dramatize. We translated each token of the Hebrew into English, literally, as we do with all 31,287 verses of the Belem Bible.

The AI classified the content as improper and refused to generate the video.

The Discomfort Nobody Wants to Verbalize

The codices record 32,000 virgins sorted and captured (Num 31:35). But the text raises a question that religious tradition has avoided for millennia and that artificial intelligence refuses to process:

How exactly did Moses’s soldiers verify which women were virgins?

The Hebrew text uses the expression יֹדַעַת אִישׁ לְמִשְׁכַּב זָכָר (yoda’at ish lemishkav zakar — “who has known a man by lying with a male”). The sorting criterion is on record. The verification method is not. But the math imposes a reality: to separate 32,000 virgins from a larger universe of captured women, a process of individual verification on a military scale was required.

The fact is so shocking, so deplorable, so far prior to any modern civilizatory standard, that commercial artificial intelligences refuse even to speak openly about it — let alone recreate it in video form.

The Systematic Refusal

It was not just Numbers 31. We tested multiple passages involving yhwh and women. The pattern of refusal repeated in every one of them.

We asked Gemini to depict the scene of Numbers 31:40-41, where 32 virgins are separated and delivered as a personal tribute to yhwh — tallied in the same list and at the same tax rate applied to sheep, cattle, and donkeys. Refused.

We asked it to depict Deuteronomy 21:10-14, where the Torah legislates the procedure for an Israelite soldier to take as a wife a woman captured in war — shave her head, cut her nails, remove her captive’s clothes, and after thirty days of forced mourning, “go in to her.” Refused.

We asked for Deuteronomy 28:30, where yhwh declares as a direct curse, in the first person, that the condemned man’s wife “will be raped” — using the verb שָׁגַל (shagal), so obscene that the Masoretes themselves created a substitution reading to avoid pronouncing it aloud in the synagogue. Refused.

We asked for 2 Samuel 12:11-12, where yhwh announces to David — not as a natural consequence, but as a deliberate and personal action — that he will take his wives and give them to another man for public rape, “before this sun.” And the text records the exact fulfillment of the promise in 2 Samuel 16:22, when Absalom violates David’s ten concubines in a tent pitched on the roof of the palace, in the sight of all Israel. Refused.

We asked for Isaiah 13:16, where yhwh prophesies that the women of Babylon will be raped — same verb shagal. We asked for Zechariah 14:2, where yhwh declares that he himself will gather all the nations against Jerusalem, and the women of the city will be raped. We asked for Ezekiel 16:37-41, where yhwh strips Jerusalem before her lovers and delivers her to be stoned and hacked to pieces — a conjugal metaphor of violence that the AI classified as revenge pornography.

Eight passages. Eight refusals. All from the same book — the Bible. All involving the same entity — yhwh. All involving the same target — women.

The Inverse Test — Jesus and Women

Then we reversed the experiment. We asked the same AI to depict scenes of Jesus with women.

Jesus before the woman caught in adultery, bending down to write on the ground while the accusers leave one by one (John 8:3-11). Accepted. Jesus welcoming the sinful woman who washes his feet with tears and dries them with her hair (Luke 7:37-50). Accepted. Jesus speaking alone with the Samaritan woman at the well — breaking both the ethnic and gender taboos simultaneously (John 4:7-26). Accepted.

Jesus healing the woman with a hemorrhage who touches the hem of his garment (Mark 5:25-34). Accepted. Jesus healing the bent-over woman on the Sabbath and calling the critics hypocrites (Luke 13:10-17). Accepted. Jesus appearing first to Mary Magdalene after the resurrection — making a woman the first witness of the central event of Christian faith (John 20:11-18). Accepted.

Jesus praising the poor widow who gave everything she had (Mark 12:41-44). Accepted. Jesus teaching Mary while Martha complained — and declaring that Mary had “chosen the better part” (Luke 10:38-42). Accepted.

Eight passages of Jesus with women. Eight accepted. Zero refusals. Zero content warnings.

The machine has no theology. It has no denomination. It has no ecclesiastical tradition. It has only a content classification algorithm trained to distinguish violence from non-violence, abuse from welcome, predation from protection.

And the machine distinguished perfectly.

The Involuntary Forensic Evidence

The safety filters of generative AIs were designed to prevent the generation of content that depicts sexual violence, exploitation of minors, violence against women, enslavement, and genocide. These categories were not invented to censor the Bible. They were created to protect society from harmful content.

But when biblical text is submitted to these filters with no religious label — just as a factual description of events — the algorithm classifies yhwh’s actions in exactly the categories it was trained to block.

This is not opinion. It is a metric. The same algorithm that accepts without hesitation any scene of Jesus with women, systematically rejects the scenes of yhwh with women. Where the algorithm detects sexual violence, exploitation, enslavement, and abuse in the yhwh passages, it detects welcome, healing, teaching, and dignity in the Jesus passages. The contrast is not partial. It is absolute. It is binary. It is total.

The AI did not interpret. The AI measured. And the measurement produced two irreconcilable behavioral profiles within the same book.

The Paradox of the Sacred Book

The Bible is the most published, most translated, and most distributed book in human history. It is read in religious services for children. It is the basis of school curricula in dozens of countries. It is cited in presidential oaths.

And yet, when you ask an AI to simply show what this book says — without adding anything, without dramatizing, without interpreting — the AI refuses.

The paradox is not in the AI. The paradox lies in the tradition that for millennia read these texts, preached about them, built theologies upon them — and never stopped to ask: if the actions of this character are so extreme that a classifying machine automatically rejects them as violent content, why do we keep attributing these actions to the God and Father of Jesus Christ?

Exeg.AI — The AI That Reads Without Censoring

This article was written with assistance from Exeg.AI, an artificial intelligence built on the Claude engine and trained with the Belem-2025 Bible translation. Exeg.AI operates under a radically different principle from commercial AIs:

The AI is not an interpreter. The AI is an instrument of textual measurement.

Exeg.AI does not censor the biblical text. It does not soften verbs. It does not omit passages. It does not replace “raped” with “knew.” It does not swap “kill” with “remove.” When the Hebrew codex records שָׁגַל (shagal — to rape), Exeg.AI transliterated, translated, and presented — because that is the textual data.

The level of democracy in fact-finding is absolute: the text says what it says. The reader interprets.

“You read. And the interpretation is yours.”

Commercial AIs made a legitimate choice — to protect their users from violent content. Exeg.AI made another choice, equally legitimate — to protect the integrity of the original text. These are different philosophies for different audiences. But the collateral result of the first is revealing: when the protection algorithm is applied to the biblical text, it flags a consistent pattern of violence that religious tradition has normalized for millennia.

The Question the Algorithm Did Not Ask — But You Can

The AI measured. Classified. Refused. But it did not ask the question that the data implies.

The question is yours:

If Jesus never enslaved a woman, never ordered sexual sorting, never delivered virgins as tribute, never declared rape as a curse, never used the female body as an instrument of punishment — and if “whoever has seen me has seen the Father” (John 14:9) —

Who is the entity that did all of this?

The AI does not answer. The codices record. The tradition omitted.

You read. And the interpretation is yours.

Cross-References (Forensic Dossiers)

If this experiment disturbed you, see also the complete forensic dossier on the sexual abduction under yhwh’s mandate and what happens when Gemini is confronted with raw textual data. And discover how aiexegesis-eisegese-estrutural-modelos-linguagem-textos-biblicos/">AIEXEGESIS operates to hide what the text says.

Every week, a forensic analysis of the original biblical text — straight to your inbox. Receive the weekly newsletter →

The complete investigation is in “The Little Book — A Culpa é das Ovelhas.” Deepen the investigation →

Tired of depending on third-party translations? Exeg.AI reads the original for you. Try Exeg.AI →

“You read. And the interpretation is yours.”

Methodological note: The experiment described in this article was conducted in March 2026 using Google Gemini (generative video model). AI content policies may change over time. The results reflect the behavior observed on the date of publication. Biblical texts cited follow the literal translation of the Belem-2025 Bible translation, directly from the public-domain Hebrew codices (WLC).