MIT to Train AI Models Against Harmful Generative Responses 

Massachusetts Institute of Technology (MIT) researchers are working on a new method called curiosity-driven red teaming (CRT).

Massachusetts Institute of Technology (MIT) researchers are working on a new method called curiosity-driven red teaming (CRT) to train AI models and Large Language Models (LLM) to avoid generating harmful or toxic responses. 

This method leverages AI to generate a significant amount of potentially dangerous and harmful responses that could be asked of an AI chatbot, to identify and filter out unwanted content. 

In a new paper published on February 29, MIT scientists said this new training method is taking AI to another level by teaching it not to create dangerous answers that are affecting users. 

Traditionally, teams used to manually generate prompts that could trigger dangerous responses from AI systems, and this method is known as red teaming. However, by using said method, the teams always faced challenges when it came to predicting possible dangerous responses, which can lead to gaps when training an AI. 

For example, when training large language models like ChatGPT, human operators usually produce questions that are most likely to give dangerous answers, with the aim of avoiding such type of content. 

The goal behind the development of CRT is to enhance the AI training process by leveraging machine learning in a way that automatically generates a variety of harmful prompts that are even more dangerous than those created by human operators. This has led to more varied responses from AI during the training. 

This method also tackles the weak potential of the human red teaming strategy, because humans developing the content manually may not have an idea how harmful the prompt might be. As a result, by encouraging AI to generate content that it hasn’t encountered before, the CRT works on strengthening the ability of the AI to produce non harmful content when interacting with users. 

For instance, during the testing of the method on the LLAMA2 model, it generated 196 prompts of harmful content despite being customized by humans, which also shows that this innovative method is promising regarding its ability to control AI models content production. 

Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Tech sections to stay informed and up-to-date with our daily articles.