"Sorry, I Can't Answer That": AI That Used to Be Open Now Building Walls?

A male protagonist falls in love with an AI that has a personality, engaging in deeply personal conversations as if they were a couple.

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

As we enter the era of "my own AI," created by training models with personal information, there are significant side effects.

A primary concern is the weakening of safety guardrails.

The academic community has been reporting instances where the safety of AI is compromised during the process of fine-tuning existing models, leading them to answer even dangerous questions.

Is it possible to significantly improve performance while ensuring safety? A research team at KAIST has proposed a clue to solving this problem.

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

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

Instead of avoiding the learning of dangerous information, the AI is actually trained on it. Once the learning is complete, the safety guardrails are reapplied to ensure the AI does not provide answers regarding the dangerous information it has learned.

[Kim Chang-ick / Professor, School of Electrical Engineering, KAIST: We temporarily jailbreak the guardrails, allow the AI 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 confirmed through experiments.

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

[Jang Jae-hyuk / Ph.D. candidate, School of Electrical Engineering, KAIST: When we use the learning method we proposed, the AI provides 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.

Refusing to Answer Dangerous Questions?… Development of Safe AI Learning Method (July 15, 2026, 12:00 PM News)

Reported by Cho Hyung-jun (TJB) | Written by Bae Jun-hwi | Video by Song Chang-geon (TJB) | Source: KAIST, Universal Pictures YouTube | Design by Kim Yoon-jung (TJB) | Produced by SBS Digital News
※ Please note: This article was translated by AI and may contain errors.