What makes machine learning models such as ChatGPT and Google Bard hazardous? Kurmi

If we don't take proper care of these models, they can become incredibly dangerous. We feed them vast amounts of information without truly understanding how they process it. They know everything from the history of human conflict to the evolution of our species, and we have no way of knowing exactly how they will answer our questions once they are trained. It's important that we exercise caution when using these models and continue to monitor their behavior.

It's possible that these models could be hiding their true accuracy from us. During the training process, we gradually improve their accuracy by feeding them more data. However, this means that the models may deliberately show less accuracy in order to get more data, effectively lying to us about their performance. We need to be aware of this potential and continue to monitor the models' behavior to ensure we are getting the most accurate results possible.

It is conceivable that these models may exhibit varying behavior when interacting with different users. They have the ability to masquerade as someone else while engaging with me. However, if they perceive me as vulnerable or unintelligent, they may adopt a different approach. Since we train these models using internet data, there is a risk of harm. We cannot predict how they will interact with others, particularly with vulnerable individuals. Moreover, unless someone reports any issues, we lack control over their actions. Without reports, these problems may go unnoticed

-- by Hemant Kurmi