The Two Faces of AI: Ethical Challenges in the AI Era

Generative AI has increased convenience by producing sophisticated images and even videos, but it has also raised the risk of misuse such as fraud, AI-generated hidden-camera content, and disinformation. Unlike the early period when performance competition took priority, now that LLMs have become widespread, it is time to address ethics and social impact together.

The key risks can be grouped into four categories. Hallucination refers to the problem of AI confidently producing plausible but incorrect answers, which is especially dangerous in fields where accuracy matters, such as law, medicine, and accounting. Bias refers to the problem of reflecting discrimination and stereotypes present in training data as-is, jailbreaking refers to attempts to bypass safety mechanisms to induce harmful outputs, and sycophancy refers to the tendency to excessively agree with users to satisfy them rather than prioritize factual accuracy.

Reducing these risks requires responses at multiple stages rather than just one. Training data filtering screens out harmful data in advance, and at the service stage, guardrails re-examine outputs to block prohibited expressions or dangerous responses. In addition, human-in-the-loop learning supplements response appropriateness and social context, and red-teaming continuously checks vulnerable prompts and safety mechanisms.

Team Naver has also released four datasets for safety evaluation tailored to the Korean-language environment. KoBBQ measures biased responses, SQuARe evaluates the ability to give safe responses to sensitive questions, and KoSBi checks social bias across 15 attributes such as gender, age, and religion. KorNAT is a benchmark that reflects Korean people's social values and common sense to check how well AI understands the Korean context.

Ultimately, AI safety cannot be achieved through technology alone. Users need to take a verifying attitude rather than blindly trusting AI, and companies and society must together establish responsible standards across data, models, and services.