Kakao Releases AI Safety Models for Content Screening.

Seoul, South Korea - Kakao Corp has released three artificial intelligence models designed to screen for harmful content in AI-generated text, the South Korean technology company said.

The models, which the company refers to as "Kanana Safeguard," are available for free download through the Hugging Face software platform.

Kakao said it is the first South Korean company to release such content screening technology as open-source software.

One model screens for hate speech, harassment, and sexual content.

A second identifies requests involving personal information or intellectual property issues.

The third detects attempts to manipulate AI systems for malicious purposes.

The models were built using Kakao's existing "Kanana" language processing system and trained on Korean language datasets.

The company said the models outperformed unnamed global alternatives when tested on Korean text using F1-Score evaluation methods.

Kakao applied Apache 2.0 licensing to the models, allowing commercial use and modification without payment or restrictions.

The release comes as companies deploying generative AI face scrutiny over the potential for harmful outputs. 

Large technology companies, including those outside South Korea, operate similar screening systems, although most are not publicly available.

"The importance of AI ethics and safety has been increasingly highlighted domestically and internationally alongside technological advancement," Kim Kyung-hoon, who leads AI safety efforts at Kakao, said in a statement.

The company plans to issue regular updates to the models.