The plan behind the story · Satoshi Tomokiyo's list of unbroken ciphers is getting shorter, and AI is the reason
Four weeks of codebreaking: by hand first, then with AI you check
A month that takes you from your first shifted-alphabet message to checking an AI's solution the way Satoshi Tomokiyo checked Napoleon's letter. New to this: you finish able to break a simple substitution cipher by hand. Already doing it: you finish with your own record of where AI gets codes right and wrong.
Why this works
When Carter Church broke a coded letter from 1809 with an AI model, AFP reported, it took about six hours. The answer still needed a person who knew the field to confirm it, and that person was Satoshi Tomokiyo, who keeps a list of unsolved historical ciphers on his website, Cryptiana.
This plan gives you both halves. You learn the oldest method by hand, so you understand what a correct answer looks like. Then you use a free AI assistant and judge its work. Ciphers are good practice for that because an answer can be tested: the decoded text either reads properly or it does not, and re-encrypting it with the key must give back the original exactly.
Two habits run through every week:
- Check what it tells you. AI can produce confident nonsense. The Notre-Dame transcription system AFP described reads 88 to 94% of characters correctly, which means some are wrong on every page.
- Keep private details out. Never paste a real password, PIN, bank code or anyone’s personal letter into an AI tool to “test” it. Use made-up messages and historical documents only.
Before you start
- Paper, pencil and squared paper. Codebreaking is easier when you can lay letters out in a grid.
- One free AI assistant. Any of the free ones will do; Your first conversation with an AI assistant covers how to start.
- A partner, if you can. Swapping ciphers with a friend or grandchild doubles the fun and gives you messages you did not write.
- A notebook page for results: date, cipher, solved by (me / AI), time taken, letters wrong.
The plan
| Week | Do this | How often | Notes |
|---|---|---|---|
| 1 (new to this) | Shift ciphers. Write a sentence with every letter moved three places along (A to D). Swap with a partner and crack theirs by trying each shift in turn. | 4 days, 15 minutes | There are only 25 possible shifts, so this always works with patience. |
| 1 (already doing it) | Write three 200-letter substitution ciphers (each letter swapped for a different one). Solve a partner’s by hand and time it. | 3 ciphers | Your hand-solve time is the benchmark for week 3. |
| 2 (new to this) | Frequency counts. Count the letters in a newspaper paragraph, then use the counts to break a short substitution cipher from a partner. | 4 days, 20 minutes | In English, E, T and A are usually the most common letters. Short messages may not follow the rule. |
| 2 (already doing it) | Try a harder kind: a cipher where common letters have several symbols each (homophonic), so frequency counts flatten out. | 2 or 3 sessions | Note what you had to do differently. |
| 3 (new to this) | Give the AI a shift cipher you made. Check its answer letter by letter against your original. | 3 times | Record every wrong letter, however small. |
| 3 (already doing it) | Give the AI fresh substitution ciphers. For each, re-encrypt its answer with the key it gives and compare with the original. | 5 ciphers | Accept only an exact match. Compare time and errors with week 1. |
| 4 (both) | Read about one real historical cipher (Klaus Schmeh’s 2020 post on Tomokiyo’s list mentions several, including letters of Charles I from 1648) and find out who solved it, how and when. | Once or twice | Write down where you found each fact, and check it in a second source. |
If it goes wrong
- You cannot crack it. Ask your partner for one word of the answer. Codebreakers call a known word like that a “crib”, and one is often enough to unpick the rest.
- The AI gives a fluent answer that is wrong. Good: that is the lesson. Mark it in your notebook and note what kind of cipher it was.
- The AI asks for more than the cipher. You never need to give it your name, address or anything personal to solve a puzzle.
- It stops being fun. Go back to week 1 with a new partner. Writing ciphers for someone else is often more enjoyable than breaking them.
Track it
Each week, add up: ciphers solved by hand, ciphers solved by AI, and letters the AI got wrong. By the end of week 4 you should have a clear sentence of your own: “AI is reliable for ___, and I always check ___.”
What next
- Read our six-step routine for checking what a chatbot tells you and use it on something that matters to you, such as an old family letter or a form.
- Keep the daily habit going with our daily puzzles.
- If you enjoy the history, visit a codebreaking museum or collection near you, and check opening times on its own website before you go.
General information, not medical advice. Check with your GP before starting a new exercise programme.