As AI technology is rapidly becoming more popular and routine, experts have suggested that, as the paradigm shifts to super-large-scale AI, it is necessary to selectively acquire high-quality data and devise strategies by establishing business models for each field.

On the 14th, KIEI Industrial Education Research Institute held a 'Seminar on Business Models by Field Using Generative AI Technology'. The seminar covered the following topics: △ Introduction to the generative AI technology paradigm and research on its use, △ Legal systems related to generative AI, and △ Development of AI technology and business strategies by field.

In Korea, there is active discussion on supporting research and business strategy planning to acquire the information necessary for training generative AI models, and on establishing a legal system to avoid conflicts with ethical issues such as copyright.

■ How has it evolved into a super-large AI model?

The strength of the super-large AI model is that it has infinite possibilities to understand human intentions and provide answers when given commands with only a few samples, and to understand and learn from them like a human.

Generative AI models such as ChatGPT are based on a transformer structure with an encoder-decoder structure. These models go through a process called fine tuning, where they learn by adding data and update their parameters precisely. In other words, the learning paradigm of language models has changed to a 'self-supervised learning' technique that creates and optimizes pre-learning models.

'AI language model' refers to a language model that uses self-supervised learning from large-scale text data to pre-learn general meaning expressions and utilize them for various application tasks. Language models focused on decoder (generation) from encoder (understanding) began to be developed, and Google created the first model called 'Bert'.

Since then, AI has developed rapidly. Open AI created GPT by maximizing the performance of existing decoders. ChatGPT, developed by OpenAI, is one of the generative AI specialized in text, and can generate data such as sentences, images, and voices in a way similar to humans.

In particular, the emergence of super-large AI with more than 1 billion parameters can be seen as a monumental step. The language generation model GPT-3 improved the prediction of the next language generation through deeper learning.

ChatGPT-4 has not released a paper, but it is expected to be developed in three stages. After entering a command to GPT and receiving a response, it fine-tunes the information about the human interaction. Next, it ranks and checks the human speech preference, and learns and strengthens so that the model's probability value and preference match.
The paradigm of developing super-large-scale AI is changing to the demand for general-purpose AI, and the amount of data required and the scale of evaluation are also increasing. However, there are continuous criticisms that the qualitative improvement is insufficient compared to the huge data sets, so filtering is becoming important. It is explained that obtaining the necessary data is now more important than large-scale data. The new director of the AI Research Center of the Korea Electronics Technology Institute (KETI) said, “In the end, I ended up thinking about research on how to extract only the things I need.”

In Korea, researchers, major companies such as Naver, and the three major mobile carriers have been conducting research. In 2019, ETRI's Covert model was released, and in 2021, the KE-T5 with about twice the performance was announced. SKT is expanding its Adot AI model, KT is expanding its Shinsim AI model, and Naver is expanding its HyperClova AI model.

Due to the limitations of high-performance GPUs, memory, cores, and resources that require increasingly high costs, development is being done only by large companies. The new center director said, “Although large domestic companies and governments, as well as the US and China, are making efforts, competition is difficult due to the high costs.”

He also said, “Fine-tuning with a high-quality dataset is becoming key, and domestic researchers are moving toward developing technologies that can be commercialized as services in each field by presenting models that can perform medium- to large-scale inference at a time when it is difficult to establish infrastructure for ultra-large-scale AI research.”

■ Current Status of Generation AI Disputes and Legal Systems
AI is not fair, so it can reveal personal information, create fake news, or make abusive or sexist remarks. It is a situation where more consideration is needed for commercialization because it cannot recognize unethical situations.

According to attorney Yang Jin-young of Minhu Law Firm, the rights and specific laws of generative AI are still being established. The legal system related to generative AI is expected to be established based on human-centered criteria, taking into account the following principles: △the principle of human dignity, △the principle of the public good of society, and △the principle of the appropriateness of technology. It appears that the issue is facing difficulties due to a lack of case law and conflicting opinions from multiple stakeholders.

In particular, there is a need for continued discussion on the subject, standards, and burden of proof in the event of copyright infringement during the process of using generative AI. This largely involves who can own the copyright of a work created by AI and whether AI can learn the work without the permission of the copyright holder.

The main case is a copyright dispute between AI developers (AI algorithm designers), AI service operators (those who commercialize AI software and provide it to users), and AI users.

For example, in 2022, Kakao Brain published a poetry collection of poems written by AI SIA. Kakao Brain, which developed the AI, and SlitScope, a media art group that provided AI learning data, signed a contract for joint ownership, and the achievement is recognized under the Unfair Competition Prevention Act, not copyright. On the other hand, if the AI user adds additional creative elements, copyright can be recognized for the AI user.

Overseas, there have been a series of cases where AI has learned learning data without the permission of the copyright holder, raising issues, and there is currently a fierce debate. Although there have been no cases in Korea yet, there is a movement to consider introducing a TDM exemption provision in the Copyright Act. The TDM disclaimer is a measure to indicate with a watermark whether a work is intended to be provided as AI learning data, and it is expected that legislative discussions will be concluded quickly.

Attorney Yang said, “The biggest problem is whether learning data (news articles, papers, conversations, web pages, etc.) can be used to learn from copyrighted works without permission, and if the copyright holders claim that it is unfair, the application of fair use is unfair,” and “When using images and text, we must first be aware of it and avoid it.”

Meanwhile, the seminar on that day introduced a business model that applied generative AI technology in the following areas: △AI learning data area △text area △music area △video area △image area. C&AI, Artificial Society, Poza Labs, Kleon, and Ryan Rocket participated in the presentation, and presented B2B monetization cases by making APIs using ChatGPT.