Bible Code — Equidistant Letter Sequences
Not a cipher — a lesson in probability. The same method finds the same "prophecies" in Moby-Dick.
Why This Exhibit Exists
The Bible Code is the museum's only exhibit on a claim that is not a real cipher. It is here because the tools used to "find" it are the same statistical tools used to break real ciphers — and misunderstanding those tools has caused enormous harm. The 1994 paper appeared in a peer-reviewed journal. Drosnin's 1997 bestseller sold millions of copies. The statistical error it embodies — confusing "pattern found" with "pattern is significant" — remains one of the most common mistakes in security and intelligence analysis today.
In 1994 Israeli mathematicians Doron Witztum, Eliyahu Rips, and Yoav Rosenberg published a paper in Statistical Science claiming to find statistically significant equidistant letter sequences (ELS) in the Hebrew text of Genesis. They claimed that the names of famous medieval rabbis appeared encoded in the text at unusually short skip intervals, clustered near their dates of birth and death.
The paper caused a sensation. In 1997 journalist Michael Drosnin published The Bible Code, extending the method to predict modern events — the assassination of Yitzhak Rabin (cited after the fact), the Gulf War, and others. Drosnin's challenge: "When my critics find a message about the assassination of a prime minister encrypted in Moby-Dick, I'll believe them."
In 1997, mathematician Brendan McKay accepted the challenge. He found messages in the English text of Moby-Dick predicting the assassinations of Indira Gandhi, Yitzhak Rabin, John F. Kennedy, Robert F. Kennedy, Martin Luther King, Abraham Lincoln, and Leon Trotsky — all using Drosnin's exact method. The findings appeared in Statistical Science in 1999, co-authored with Dror Bar-Natan, Maya Bar-Hillel, and Gil Kalai. The original Witztum-Rips experiment was independently shown to depend entirely on a specific controversial list of rabbis; when an independent list was substituted, the effect disappeared entirely.
Take any text. Remove all spaces, punctuation, and numbers, keeping only letters. This gives you a letter string. Now pick a target word and a skip value d. Starting at position p, read every d-th letter: positions p, p+d, p+2d, and so on. If the result spells your target word, you have found an ELS at (start=p, skip=d).
Example: text = THEQUICKBROWNFOX Word = "TOX", skip = 4 Position 0 → T Position 4 → U ✗ (not O) Word = "TRX", skip = 4 Position 0 → T Position 4 → U ✗ Try skip = 5: Position 0 → T Position 5 → I ✗ Finding real words requires scanning ALL (start, skip) pairs. In a 100,000-letter corpus with skips up to 1,000, that is 100,000 × 1,000 = 100,000,000 pairs per target word. Any 4-letter word with average letter frequencies is expected to appear hundreds to thousands of times.
The key statistical fact: for a 100,000-character text searched with skips up to 1,000, a 4-letter word like EDEN or FIRE has an expected hit count of roughly 400–2,000 depending on its letter frequencies. Finding hundreds of hits is not evidence of hidden messages — it is arithmetic.
The two boxes below search the same word in two texts of identical length: a passage of English prose and a randomly generated text with the same letter frequencies. The hit counts are statistically indistinguishable. Finding a word in a long text by ELS is not evidence of intent — it is the expected consequence of searching enough (start, skip) pairs.
Try target words: EDEN, DEATH, WAR, KING, ASSASSIN.
Every one appears hundreds of times in any long text. The "Bible Code" method gives you no way to distinguish these hits from noise.
McKay's team applied Drosnin's own method — identical parameters, identical search procedure — to Hebrew translations of War and Peace and to the English text of Moby-Dick. Both texts produced equally "remarkable" clusters of famous names near death dates. The original Witztum-Rips result also proved sensitive to which list of rabbis was used: the specific list in the 1994 paper was a version that had been informally optimised for the experiment over several years. When an independent scholar's list was substituted, the result vanished completely.
A 100,000-letter text, words of length 3–6, skips up to 1,000, forward and backward: the total search space exceeds 109 (start, skip, direction, word) combinations. If any combination counts as a "hit," and you search widely enough, hits are certain. The post-hoc selection of "remarkable" hits from a universe of a billion candidates is not statistical significance — it is the Texas Sharpshooter Fallacy at scale. Modern cryptanalysis requires pre-specifying the hypothesis before searching.
| Mistake in the Bible Code claim | Modern cryptographic parallel |
|---|---|
| Searching billions of (start, skip) pairs post-hoc and reporting only the hits | Statistical significance in security audits: running 1,000 statistical tests and reporting the 50 that pass at p<0.05 produces 50 false positives by construction — the same error that plagued early side-channel research |
| Treating a pattern found in scripture as non-accidental without comparing to a random baseline | Distinguisher attacks: a valid linear or differential distinguisher requires showing the bias in the cipher is greater than expected from a random permutation, not merely that a bias exists |
| The "remarkable" result depended on a specific optimised dataset the authors had adjusted over years | Chosen-input bias: constructing a test set to fit a desired result is the practitioner error behind broken PRNG certification tests — e.g. the Dual EC DRBG backdoor design |
| Short ciphertext + flexible parameters = near-certain matching | Over-parameterised models find anything: a cipher analysis with enough free parameters can "fit" any plaintext hypothesis to any ciphertext — the same reason differential cryptanalysis requires strict key and round constraints |
| Peer review passed the original claim; independent replication killed it | Security proofs require adversarial review: the NSA's BSAFE RNG and WEP both passed internal review; independent public cryptanalysis found flaws within months. The parallel refutation is why open cryptographic standards survive and secret ones fail |
On the original 1994 paper. Witztum, Rips, and Rosenberg were not fraudsters — they were mathematicians who made a genuine statistical error of a kind that was not well-understood in 1994. The journal editors published the paper as a challenge to the community to find the flaw. The community found it. That process — publication, replication, refutation — is exactly how science and cryptanalysis are supposed to work. The exhibit is critical of the claim, not of the people who made it.
- Voynich Manuscript — genuinely unsolved; statistical analysis has proved its letter distributions are unlike any known cipher or natural language
- Beale Ciphers — another case where "finding" a partial decode via book cipher may be pattern-matching rather than cryptanalysis
- Kryptos — four panels, three solved; the fourth is a genuine open problem, not a statistical artefact
- Caesar Cipher — chi-squared scoring is the correct way to distinguish signal from noise in a short substitution cipher
| Exhibit | 140 of 140 |
| Claimed | 1994 (Witztum et al.); 1997 (Drosnin) |
| Refuted | 1999 (McKay et al., Statistical Science) |
| Method | Equidistant Letter Sequences (ELS) |
| Error type | Multiple comparisons / Texas Sharpshooter Fallacy |
| Hall | VIII — Puzzle & Novelty Ciphers |
For a 78,000-letter Hebrew text (the Torah), searching words of length 5 at skips up to 1,000:
Before asking whether a hit is "significant," you must ask: how many hits were expected? In a 780M-pair search, finding 50 hits for a specific name is exactly what chance predicts.
- The claim: Witztum, Rips & Rosenberg — "Equidistant Letter Sequences in the Book of Genesis." Statistical Science 9:3 (1994), pp. 429–438.
- The popular version: Michael Drosnin — The Bible Code. Simon & Schuster, 1997. (This exhibit is a direct response to this book's methodology and conclusions.)
- The refutation: McKay, Bar-Natan, Bar-Hillel & Kalai — "Solving the Bible Code Puzzle." Statistical Science 14:2 (1999), pp. 150–173. Freely available online.
- Moby-Dick experiment: Brendan McKay's personal site preserves the analysis: users.cecs.anu.edu.au/~bdm/dilugim/moby.html