AI can help solve a historical cipher when the cipher leaves patterns that software can exploit and there is enough reliable text to test a solution. It is much less dependable when the method is complex, the message is short or damaged, or the language and historical context are uncertain. A fluent plaintext is only a candidate: a convincing decipherment should also fit the cipher’s mechanics and withstand checks against independent historical evidence.
What makes one cipher easier for AI than another?
“AI” is not a single cipher-breaking method. Depending on the task, software may help transcribe a manuscript, identify a likely cipher family, search possible keys, or score candidate plaintexts for how well they fit a language. Those are separate jobs; generating plausible prose is not the same as recovering a cipher’s method and key.
As an Amazon Associate I earn from qualifying purchases.
Many successful approaches exploit structure. In a simple monoalphabetic substitution, each ciphertext symbol consistently represents one plaintext letter. Repeated symbols and recurring sequences therefore constrain possible readings. A Transformer-based language model described by Kambhatla, Born and Sarkar learns symbol recurrences and performed strongly on synthetic 1:1 and homophonic substitution ciphers; the authors also reported successful solutions for several real historical homophonic ciphers. These findings support those cipher families and experiments, not a claim that AI can break every cipher. Read the 2023 ACL paper.
A homophonic substitution may use several different ciphertext symbols for the same plaintext letter, making symbol-frequency clues less direct than in a 1:1 substitution. It still has exploitable structure, but a method must account for the cipher’s rules. Other systems preserve different patterns or require a much broader search, so success on substitution does not transfer automatically.
The cipher family sets the problem
A 2018 HistoCrypt paper by Nils Kopal contrasts monoalphabetic substitution, which can be solved by hand, with Vigenère, which takes more work, and Enigma, which is nearly impossible to solve by hand. The distinction is practical: different systems demand different analyses, and machine ciphers require modeling the machine and searching settings rather than simply matching recurring letters. Kopal also presents CrypTool 2 as software for automating analysis of classical and modern ciphers. Read the HistoCrypt paper.
Why language and historical period matter
Many decipherment systems rank possible readings by how plausible they are in a target language. That makes language knowledge part of the attack. An old text may use spellings, vocabulary, grammar, or conventions that a model trained on modern writing does not expect. A candidate can be historically plausible but score poorly under modern-language assumptions—or sound smooth while being wrong.
Experiments by Megyesi and co-authors compared modern-text and historical-text language models on English and German homophonic substitution ciphers. They reported that historical models significantly improved performance on ciphertext produced in the 17th century or earlier; century-specific models helped more with longer and older samples. This is a result for the studied languages, cipher family, and experimental conditions, not a guarantee for every archive. Read the study.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteHow much text—and how clean a transcription—does AI need?
More ciphertext generally offers more repeated patterns and more evidence for comparing candidate readings. Short messages leave fewer clues, while transcription errors can make real patterns harder to detect. Missing spaces add another uncertainty because the system must infer word boundaries as well as letters.
Rank #3
Aldarrab and May studied a multilingual sequence-to-sequence system for 1:1 substitution. They tested 14 languages under conditions including different ciphertext lengths, missing spaces, and transcription noise, then applied the system to the historical Borg cipher using its first 256 characters. Those details describe that study and case; they do not establish a minimum length that applies to other ciphers. Read the paper.
What historical cipher examples show—and what they do not
The 1809 letter: context can check a candidate
Live Science reported on 2 October 2026 that AI engineer Carter Church used an AI-assisted workflow to decipher an 1809 letter associated with Eugène de Beauharnais and Marshal Marmont. The report says the system found and applied a partially identified cipher table. Historian Michael Rowe said related correspondence made the proposed contents easier to check because much of the message was already expected. He compared the material to “a kind of Rosetta Stone.” The case illustrates how partial cipher information and historical context can help validate a reading; it should not be described as a wholly blind discovery proved by a model alone. Read Live Science’s report.
The 82-letter Enigma message: a reported candidate, not independent verification
Tom’s Hardware reported on 26 September 2026 that an AI system selected an 82-letter MVUEH Enigma message, developed simulator and search tools, and produced a candidate plaintext. The article renders it as “BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH” and gives an approximate English translation, while noting an apparent spelling error. The report is secondary coverage; the material cited here does not independently verify the key or plaintext. Enigma is a reminder that a language model’s output is not enough: a strong solution needs a reproducible machine model and settings that generate the proposed text. Read Tom’s Hardware’s report.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
How to judge a claimed AI decipherment
When someone says AI “cracked” a historical cipher, ask whether the result is a mechanically reproducible solution or merely a plausible interpretation. A credible claim should make clear what method was assumed, how a key or settings were recovered, and how well the resulting plaintext fits the actual ciphertext.
- Cipher family: Does the proposed method match the system’s rules and the patterns it preserves?
- Text quality and amount: Is the message long enough to constrain a reading, and could damaged writing or transcription mistakes explain anomalies?
- Language assumptions: Is the target language known, and does the scoring account for historical spelling and usage?
- External clues: Were a partial key, crib phrase, cipher table, or related correspondence available? These can strengthen a check while making a supposedly blind discovery less surprising.
- Reproducibility and validation: Can the method and key or machine settings reproduce the ciphertext-to-plaintext match, and does independent historical evidence support the interpretation?
There is no single cross-family benchmark in the cited studies that ranks all AI methods. Results from substitution experiments cannot establish how well AI handles Enigma, an unknown language, or modern encryption. Historical classical-cipher work should not be treated as evidence that AI can break modern encryption systems.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
