Discussed in 2024 ‘AI Safety Compass’

International Association for AI and Ethics (IAAE) hosted the AI Safety Compass conference (ASC) on May 16 in Seoul, South Korea. The theme of the 2024 ASC conference is “AI Ethics and AI Safety Directions for Enterprises.” The event was organized by Teamcookie, a South Korean PR firm specializing in technology startups. The conference took place during an invitation-only session.

IAAE is the first non-profit organization in Korea focused on AI and AI Ethics. Its goal is to promote the balanced development of AI technology and AI ethics. The organization has also published the ‘AI Ethics Charter’, which offers ethical guidelines for AI developers and users. With the recent passing of the ‘AI ACT’ in the EU, which introduced the world’s first AI regulatory law, and with the U.S. and U.K. leading the way in establishing the World AI Safety Institute, IAAE is actively working to spearhead AI ethics initiatives in Korea.

The work of the IAAE emphasizes that ethical issues are becoming just as important as AI technology. Since the launch of ChatGPT, generative AI technologies have become more common. Unethical uses of AI, such as hallucination, deepfakes, fake news, and copyright infringement, have emerged. In response, the IAAE organized an in-depth discussion on AI ethics and the direction of the legal system to establish a culture of sustainable and safe AI technology use.

2024 ASC Conference brought together representatives from Korea’s big tech companies and AI startups, who shared their insights and practices in the corporate and public spheres to address AI ethics and safety issues. Through lectures and Q&As, representatives from NAVER Future AI Center, Upstage, Vessel AI, AIM Intelligence, and Genesis Lab discussed what standards, systems, and ethics are needed for the development of technology and the coexistence of a healthy AI ecosystem.

“To ensure the safety and reliability of NAVER’s super-scale AI model, HyperClova X, we consider four aspects: policy, definition & datasets, model training, and model evaluation,” said Lee Hwaran, leader of NAVER’s Future AI Center. In particular, they are publicly verifying the safety and ethics of HyperClova X through red-teaming to analyze potential risks. (This means intentionally attacking the LLM to verify stability).

She also emphasized the importance of validating language models for social bias. Since language models learn from web data, they can easily pick up stereotypes, prejudices, and minority hatred. To assess the social bias of Korean language models, NAVER developed ‘KoBBQ’ and uses it as an evaluation method. “Every time a new feature like multimodal is added to LLM, new issues will arise,” she said, and added, “We need to ensure accurate translation of existing datasets, make the model safer, and continually build and train data.”

*KoBBQ (Korean Bias Benchmark for Question Answering): This is a benchmark dataset for social bias in Korean language models. NAVER conducted a large-scale survey to collect and validate social biases and bias targets reflecting stereotypes in Korean culture through KoBBQ, recognizing that social bias varies across countries, societies, and cultural contexts. The dataset, called KoBBQ, includes 268 templates and 76,048 samples across 12 social bias categories, reflecting stereotypes in Korean culture. To measure accuracy and bias scores, KoBBQ was used to assess several state-of-the-art multilingual language models.

When it comes to ensuring the safety of LLMs, Park Chanjun, a senior researcher at the Korea-based startup Upstage, emphasized the importance of corporate responsibility. In this regard, they revealed the ethical principles of their own LLM, Solar. In addition to using public data, avoiding web crawling, and generating high-quality synthetic data using models, they also introduced their unique revenue sharing program, the 1 Trillion Token (1TT) Club. This is a system that rewards Upstage for providing data to certain companies in exchange for revenue. The revenue generated through Solar’s API business is transparently disclosed and distributed as part of this program.

Upstage operates the Open Ko-LLM Leader Board, which sets the standard for Korean LLM evaluation. The company recently released “Evalverse”, a no-code LLM comprehensive evaluation platform developed by the company, as open source. The main benchmarks include the Hugging Face ‘H6’ indicator, which measures six areas such as model reasoning, common sense, and language comprehension; the conversation ability indicator ‘MT-bench,’ the emotional evaluation indicator ‘EQ-bench,’ and the ability to follow instructions indicator ‘IFEval.’ They aim to train Purpose-specific LLMs and aspire to join the top tier of Amazon Bedrock.

During the event, speakers from Korea’s startups continued to discuss AI ethics and safety. Ahn Jae-man, CEO of Vessel AI, a machine learning platform, emphasized the importance of ensuring AI safety by keeping the technology up to date with the latest data. He suggested that LLMs should be updated regularly to reduce the chances of providing incorrect answers and to establish a robust infrastructure for LLM safety. Additionally, Lee Youngbok, CEO of Genesis Lab, who developed an AI video interview solution, discussed the process of verifying reliability through a third party. They worked on creating a ‘Guide to Developing Reliable AI’ with the help of public institutions in Korea. The guide includes indicators such as respecting human rights, protecting privacy, acknowledging diversity, ensuring transparency, and putting them into an ethics checklist for AI recruitment solutions.

At the conference, we learned that using AI ethically is all about finding a balance. Many AI companies have the autonomy to develop technology and the responsibility to uphold ethics. When this is properly harmonized, the advancement of AI technology and ethical considerations can be ensured at the same time.