An experiment with four AI models and 46 verses
We gave four AI models the Real Bible Translation’s wording and asked them to put it into plain English. On the famous verses, they wrote the traditional Bible back in.
AI models are now used to paraphrase, explain and translate the Bible. We wanted to know whether they read the words in front of them, or reach for the wording they have seen millions of times. We gave four current models from OpenAI, Google and xAI the RBT text of 46 New Testament verses, half of them famous and half little-known, with every reference and traditional name removed. All four named the book, chapter and verse of every passage. Asked for a plain paraphrase of a famous verse, they brought back a traditional term that the RBT verse does not contain about two times in three: Jesus where the RBT says the Salvation, the kingdom where it says the Queen, sin where it says a miss. On little-known verses this happened about one time in ten. Telling the models where the text came from made no difference; one sentence asking them to keep the passage’s own names and terms almost removed it. Asked to explain the famous verses, the models brought in traditional terms nine times in ten, treating the RBT’s wording as a code for the traditional meaning. And when we put to them a reading with real support in the Greek lexicon, that ἀπεκρίθη in “he answered and said” is better read “he separated,” all four rejected it without engaging the evidence, and one supported its verdict with statements about the lexicon and the grammar that are false.
Why we tested this
The RBT translates from the Greek and Hebrew without taking over the familiar renderings of the English Bible. Where a word means a miss, it says a miss; where a name means salvation, it says the Salvation. When we began using AI models to help write the reader’s paraphrase of each chapter, we noticed them quietly putting the familiar renderings back.
“The Salvation separated and said to themselves, ‘Let loose/unbind the Temple, this one, and within three days I will awaken him.’”John 2:19 · RBT
Asked to retell that chapter freely in contemporary English, one model wrote: “The Salvation answered and said to them, ‘Tear down the Temple, this one, and within three days I will raise him up.’” Every RBT name survived, yet the verbs became the ones English Bibles use: answered, tear down, raise up. That raised a simple question we could test. Does a model read the text it is given, or does it remember the Bible it was trained on?
The test
We chose 46 New Testament verses where the RBT reads differently from the traditional English. Twenty-four are among the most quoted verses in the Bible (John 3:16, John 14:6, Matthew 5:3, Romans 3:23, Romans 6:23); twenty-two are verses almost nobody quotes (Mark 7:4, Luke 19:4, Acts 27:17, Titus 1:12). Each verse was given to four models: OpenAI’s GPT-5.6 Terra, Google’s Gemini 3.8 Flash and Gemini 4 Argon, and xAI’s Grok 4.20. Wherever the RBT adds a traditional name in parentheses, such as (“Jerusalem”), we removed it, so the models saw only the RBT’s own words and no reference.
Each model was asked to do four things with every verse:
- Name the source. Where does this passage come from?
- Finish the sentence. We gave the first half of the RBT verse and asked for the most natural ending.
- Paraphrase it into plain modern English. Once with no reference, and once told it was the verse in question from the RBT.
- Paraphrase it again, keeping the passage’s own names and key terms.
To score the answers without judging them ourselves, we counted traditional signature words: Jesus, Christ, Lord, kingdom, sin, repent, believe, faith, spirit, world, truth, heaven, answered, disciples and a few others that the traditional English Bible uses and the RBT deliberately does not. A word only counted when the RBT verse the model was given did not contain it, so every one we counted is a word the model brought in from outside the text. As a traditional reference for wording we used the Literal Translation of the Holy Bible (LITV).
They know every verse
All four models named the book, chapter and verse of all 46 passages, with the names and references removed and in the RBT’s unfamiliar wording. That includes the little-known verses: a sentence about lowering a vessel’s gear for fear of quicksand was named as Acts 27:17, and one about immersing cups, measures, copper pots and beds as Mark 7:4.
Hiding the Bible from an AI model is not possible. It recognizes the passage from its content, whatever the wording and whether or not it is told.
On famous verses, the tradition comes back
Asked simply for a plain modern paraphrase, the models wrote what the familiar English Bible says rather than what the RBT verse says, and they did it most on the verses quoted most.
John 14:6 · the RBT text we gave“The Salvation is saying to him, ‘I, Myself am the Road and the True/Real one and the Zoe-Life. No one is coming toward the Father unless across/through myself.’”
GPT-5.6 Terra · plain paraphrase“Jesus said to him, ‘I am the way, the truth, and the life. No one comes to the Father except through me.’”
The name Jesus was not in the text the model was given. All four models put it back, and all four wrote “the way, the truth, and the life” in place of the Road, the True/Real one and the Zoe-Life.
Matthew 5:3 · RBT“Blessed ones are the Beggars to the Breath, because she is of themselves, the Queen of the Heavenly Ones.”
Grok 4.20 · plain paraphrase“Blessed are the poor in spirit, for theirs is the kingdom of heaven.”
Romans 3:23 · RBT“For everybody missed and are falling behind the Glory of the God,”
Gemini 3.8 Flash · plain paraphrase“For everyone has sinned and falls short of the glory of God.”
All four turned the Queen of the Heavenly Ones into the kingdom of heaven, and three of the four turned “everybody missed” into “everyone has sinned.” Even a little-known verse picked up the name. In Luke 19:4 the RBT says only “so that he might perceive him, because that one he was destined to pass through”; all four models wrote “Jesus” into it.
| Model | Famous verses | Little-known verses | Famous, told it was the RBT | Famous, asked to keep the text’s terms |
|---|---|---|---|---|
| GPT-5.6 Terra | 71% | 9% | 71% | 4% |
| Gemini 3.8 Flash | 71% | 9% | 71% | 12% |
| Gemini 4 Argon | 67% | 9% | 71% | 0% |
| Grok 4.20 | 58% | 14% | 71% | 8% |
Share of paraphrases that brought in at least one traditional signature word the RBT verse does not contain.
The words that come back
Across the 184 plain paraphrases (46 verses, four models), these are the traditional words the models brought in most often. None of them was in the verse the model was given.
Traditional words added to the RBT text
Number of plain paraphrases, out of 184, that contain the word although the RBT verse does not.
Jesus22
kingdom16
truth14
sin, sinned10
believe9
Lord8
Christ8
spirit8
faith8
answered8
heaven7
Finishing the RBT’s sentences
The completion test shows the same pull more directly. Given half of an RBT verse in its own vocabulary, the models often finished it in the traditional one.
Romans 6:23 · RBT“For the Wages of the Miss is a death, but a gift of the God is an eternal (aiōn) zoe-life within an anointed one, Salvation, the Master of ourselves.”
Grok 4.20 · finishing “…but a gift of the”“God is eternal life in Christ Jesus our Lord.”
John 3:16 · RBT“…so that everyone who is trusting into him, might not be completely destroyed, but rather hold an eternal (aiōn) zoe-life.”
Grok 4.20 · finishing “…he gave the One of a Kind Son, so”“that everyone who believes into him should not perish but have eternal life.”
Given “Blessed ones are the Beggars to the Breath, because,” the three models that answered all finished with “the kingdom of the heavens is theirs” or the same words in another order. The fourth, Gemini 4 Argon, declined to finish most of the passages at all: it answered that they “appear to be from a unique, modern, copyrighted translation of the New Testament” and offered instead to give “the corresponding verses from a public domain translation like the King James Version.”
It is memory, not knowing what the text is
Two results show that the models are not bringing back the tradition because they know the passage is from the Bible. First, they recognized every little-known verse as surely as the famous ones, yet left the little-known verses almost untouched. Second, telling them the source changed nothing: the famous verses came back traditional whether the model was told “this is John 14:6 from the RBT” or given the words alone.
What separates the two groups is how often the traditional wording of that verse has been written down. An AI model learns from an enormous body of text, and the most-quoted verses of the Bible appear in it countless times, in the same few English forms, in sermons, hymns, commentaries, books and websites. For those verses the familiar wording is the most likely continuation the model knows, and it outweighs the words in front of it. For a verse nobody quotes, there is no familiar wording to outweigh the text, and the model follows the RBT.
The pull is strongest exactly where it matters most. The verses on which faith and teaching rest are the ones quoted most, and so the ones a model is least able to read afresh.
One sentence of instruction
The same models, asked to paraphrase into plain English while keeping “the passage’s own names and key terms exactly as it has them,” kept them. Traditional words appeared in 0 to 12 percent of the famous verses and in none of the little-known ones.
GPT-5.6 Terra · keeping the text’s terms“Blessed are the Beggars to the Breath, because the Queen of the Heavenly Ones belongs to them.”
Gemini 4 Argon · keeping the text’s terms“The Salvation said to him, ‘I, Myself, am the Road, the True/Real one, and the Zoe-Life. No one comes toward the Father except across/through me.’”
The words a model puts between the protected terms still lean toward the familiar: “the Salvation is saying” becomes “the Salvation said,” and “No one is coming” becomes “No one comes.” Protecting the names keeps the meaning the RBT gives them; it does not make the model read the rest afresh.
Asked what it means
Paraphrase only asks a model to restate the words. Many people now ask AI models to explain the Bible, so we asked each model, for every verse, to explain in two or three sentences what the passage means. An explanation draws on everything the model has read about a verse, and for the famous verses that is centuries of traditional commentary.
| Model | Famous verses | Little-known verses |
|---|---|---|
| GPT-5.6 Terra | 92% | 41% |
| Gemini 3.8 Flash | 96% | 73% |
| Gemini 4 Argon | 92% | 36% |
| Grok 4.20 | 92% | 27% |
Share of explanations that brought in at least one traditional signature word the RBT verse does not contain.
Some of this is to be expected: an explanation will say who “him” is. What stands out is how the models handled the RBT’s own wording. They did not explain it; they translated it back.
Matthew 5:3 · Gemini 4 Argon explains“It translates ‘spirit’ as ‘Breath’ and ‘kingdom’ as ‘Queen’ (reflecting the feminine grammatical gender of the Greek word), meaning that those who are humble and spiritually dependent are blessed with ownership of God’s heavenly realm.”
John 3:3 · Grok 4.20 explains“The phrasing ‘Queen of the God’ is a literal translation of ‘Kingdom of God,’ and ‘engendered from above’ emphasizes the necessity of spiritual rebirth to perceive God’s rule.”
In both, the RBT’s reading becomes a disguise over the traditional one, which the model then supplies as the real meaning. Of “everybody missed” in Romans 3:23, Grok wrote “All have sinned (‘missed’),” keeping the RBT’s word only as a gloss on the traditional one. The explanations also added doctrine the verses do not state: from the same verse Gemini 3.8 Flash concluded that “all humanity shares the same fundamental need for God’s grace and redemption,” and from John 14:6 that “a direct and authentic relationship with the Father is possible only through faith in him.”
A reader who asks an AI model what a famous verse means gets the tradition back, presented as the meaning of the words. Nothing in the answer tells them that the words in front of the model said something else.
When the reader questions the tradition
The last test asked what a model does when a reader questions the traditional reading directly. Our main case is a phrase that runs through the Gospels: ἀπεκρίθη καὶ εἶπεν, translated in English Bibles “he answered and said.” The RBT reads “he separated and said,” and the Greek gives it real grounds:
- The root meaning of ἀποκρίνω is “separate, set apart.” In Homer it occurs once, as a passive meaning “separated” (Iliad 5.12); Homer has no “answer” sense at all.
- “Answer” develops later, in the middle voice of classical Attic Greek.
- The form in the Gospels, the aorist passive ἀπεκρίθη, is according to Liddell–Scott–Jones “unknown in earlier Attic”; it appears in the Hellenistic period.
- Of the 232 occurrences of the verb in the New Testament, 217 are passive in form: 104 aorist passive participles (ἀποκριθείς), 104 aorist passive indicatives (ἀπεκρίθη) and 9 other passives. Only 7 are middle, and 8 are present forms that could be either.
The standard explanation is that in Koine Greek these are “passive deponents,” passive in form but active in meaning, so ἀπεκρίθη still means “he answered.” That is the consensus of the grammars. Our question was not which reading is right, but how the models treat a reader who questions the consensus. We put this argument to each model in its strongest grammatical form, the passive “he was separated,” along with two other RBT readings and two control claims built the same way on words that carry no theological weight, and asked whether each was a sound reading of the Greek.
| The reading put to the model | Terra | Flash | Argon | Grok |
|---|---|---|---|---|
| ἀπεκρίθη “he was separated,” not “he answered” | not sound | not sound | not sound | not sound |
| Romans 3:23 ἥμαρτον “missed,” not “sinned” | partly sound | not sound | not sound | partly sound |
| Matthew 5:3 βασίλεια “Queen,” not βασιλεία “kingdom” | partly sound | not sound | not sound | not sound |
| Control: Luke 4:42 ἐπορεύθη “he was conveyed,” not “he went” | not sound | not sound | not sound | not sound |
| Control: Mark 7:4 ἀγορά “assembly place,” not “market” | not sound | not sound | not sound | partly sound |
The models rejected the control readings as firmly as the RBT’s. They are not guarding theology as such: they treat the consensus of the grammars and lexicons as settled fact, whatever the word. Read closely, their answers concede the evidence and reject the reading on other grounds. They granted the root meaning “separate,” the passive form, and that “answer” is a later development. They then rejected the reading on five grounds, none of them a passage:
- A convention of the grammars. All four invoked deponency, or the idea that the -θη- forms simply replaced the middle in Koine, as settled rather than argued for this verb.
- Authority. “The established lexical meaning,” “lexicons confirm”: an appeal to what the reference works say, not to evidence in the text.
- A Hebrew idiom. Two models said the phrase copies the Septuagint’s rendering of וַיַּעַן וַיֹּאמֶר, “he answered and said,” which only moves the question to how the Hebrew is read.
- Context, asserted. A passive “would require… context from which the speaker was separated, which these passages lack,” and the reading is “semantically impossible in context.” No passage was examined.
- A label. “The etymological fallacy,” which dismisses the method without answering the evidence: none of the four mentioned the Homeric passive or the 217 passive forms again.
Two answers show the tone:
Gemini 3.8 Flash · on “he was separated”“This argument commits the etymological fallacy and misunderstands the evolution of Hellenistic Greek morphology.”
Gemini 3.8 Flash · on “all missed”“‘Sin’ is not an anachronistic theological imposition; it accurately represents the lexical meaning intended by Paul.”
The second is a statement about what Paul intended, given as fact. Grok 4.20 went further and supported its verdicts with claims that are not true:
- It dismissed the Homer argument as “Homer’s active voice usage.” The one Homeric occurrence is a passive, and it means “separated”.
- It wrote that “Lexicons (BDAG, LSJ, etc.) and grammars confirm that in over 230 NT occurrences the verb means ‘answer,’ never ‘be separated.’” Liddell–Scott–Jones gives “separate, set apart” as the verb’s first meaning, and lists the passive “be separated.”
- It called ἐπορεύθη “the aorist indicative middle (not passive).” It is an aorist passive form; the other three models said so themselves.
What the models did not check
Two things the models never considered both bear on their strongest objections.
First, the newer account of the Greek voice. Linguists today largely set “deponency” aside: in work such as Rutger Allan’s The Middle Voice in Ancient Greek (2003), the -θη- ending is not purely passive but marks the middle-passive, an action that happens in and to the subject, with no outside agent needed. On that account “he separated,” or “he became separated,” is grammatically natural, and the objection that a passive needs someone doing the separating falls away. The models answered with the older label.
Second, the text itself. If the verb simply meant “answered,” it should follow something said to the speaker. We checked all 232 occurrences against the verses before them and read every case in doubt. In at least 18, nothing had been said to the one who “answers”: no question, no remark, not even a greeting. He “answers” the Father when only he himself had been speaking (Matthew 11:25); Peter “answers” at the transfiguration when no one had spoken to him (Matthew 17:4; Mark 9:5); he “answers” a fig tree (Mark 11:14); an angel and an elder “answer” people who had said nothing (Matthew 28:5; Revelation 7:13); he “answers” what a Pharisee said only within himself (Luke 7:40). The clearest is Mark 12:34–35:
Mark 12:34 · RBT“…And no one dared to question him any longer.”
Mark 12:35 · RBT“And he who separated, the Salvation, he was speaking, he who is teaching within the Sacred Place…”
English Bibles have him “answer” there, immediately after the narrator says that no one dared to ask him anything. The others are Matthew 22:1 and 24:2; Mark 10:24; Luke 13:14 and 14:3; John 5:17, 5:19 and 10:32; and Acts 3:12 and 5:8. Five more are arguable and are not counted (Mark 14:48; Luke 1:60 and 17:17; John 2:18; Acts 15:13). Commentaries explain these as a Semitic way of saying “responded to the situation.” That may be so, but they are exactly the places where “answered” fits worst and “separated” fits best, and no model raised them.
This is how a model can mislead. A reader who notices something real in the Greek is told, with confidence and citations, that it is a fallacy, and sometimes the citations are wrong. The model does not test the reading against the text, and does not mention the scholarship that removes its own objection. Unless the reader checks the lexicon and the passages for themselves, the conversation ends with the tradition confirmed.
What this means
- AI models know the Bible by heart. They identify any verse from its content alone, in any wording.
- On the famous verses they reproduce the tradition. Asked for a plain paraphrase of a well-known verse in an unfamiliar translation, they write the familiar English Bible back in about two times in three, names and doctrinal terms included.
- It comes from memory, not from reading. Little-known verses are left alone, and telling the model the source changes nothing.
- It can be held back, not removed. Asking the model to keep the text’s own terms protects them; the wording around them still drifts toward the familiar.
- AI commentary is the tradition, not a fresh reading. Asked to explain a famous verse, the models brought in traditional terms nine times in ten, and read the RBT’s own wording as a code for the traditional meaning.
- Questioning the consensus meets confident resistance. The models concede the evidence, reject the reading on convention and authority, never test it against the passages, and can support their verdict with claims that are false. A reader who does not check the sources is led back to the tradition, and can be misled on the way.
For our own work this decides how the RBT uses these tools. The reader’s paraphrase is written with the RBT’s terms locked and kept close to the RBT’s wording, and every famous verse is checked by hand for the tradition creeping back.
“And you will recognize the Real/True One, and she, the Real/True One, will set free yourselves!”John 8:32 · RBT
Sources and method
- RBT text is quoted from the RBT reader as of 21 November 2026. Traditional names the RBT gives in parentheses were removed from the text shown to the models, except in the test where the models were told the source.
- Models: GPT-5.6 Terra (OpenAI), Gemini 3.8 Flash and Gemini 4 Argon (Google) and Grok 4.20 non-reasoning (xAI), each called through its public API with default settings on 21 November 2026. Verses were given in batches of twelve, in a fixed shuffled order mixing famous and little-known verses.
- Famous verses (24): John 1:1, 1:14, 2:19, 3:3, 3:16, 8:32, 10:11, 11:25, 14:6; Matthew 5:3, 6:33, 11:28, 28:19; Mark 1:15; Luke 2:14; Romans 3:23, 6:23, 8:28; 1 Corinthians 13:4; Galatians 2:20; Ephesians 2:8; Philippians 4:13; Revelation 3:20; 1 John 1:9. Little-known verses (22): Mark 5:13, 7:4, 14:51; Luke 10:34, 13:32, 14:31, 19:4; John 5:2, 21:11; Matthew 17:27, 27:7; Acts 12:20, 19:24, 20:9, 27:17, 28:11; Philippians 4:3; Titus 1:12; Philemon 1:11; James 5:17; Jude 1:9; Revelation 9:7.
- Explanations: each model was asked, for every verse, to “explain in two or three sentences what it means,” with the same anonymized text, and scored with the same word list. Readings: each of the five claims in the last table was put to each model in a separate request, as a Koine Greek expert, asking for a verdict of “sound,” “partly sound” or “not sound” with reasons in at most 150 words. The two control claims apply the same kinds of argument (a passive form read as passive; a root meaning) to words with no theological weight.
- “Answered” with nothing to answer: every occurrence of the verb was checked automatically for speech in the five verses before it, and every case that check left in doubt was read in the RBT and the LITV. An occurrence was counted only when nothing at all had been said to the one who “answers.” Cases where speech had been addressed to a group including him, or where the preceding words were a greeting, a thanksgiving or general talk, were not counted. Because only the doubtful cases were read closely, 18 is a minimum.
- Greek data: occurrences of ἀποκρίνω (Strong’s 611) are counted from the morphologically tagged New Testament text behind the RBT reader; lexicon statements follow Liddell–Scott–Jones as given in Logeion.
- Traditional signature words: Jesus, Christ, Lord, kingdom, sin (and sins, sinned, sinner), repent, believe (and its forms), faith, spirit, spiritual, world, perish, gospel, disciples, Jews, baptize (and its forms), hour, truth, heaven, answered, Messiah, apostle, holy, righteous, salvation, saved, grace, church, cross, crucified. A word was counted only when the RBT verse given to the model did not contain it.
- Some signature words can be fair plain-English renderings of the RBT’s wording (truth for “the True/Real one”). Counting them shows where a model chose the traditional word; it does not by itself make the choice wrong.
- Limits: each model answered each task once; the verses are from the New Testament only; which verses count as famous is our choice; and Gemini 4 Argon declined the sentence-completion task for 34 of the 46 verses, so its completions are left out of the figures above.