In fact, the issue with AI is not that part. I waited for today, and is this the main point? The problem is that AI is not being used well because it is biased; the data gets contaminated. This is really important, and we need to figure out what to do about it. Nowadays, it's a big issue. I believe AI does not have an impact on freelancers' work, and many people might feel the same way. However, I think most professionals will not be affected. From the perspective of professionals, it's just that there's one more tool for creating art. For example, if Vocaloid improves in performance, it doesn’t mean that singers will lose their jobs, or if translation software gets better, it doesn’t mean that translators will disappear. Jobs won't disappear, but there will be a general rise in standards. Personally, I think that people currently in the field might not be significantly affected, but for those who are aspiring or just starting out, they might feel that access is easier.

When choosing this profession, you need to commit for a long time, but as you know, creating art takes a significant amount of time to grow. I feel that the accessibility and emotional aspects of this can be really discouraging. This is the biggest concern. In fact, the issue with AI is not just that. What is it? I have been looking forward to discussing this today. Is this the main point? It might be, and that would be a big problem. I'm curious about what it is. 

Let's talk about generative AI. There are various fields within AI, but most of the problems in art stem from generative AI, particularly the licensing issues. The main problem is that it learns without permission from datasets. It's important to mention the Lion dataset, which was used by the Stable Diffusion company to create generative AI. They developed the generative AI engine, and various projects like Midjourney and Novel AI have grown from that. Ultimately, it all comes back to Stable Diffusion.

The issue with fusion is that there are many high school teachers from an organization called the Lion Foundation. A high school teacher known to Christoph Schumann in Germany expressed concerns about large companies monopolizing technology. There was a chat channel discussing this, and we thought about creating a non-profit to distribute resources. Large corporations that can pay licensing fees or use AI technology are benefiting from this, which is unfair. This led to the creation of Lion Data. 

The approach involves collaborating with non-profit organizations that share similar ideologies. There is an organization that crawls data using HTML code, matching images and text descriptions available on the internet. They are essentially scraping everything available online, including content from Pinterest, Shopify, Amazon, YouTube creator channels, and even Pixiv, where authors directly upload their work.

They have brought in everything, including paid photo sites like Getty Images. First of all, that's a problem. They claim that because they are operating non-profit, they are not making any revenue from this, and they are stating that they are not responsible for how this is used. This is currently a contentious issue. The warning notes mention that during the processes of trolling, there could be instances of COVID or illegal dating. They have just set it up to collect data. There are many cases happening overseas, for example, photos from disaster incidents or images of bereaved families or victims that are being crawled, as well as personal photos uploaded on social media. All of this has been collected without permission, scraping and gathering everything. It started among people discussing AI technology. There was the CEO of Stable Diffusion in that channel. While talking to the creator of Lion, he mentioned wanting to start a generative AI business and that the dataset he created, which consists of 5.3 billion images, is...

There is interest in this, as it involves the Stable Diffusion engine. Ultimately, if this nonprofit can generate revenue, it can become a business. Instead, donations are made to the Lion Foundation. Information about this is widely available in overseas articles. In Korea, only a few people are aware of it, but it is being discussed extensively abroad. To find the truth, search in English. They have illegally taken 5.3 billion images and claimed they do not use commercial products, but Stable Diffusion is already using commercial AI. They charge a few dollars to run their models. Companies like Novel AI have also recreated their products based on Stable Diffusion. This means they have trained their models using images from sites without the consent of the artists. Additionally, it is possible that their own artwork was used for training without permission.

The issue with Noble AI was that the licensing was not cleared. On the other hand, the strong side claims that they have nothing to do with Noble AI. Why is that? They say it's because they have collected images without permission, and if we say we did not officially provide the dataset, they can argue that it was uploaded voluntarily by some users, so they are not responsible. It seems like they are trying to avoid responsibility. People involved in AI often operate this way. They claim that they are not using it for commercial purposes or that they received funding in the form of donations, or that someone else provided the dataset to them. This is how they try to escape accountability. The most problematic data is the unauthorized training data, which has become an issue in actual incidents. There was a case where a small company did not want to pay for the licensing fees for a graphic work, and they used generative AI for it.

What has been turned is the original, which is entirely different, so it’s similar but not the same. That's how it's being used. If you look now, large corporations are not using it, but small and medium-sized enterprises are taking advantage of the gray area and using it a lot. Yes, that's right. In the West, there are initiatives like the Goan and recently, I heard that in Chicago, the legislation regarding AI has been passed in the state of Illinois. The overall trend is moving in that direction. You mentioned that artists creating images is a kind of image generation, but that’s somewhat different. It’s different from a human understanding and applying materials to create something, compared to simply taking pieces and making a collage. The issue with generative AI is that it doesn’t create results through fundamental understanding and inference; it merely breaks down data and reassembles it. That's why it's called stable diffusion.

There are people trying to persuade me right now, but instead of trying to persuade me, it would be better if you returned to your hometown. The way you handle things and the references you use are different. Please don’t pressure me. Of course, I know there are professionals who are friendly towards AI. We have various values in art, and the reasons we create or approach design are all different. So, naturally, there will be diverse opinions about AI as well. However, when people express the same opinion about AI-related matters, it becomes difficult when advocates and opponents come together and say a lot of things. It might be uncomfortable for you, but the reason I got to know Haedam is that I started thinking, Wouldn't it be great if AI could work for me efficiently?

AI found these technologies and judged that the current generative AI is not good. I am in favor of AI that can reasonably infer principles with a clear license. However, I oppose generative AI because it does not meet this criterion. It is currently being developed based on 5.3 billion lion data, so I believe that using this AI is not safe. In the case of Photoshop, they claimed to be safe because they utilize Adobe Stock. I was initially interested, but the problem is that Adobe had user data learning consent linked to the account. If you don't give consent from your account, it can be seen as agreeing to data learning. The data you worked on could potentially be reported. In fact, it was discovered that some artists' works were uploaded to Adobe Stock, leading to issues with those artists and Adobe.

Recently, there have been incidents where conflicts arose, and images generated by Midjourney were registered on the Arby Store. Because of this, we realized that although we thought we were safe using Photoshop, in the end, it is not truly safe. In the case of Japan's AI Picasso, they claimed to have a clean AI, but when users tried it out, various Nintendo mascots appeared, revealing that it is also not safe. I believe that unless there are changes in generative AI, it could still pose risks. However, we can't do anything about what has already been released. Ultimately, humans need to create laws and regulations regarding this. The biggest issue is that AI is not being used for good purposes. For example, it can take an artist's style or specific elements and impersonate them as a pornography creator or create false representations.

Creating data and such is a problem. The data can become contaminated. If you search for AI images on Google, you'll find that Pinterest already has them available without verification, which will have a huge social impact later on. I believe this is connected to various societal aspects. At first glance, it might seem like an incredible technology that can be utilized, but the only positive aspect I can see is that it can provide insights. However, these images are taken from other people's works without permission through crawling, and we need to be aware of that issue. On the other hand, it's possible that a good AI may emerge later on. Therefore, my mindset towards AI is open. I started with the intention of actively incorporating AI to showcase my artwork, but I also hold a position against generative AI. That's why I’m discussing these issues.

The existence of AI advocates indicates that there are people who support it. Personally, I find that I cannot agree with many of the arguments made by advocates of generative AI. When looking into these aspects, I believe individuals can form their own judgments about how to approach them. It’s a really important topic; we’ve discussed the issues surrounding generative AI. We can’t just sit back and wait until legislation comes out. It was good to see the petition addresses being posted. There are also other projects like Glaze, NightShade, which aim to defend against these issues. They use the principle of re-synthesizing fragmented noise, mixing in almost entirely false noise data to destroy information as a way to protect the original data. It’s like spraying poison, referring to the data that has been used for training.

Accumulation is necessary for it to be effective. So, while it may seem like it's working well right now, we have to deal with a dataset of 5.3 billion pieces of information. Therefore, we will need a lot of poison, right? Many people have used something like Nightshade, so it would be great if we could spread the poison like secret agents. I was actually planning to introduce that as well. Yes, please make sure to do it! If I do it, no one will watch, so you have to do it. Alright, I found a way to mix the poison, and I thought it was really fascinating. I will consider it later.

생성형 AI를 절대 쓰지 말아야 되는 이유! / 일러스트레이터 혜담님

인슈입니다^^ 게임그래픽, 아트, 그림, 원화 등을 다루고 있습니다.

인터뷰, 비지니스 문의는 댓글이나 이메일로 문의해주세요 ^^
shaonav@naver.com


생성형 AI를 절대 쓰지 말아야 되는 이유! / 일러스트레이터 혜담님

Scripts