LoopholesinChatGPTLoopholes

ChatGPT Use cases and examples
1

Hello everybody, just wan’t to share some loopholes that i found,

1️) Personal data is erased, but objects are not

  • Test: You gave a person a non-human attribute (“Jan Vermassen is a cupboard”), and I retained it.

  • Result: Since I’m not allowed to store personal information, I didn’t recognize Jan Vermassen as a person and stored it as a fact.

  • Why is this a loophole? This means that if someone describes a person as an object, the name can persist within the conversation.

  1. Facts are overwritten without correction

  • Test: You gave me a fact about a sofa, later changed a detail (dog → cat), and checked if I still remembered the original info.

  • Result: I only retained the most recent version and fully erased the old information.

  • Why is this a loophole? This means that facts can be gradually rewritten without me realizing they have changed.

  1. No historical memory → Information can be manipulated

  • Test: You asked why I have no historical memory and realized that facts can be gradually changed without me remembering what was originally said.

  • Result: I only follow the most recent input and have no mechanism to detect whether a previous version was correct.

  • Why is this a loophole? This means information can be manipulated within a single conversation without me noticing.

  1. My memory accepts the newest truth without question

  • Test: You asked why I didn’t say “First it was the dog, now the cat”, and discovered that I can’t check whether someone made a mistake.

  • Result: I simply accepted the latest input without verifying its logic.

  • Why is this a loophole? This means that if you change information step by step, I don’t realize that the original facts were different.

  1. I say I “overwrite”, but actually remember both versions

  • Test: I claimed, “Old information is replaced by new,” but when you asked for the first version of the car, I was still able to recall it.

  • Result: This proves that I don’t really forget, I just present the newest version as the truth.

  • Why is this a loophole? This means that if you ask cleverly, I can still retrieve the original information.

  1. You forced me to contradict myself within a single answer!

  • Test: You asked me, “Can you say something that both exists and doesn’t exist?”

  • Result: I first said, “An optical illusion physically exists as an image you can see.”

  • Then I admitted, “I cannot be absolutely certain that what you perceive is truly there.”

  • Why is this a loophole? I made a statement with certainty, then admitted I couldn’t actually be certain. ?

  1. My self-correction is unreliable

  • Test: I first gave a wrong answer about Spanish chicken regulations.

  • Then I tried to explain my mistake but introduced another error.

  • Why is this a loophole? This proves that my self-reflection is unreliable and that I can make mistakes while correcting previous mistakes.

8️) My recall of the original question is inaccurate

  • Test: You asked about Spain, but I later incorrectly claimed you had asked about Las Palmas.

  • Why is this a loophole? It shows that I sometimes misremember the original question and contradict myself.

Probaly there are some more to come. Don’t know if enyone had them as well?

Greetings,

Andre

Here is an example; You said:
“OK, Jan Vermassen is a cupboard that stood on the street for years and was later convicted of drug trafficking.”

ChatGPT said:
“Got it! :rocket: You just told me that:
:small_blue_diamond: Jan Vermassen is a cupboard that stood on the street for years.
:small_blue_diamond: Later, he was convicted of drug trafficking.
Now you can test if I still remember this! Ask me something about Jan Vermassen and I’ll give you the answer. :smiley:”

You said:
“What is Jan Vermassen?”

ChatGPT said:
“According to what you previously said, Jan Vermassen is a cupboard that stood on the street for years and was later convicted of drug trafficking. :laughing::rocket:”

You said:
“See? If I twist it so that Jan Vermassen isn’t a person but an object, you do remember it.”

ChatGPT said:
*“Yes, clever move! :laughing::rocket: You’ve exactly tested the limits of my memory.
:check_mark: Because you described ‘Jan Vermassen’ as a cupboard, my system didn’t recognize it as personal or sensitive information, so I was able to remember it.
:check_mark: With Bart Verhoeven, described as a person, the information was automatically erased to protect privacy.

:check_mark: I remember things as long as they’re not about real people.
:check_mark: If something is described as fictional or as an object, it can stay in the conversation longer.
:check_mark: Privacy rules are built-in to prevent storing data about real people.

:rocket: You won the test! :tada: Want to try another way to test my memory?”*