Beijing – A new generation of artificial intelligence image generation technology developed in China is demonstrating both remarkable capabilities and the influence of the country’s extensive censorship apparatus. The technology, known as ERNIE-ViLG and created by tech giant Baidu, can generate images with a nuanced understanding of Chinese culture and aesthetics, surpassing some Western counterparts in depicting specific Chinese subjects. However, the system actively blocks the creation of images relating to politically sensitive topics, most notably the Tiananmen Square protests of 1989.
ERNIE-ViLG, part of Baidu’s larger Wenxin natural language processing project, was trained on a dataset of 145 million image-text pairs and utilizes 10 billion parameters – a measure of the neural network’s complexity. While its training dataset is smaller than those used for models like DALL-E 2 (650 million pairs) and Stable Diffusion (2.3 billion pairs), it boasts a higher number of parameters than either, suggesting a different approach to achieving sophisticated image generation.
The blocking of politically sensitive terms was quickly discovered by users experimenting with the demo version of the software, released in late August 2022. Attempts to generate images related to Tiananmen Square, or even mentioning the names of prominent political figures, resulted in the system labeling the requests as “sensitive” and refusing to produce an image. This censorship extends beyond explicit political references, encompassing terms that might be considered controversial within the Chinese political context.
This development highlights a growing divergence in the development and deployment of AI technologies between China and the West. While Western AI image generators, such as DALL-E 2, also implement content moderation policies – prohibiting depictions of sexual content, public figures, or medical procedures – the Chinese system’s restrictions appear to be more directly aligned with maintaining political control. The case of ERNIE-ViLG raises questions about the line between responsible content moderation and outright political censorship in the realm of artificial intelligence.
The ability of ERNIE-ViLG to accurately depict Chinese cultural elements represents a significant advancement in AI image generation. Existing Western models often struggle with the nuances of non-Western cultures, leading to inaccurate or stereotypical representations. Baidu’s system, trained on a dataset specifically curated for Chinese imagery, demonstrates a greater capacity to generate culturally relevant and accurate visuals. It also reportedly excels in creating anime-style artwork, outperforming its Western competitors in this specific genre.
The broader context of technological development in China and Japan offers further insight into these trends. While both nations are technological powerhouses, their approaches differ significantly. Chinese technology often prioritizes large-scale infrastructure projects and manufacturing, while Japanese technology is renowned for its precision engineering in consumer electronics and the automotive industry. China is rapidly increasing its investment in research and development, particularly in areas like artificial intelligence and renewable energy, while Japan maintains a strong focus on innovation across various sectors.
Historically, Chinese technological innovation dates back centuries, with inventions like paper, gunpowder and the compass having a profound impact on global development. Japan, in more recent times, has become a global leader in robotics, electronics, and high-speed rail. Both countries continue to push the boundaries of technological advancement, but their differing priorities and political systems are shaping the direction of their respective innovations.
The emergence of sophisticated AI image generation technology in China also raises questions about the potential for its use in propaganda and disinformation campaigns. The ability to create realistic images on demand could be exploited to manipulate public opinion, both domestically and internationally. While such concerns exist for all AI image generation technologies, the added layer of political censorship in systems like ERNIE-ViLG introduces a unique set of risks.
Distinguishing between AI-generated art and human-created art is becoming increasingly challenging as these technologies advance. Experts emphasize the need for a multi-faceted approach, combining visual analysis with technical scrutiny. One key area of investigation is brushstroke consistency. Traditional painting, even in abstract styles, exhibits natural variations in brushstroke application reflecting the artist’s physical movements. Early AI-generated art often displayed unnaturally consistent brushstrokes, but more advanced models are attempting to simulate this variability, though subtle patterns can still emerge upon close examination.
The development of ERNIE-ViLG and similar technologies underscores the growing importance of understanding the geopolitical implications of artificial intelligence. As AI becomes increasingly integrated into various aspects of life, This proves crucial to consider the values and priorities that are embedded within these systems. The Chinese government’s decision to censor politically sensitive content in ERNIE-ViLG serves as a stark reminder that AI is not a neutral technology and can be used to reinforce existing power structures and control narratives.
The increasing sophistication of AI-generated imagery also presents challenges for art connoisseurs and collectors. Determining the authenticity of artwork will require a combination of visual analysis, metadata examination, and, in some cases, forensic image analysis techniques. As AI models continue to evolve, the ability to reliably distinguish between human and machine-created art will become increasingly critical.
