The Day Turing Never Lived to See: AI Completes His Unfinished WWII Code-Breaking Mission

Imagine this scene: 1944, Bletchley Park. Cryptanalysts work around the clock staring at intercepted German telegrams, knowing the messages contain intelligence that could turn the tide of the war — yet, constrained by the limits of computational power, they watch helplessly as a portion of the ciphertext remains forever locked in mystery.

Over eighty years later, that very same ciphertext was cracked by two language models in a matter of hours.

Not That Turing Test — The Other One

Most people know the "Turing Test" as a measure of whether a machine can converse like a human. But Turing left behind more than one unsolved puzzle. TechCrunch reported on September 25 that Google DeepMind's Astra and Anthropic's Claude Opus — two frontier models — have just completed another of Turing's unfinished endeavors: breaking the Lorenz SZ cipher system, the very system Nazi Germany used to encrypt communications at its highest command levels during World War II.

This was no textbook reconstruction. The research team worked with genuine historical intercepted ciphertext, not practice problems with known answers. Both Astra and Opus, without prior knowledge of the keys, recovered readable German from the ciphertext through statistical analysis and pattern recognition.

In plain terms: these two models accomplished what the Bombe and the Colossus computer never fully managed to do.

Why Lorenz Was Harder Than Enigma

Most people are more familiar with Enigma, thanks to the film The Imitation Game. But Lorenz was significantly more difficult. Enigma required operators to manually type and encrypt messages letter by letter; Lorenz was fully automated teletype, with a far more complex key-wheel system and a combination space orders of magnitude larger.

Turing and his colleagues did crack Enigma, but the complete decryption of Lorenz was fragmentary. Many intercepted messages went unrestored for decades after the war. These "cold case ciphertexts" sat quietly in the UK National Archives.

The batches Astra and Opus worked with came from exactly that collection.

How the Models Actually Broke It

There's a detail here worth pausing on. Breaking Lorenz is not a matter of brute-force enumeration — the combination space is simply too vast to exhaust. What Colossus could do was use statistical properties to narrow the search space, but that still required human analysts to make a great deal of manual judgments.

What Astra and Opus did was, in essence, apply both "statistical analysis" and "language comprehension" simultaneously within the same model. They weren't merely finding patterns — they were concurrently verifying whether the recovered German text made semantic sense. That dual-track validation is something pure statistical tools cannot do.

Think of it this way: the models were simultaneously playing the roles of cryptomathematician and linguist, each role correcting the other. This is precisely why frontier language models suddenly have an advantage in this task — not because they have more compute, but because they arrive with linguistic knowledge already embedded.

What Would Turing Think, If He Were Alive Today

This question has been turning over in my mind for a while.

Turing was the kind of person who was captivated by the question of where the limits of computation lie. The deeper motivation behind the Turing Test was not to prove that machines could fool humans — it was to force everyone to confront a question: if a machine's behavior is indistinguishable from intelligence, on what grounds do we say it has none?

Now Astra and Opus are not merely imitating language. They are solving a historical problem that Turing himself never fully resolved. The reversal is historically striking — the student's tools completing the homework the master never finished.

My guess is that Turing would not feel supplanted. He would probably immediately ask the next question: "All right — beyond Lorenz, what other cipher systems haven't we broken yet?"

What This Means for AI Development Today

There are several dimensions worth unpacking:

• A milestone in reasoning capability: Breaking Lorenz requires not memorization, but the ability to converge on correct answers under conditions of uncertainty. This kind of task was once considered to require "human intuition." Models can now do it.
• New uses for historical data: Vast archives of dusty historical records have suddenly become resources that AI can mine anew. Not just ciphers — the same logic applies to archaeology, legal documents, and ancient scripts.
• The evolution of evaluation benchmarks: This also suggests that AI assessment cannot rely solely on language comprehension scores or math benchmarks. We need touchstones closer to genuinely complex real-world tasks.

As a side note, if you're interested in how models handle domain-specific knowledge, the difference between fine-tuning and RAG is another topic worth getting clear on — especially when you're thinking through architectural questions like "should I have the model memorize a domain's knowledge, or have it retrieve?"

The Most Historically Resonant Line in All of This

This story brings to mind a striking contrast: when Turing worked at Bletchley Park, it was to let the Allies read Hitler's orders one step ahead; now, what reads those messages is a language model with no "purpose," no "fear."

It has no knowledge of the weight of the war, no awareness of how many people died behind those ciphertexts — and yet it produced the answers.

That image leaves you somewhat speechless. Not because AI has become something magnificent, but because the arc of history sometimes bends back on itself in places you never expected.

If you'd like a more systematic look at where frontier models' capability boundaries currently stand, the comparison between ChatGPT and Gemini is a good entry point — the gap in design philosophy between the two companies actually maps onto something in how Astra and Opus each approached the decryption task.

---

One thought to take with you: what Turing left behind was not only tests and algorithms — he left behind a set of problems not yet solved. AI is now taking those problems over, one by one. Not to replace him, but to keep walking the same line forward.

References

• Astra and Opus just passed Turing's other test — TechCrunch (2026-09-25)