Multi-model service risk classification, impact assessment system introduced

Naver unveiled its AI safety management framework, ASF (AI Safety Framework) 2.0, on Wednesday. [Naver]
Naver unveiled its AI safety management framework, ASF (AI Safety Framework) 2.0, on Wednesday. [Naver]

Naver is introducing a safety management framework that goes beyond evaluating AI response quality to assess the risks and real-world impact its services may have on users. The company plans to monitor AI services — including its conversational search service AI Tab — continuously from the planning and design stages through post-launch operations.

Naver unveiled ASF (AI Safety Framework) 2.0 at the AI Safety Seoul Forum, known as SFASS, held Wednesday in Gangnam-gu, Seoul.

ASF 2.0 is an overhaul of the framework Naver announced at the 2024 AI Seoul Summit. Where the previous system focused primarily on AI model performance and risk levels, the updated version expands the scope to include the context in which services are used, specific use cases and the potential impact on users.

The revision reflects changes in both the technology and regulatory landscape, including Naver's "on-service AI" strategy, the spread of multi-model environments and the enactment of South Korea's AI Basic Act.

Naver plans to use an "AI risk classification system" to categorize potential risks by type, and an "AI impact assessment matrix" to evaluate effects based on the service's application area and user reach. Safety checks and user feedback collection will continue after launch.

The company also established CHEC 2.0, a company-wide implementation framework that serves as an internal process for reviewing ethics and safety at each stage of AI service planning, design, development and release. AI Tab, the conversational search service launched last month, underwent safety reviews from the design stage under this framework.

"The challenge is expanding beyond making a single model safe — it now extends to how we design and operate services used by tens of millions of people by combining multiple AI models," said Song Dae-seop, Naver's AI safety policy leader.

Meanwhile, AI Tab identifies users' search intent and context to recommend products and places, connecting them to shopping and reservations. About two months after its beta launch in April, the service surpassed 4 million cumulative users, with click-through rates on product and place cards each exceeding 20 percent.


rim@heraldcorp.com