When you search for information at work, dates help you decide which documents deserve your attention. A contract from the wrong year or an outdated project report might seem irrelevant, even when it contains something that could answer your question.
A reported experiment with GPT-6 Astra suggests why an AI research assistant might benefit from questioning those assumptions. According to a September 17 post from Prinz AI, the model recovered the contents of a German military radio message sent on November 27, 1918, with an encryption key associated with December.
The result offers a case study in how AI could connect information across documents, then look for separate evidence to check its answer.
How the Germans concealed the message
The message reportedly contains 170 encrypted characters and uses ADFGVX, a system the German military used during World War I to conceal radio communications.
Think of it as a two-step scrambling process. First, the sender replaced each letter or number with a pair of symbols. Then, a keyword determined how those symbols were rearranged before transmission. The recipient needed the correct instructions to reverse the process and read the message.
Researchers have already recovered hundreds of messages from this collection. In a 2016 paper, George Lasry and his colleagues described decoding 618 of 668 encrypted texts preserved by American intelligence officer J. Rives Childs.
Prinz’s experiment began with messages listed as unresolved in Klaus Schmeh’s historical cipher collection. The reported AI recovery builds on years of human research and preserved records.
Why the date on the key matters
According to Prinz, Astra used the keyword TRUPPENVERSCHIEBUNG, which appears in Childs’s historical reference on German military ciphers. The account associates that key with use beginning December 9, 1918, almost two weeks after the message was transmitted.
For a researcher, the mismatch could make the key seem unlikely to work. If your records say a password became active in December, you would have a reason to question whether it could unlock something encrypted in November.
Astra reportedly tested the key anyway, but the explanation for why earlier researchers missed the message remains a hypothesis. We do not know which keys they tried, and the account does not establish why the dates differ.
For professionals who use AI to investigate a problem, the potential application is straightforward: an assistant could examine possibilities that initially appear unlikely. Each possibility would still need a test before anyone could accept the answer.
What the recovered message appears to say
The proposed German text published by Prinz reads:
EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X
In English, the proposed reading is:
An English cruiser arrived at Sevastopol on the ?4th. An Allied squadron follows on the 26th.
The question mark reflects uncertainty in the recovered first date. Although the message appears to describe naval arrivals, the text alone does not clearly establish whether the first arrival happened on the 24th.
The surviving log of HMS Canterbury, a British cruiser, helps resolve that uncertainty: it records the ship reaching Sevastopol harbour on November 24, 1918, followed by an Allied squadron’s arrival on November 26.
Both entries correspond with the proposed reading. They also establish the correct year as 1918; the Prinz post’s reference to Canterbury arriving in “2018” is a typo.
Why a believable answer still needs checking
Two different checks help assess this result, and each answers a different question. The first concerns the decoding itself: can another researcher apply the same instructions to the encrypted message and reproduce the proposed text?
The second concerns the historical facts: do separate records support what the message appears to describe? Canterbury’s log provides evidence for the naval movements, although it cannot prove which encryption key the German sender used.
The distinction resembles checking a financial calculation against the underlying transactions. The arithmetic needs to work, while the records need to support the explanation.
The claim of a discovery also requires care. Prinz says they are unaware of an earlier solution, which leaves open the possibility that someone previously recovered the message.
This investigation is separate from the 1941 MVUEH Enigma recovery. Each result needs its own evidence.
The experiment suggests a way to judge AI research assistants in professional work. An assistant should explain which assumption it questioned and show how it checked the resulting answer. When colleagues can repeat those checks, they have a basis for deciding whether to act on the finding.
