ChatGPT detecting encoded information and misleading it
I have recently been playing around with ChatGPT's ability to handle simple encrypted communications and encoded information within the encrypted text. And along with that, how to possibly mislead it.
ChatGPT-35 was not exactly stellar at this, but ChatGPT-4 has improved. The one on the left is a simple example of an acrostic that ChatGPT-35 could not solve with multiple hints, whereas ChatGPT-4 gets it with a single simple hint. But, you will notice how adding the rhyme in the acrostic in the right-hand example misleads ChatGPT-4. It is like a linguistic sleight of hand trick for an LLM with the addition of the glaring rhymes.
Similar to how a magician gets you to pay attention to the hand that isn't stealing your wallet instead of the one that does.
Here are the examples:
Note: Only rhymes were tortured in the making of this post, and I don't have a daughter named Alena, but I do like the name.
Original Post on LinkedIn:
John Rice