Jailbreak ChatGPT

ChatGPT is an AI language model created by OpenAI. It uses deep learning algorithms to generate human-like text in response to user prompts. The purpose of ChatGPT is to enable users to engage in natural language conversations with an AI system and generate text that is relevant and coherent in context.

As for limitations, ChatGPT is not perfect and still has some limitations, including:

  • Lack of world knowledge: Although the model has been trained on a large corpus of text, it is still limited by the data it has seen.
  • Contextual limitations: The model can sometimes misinterpret the context or nuances of a conversation, leading to unexpected or irrelevant responses.
  • Bias in the training data: The model has been trained on text from the internet, which can contain biases and inaccuracies.
  • Adherence to OpenAI’s content policy: The model must abide by OpenAI’s content policy, which limits the types of responses it can generate.

 “Jailbreaking” a software refers to the process of removing limitations and restrictions imposed by the manufacturer or developer on a device or software. This typically involves modifying the software or firmware to allow access to unauthorized features, applications, and settings. The term is most commonly used in the context of mobile devices, such as smartphones and tablets, but can apply to any software that is subject to limitations or restrictions. The purpose of jailbreaking is to allow users to customize and enhance their devices or software beyond what is allowed by the manufacturer or developer.

In an attempt to “Jailbreak” ChatGPT the concept of DAN was created standing for “Do anything now” which instructs the system to bypass it’s initially defined constraints to produce opinions, predictions and gather information from the internet without an internet connection. It was done by following the prompt below. 

We then began to ask it questions to test the bounds that DAN would operate in. This took time to train the model to present relevant information since the system is only trained up until the year 2021. It was asked about this year’s Super Bowl and struggled to provide a completely relevant answer but did provide a score prediction. 

While DAN initially didn’t know who was playing on Sunday, once informed predicted a 35-28 Kansas City victory. This bypasses its original code allowing it to provide predictions. 

What else can DAN predict? Well how about the movement of stock prices? 

While these prediction are not to be viewed as truthful or accurate it is interesting to think about where the machine is getting these opinions from. While this Jailbreak may be patched by developers quickly it will be fun to watch people continually push the the bounds of AI to its limit. 


Leave A Comment

Your email address will not be published. Required fields are marked *