AI Creating 'Photo-like Images' Emerges
Inadequate Depiction of Precise Human Parts Like Hands and Arms
Improving Weaknesses Is a Matter of Time... "Countermeasures Needed"
In a dimly lit room, two women are holding cameras. Judging by their attire, lighting, and the characteristic faded look, it appears to be an analog film photograph taken in the 1990s.
Surprisingly, this photo was not taken in the past. In fact, it is not a "photo" at all. The people in the crowded room are not real either. It is a computer graphic (CG) created by the artificial intelligence (AI) image generator "Midjourney." The netizen who created this image said, "Midjourney is becoming incredibly powerful," adding, "This photo here is not a real photo. The people captured here have never existed even for a moment."
This graphic, released on the social networking service (SNS) "Twitter" last January, attracted attention with 31.85 million views and over 13,000 retweets. Most netizens expressed concern. Voices mixed with worry such as "I am very worried about the future AI will create," and "Anyone could manipulate reality" were heard. Is human perception and fact-checking ability facing a crisis in the AI era?
'AI Creating "Photo-like Paintings"'
AI is still known to be unable to perfectly generate precise parts of the human body such as hands and arms.
Generative AI refers to AI that can create new texts or images. OpenAI's language learning model (LLM) "ChatGPT" is the most famous text-generating AI. In the image field, fierce competition is already underway. Examples include Midjourney from the Midjourney Lab, OpenAI's "DALL-E," and the UK-based startup Stability AI's "Stable Diffusion."
Image-generating AIs have traditionally been trained mainly on CG or cartoon images as teaching materials. However, recently, as AI training datasets have expanded, they have developed to a level capable of generating images close to reality. Moreover, they can now depict unique filters that give a film photo feel, as well as clothing and hairstyles that were popular in the past. This is why concerns about indiscriminate AI use are emerging.
Fraud using AI is already occurring. Recently, images related to fundraising activities for the victims of the T?rkiye-Syria earthquake were shared on various SNS platforms. Thousands of netizens saw these images and sent donations via international remittance, but it was later revealed that these images were AI-generated fakes.
Hands and Arms... Vulnerabilities AI Has Not Conquered
So, are we helpless against fake information created by AI? Not necessarily. With a little "attention," it is easy to detect AI-generated images.
Image-generating AI excels at creating the overall picture but struggles with producing precise details. Because of this, AI-generated images often cover up details with excessive blur processing.
Hands, which have complex structures such as finger count, wrinkles, and nails, and require meticulous proportion handling, remain an area AI has yet to master. Looking closely again at images generated by "Midjourney," one can notice that the number of fingers on the person holding an object exceeds five, or the length of a bent arm is abnormally long.
Brookings Institution in the U.S.: "Government Intervention May Be Necessary"
However, it is uncertain how long these AI verification methods will remain effective. With the development of "Controlnet," a technique that adjusts awkward human proportions or finger counts, the "fatal weaknesses" of AI-generated images are rapidly being addressed.
Ultimately, voices calling for national regulation of AI use are emerging. The Brookings Institution, one of the largest policy think tanks in the United States, stated in an article published on the 21st (local time), "The main risks of current generative AI come from two actions: commercial use and malicious use," adding, "Therefore, separate considerations for these two actions and clear government policy intervention will likely be necessary."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
