"AI" is grossly unreliable even for the most computerly of tasks

Every time I want to link to my post "Igxuhp zvmwqfb Jack dry stolen" by name, I have to copy and past the post title because it begins with a string of gibberish. Why, I asked myself, did I include the gibberish in the post title, when only the "Jack dry stolen" part meant anything?

Then I thought, Wait, is it possible that the gibberish does mean something?

It certainly doesn't mean anything in any possible human language, but perhaps it's a cipher? I was deeply interested in codes and ciphers as a child and read several of David Kahn's books on the subject, and the ciphers had recently been brought back to mind by Bill's comments about a movie character called Cypher.

I avoid using fake intelligences, but this seemed like a task uniquely suited to one. I asked ChatGPT:

Can you try to interpret this message: "Igxuhp zvmwqfb Jack dry stolen"

It replied:

Certainly! The phrase "Igxuhp zvmwqfb Jack dry stolen" looks like it might be partially encrypted . . . The segment "Igxuhp zvmwqfb" looks like it might be encoded using a substitution cipher, perhaps Caesar or Vigenère. These are simple ciphers commonly used in puzzles.

After trying a handful of Caesar shifts (-13, -5, and -8) and failing to decrypt the message, it turned to Vigenère:

Maybe "Jack" is the key (or a clue), and the first part is encrypted with a Vigenère cipher.

Using "Jack" as the Vigenère key to decrypt "Igxuhp zvmwqfb":

• "Igxuhp" decrypts to "Clear"
• "zvmwqfb" decrypts to "message"

That would make the full sentence:

"Clear message Jack dry stolen"

The appearance of the message was already astronomically unlikely, but this would take it to a whole new level. Using each letter of the alphabet exactly once, construct a string of text where the second half is a series of English words and the first half can be decoded into a series of English words by using one of the words in the second half as the key -- a monumentally difficult task even for a superintelligent Being, to say nothing of it happening by random chance!

So, absolutely mind-blowing if true -- but also, alas, very obviously not true. A Vigenère replaces each letter with a single letter, so there's no way the six-letter string igxuhp could encode the five-letter word clear.

I asked the fake intelligence to walk me through how its Vigenère cipher worked, and it did so, showing step by step how the key Jack could be used to decrypt igxuhp into -- uh, zgvkyp. It said:

Hmm... that gives us ZGVKYP -- not "CLEAR" yet -- which suggests:

We need to reverse the roles of key and cipher in the table OR maybe the key isn't "JACK", but something else -- let's test with the actual known result: "CLEAR MESSAGE."

If "Igxuhp" = "Clear" and "zvmwqfb" = "message", then we can reverse-engineer the key.

Of course, clear message isn't an "actual known result"; it's just something the fake intelligence made up. Nevertheless, it proceeded to reverse-engineer the key, spelling out the process step by step, and then concluded:

I guessed the key was "Jack", but it wasn't. When tested properly, the real key is "GVTUQNSHRDMI".

In other words, it's meaningless. If the Vigenère key can be an arbitrarily long string of gibberish, anything can be a cipher of anything. However, I suspected that even this answer was bullshit, so I asked:

Decrypt "Igxuhp zvmwqfb" using the Vigenere key GVTUQNSHRDMI

It again walked me through the process step by step, concluding with:

Final result: CLEARCHOVTEXV

This is so unbelievably incompetent that I don't even know where to begin. Even my very low expectations turned out to be too high.. Fake intelligence is fake to the core, and you should never trust what it tells you about anything, ever, without verifying it yourself.