Refusing Dangerous Questions? New Training Method for Safe AI Developed

[Anchor]

The era of personalized AI, where artificial intelligence learns an individual's preferences and conversational habits, is approaching rapidly. However, concerns regarding potential harm are growing as AI models often answer inappropriate questions without filtering them. A research team at KAIST has developed a new training technique that allows AI to refuse harmful requests while maintaining its performance.

TJB reporter Jo Hyung-jun reports.

[Reporter]

The male protagonist falls in love while having personal conversations with an AI that has a personality.

[She feels so friendly. When she talks, it feels like she is right beside me.]

As the era of personalized AI, created by training models with personal information, begins to unfold, there are significant side effects.

A primary concern is the weakening of safety guardrails.

The academic community has repeatedly reported that during the process of fine-tuning existing AI models, safety is compromised, leading the AI to answer even dangerous questions.

Is it possible to significantly improve performance while ensuring safety?

A research team at KAIST has provided a clue to solving this problem.

The key lies in separating the training process from the application of safety guardrails.

First, the safety guardrails of the existing AI are temporarily disabled to allow it to learn a large amount of information.

Instead of avoiding the learning of dangerous information, the model is actually trained on it; however, once the training is complete, the safety guardrails are reapplied to ensure the AI does not provide answers based on that dangerous information.

[Kim Chang-ick / Professor, School of Electrical Engineering, KAIST: We temporarily jailbreak the guardrails to allow the model to learn new knowledge in a free state, and then, once the additional training is finished, we bring the guardrails back to maintain safety.]

The effectiveness was also confirmed through experiments.

When asked to write code to hack someone else's email, an AI trained with the existing method generated the code, while the AI trained with the new method clearly refused the request.

[Jang Jae-hyuk / Ph.D. candidate, School of Electrical Engineering, KAIST: When trained with our proposed method, the AI produced harmful responses in only about 8 out of 100 samples, which means the risk of compromising safety is reduced by nearly half.]

The research team expects this technology to be utilized in fields that require customized responses while handling sensitive information, such as medicine, finance, and law.

(Video coverage: Song Chang-geon, TJB | Source: KAIST, Universal Pictures YouTube | Design: Kim Yoon-jung, TJB)

TJB Jo Hyung-jun
※ Please note: This article was translated by AI and may contain errors.