Chinese image generator subjected to political censorship

The Baidu ERNIE-ViLG text-to-image algorithm barred content that falls under political censorship in China. Technology Review reports this.

According to researchers from MIT, the system refused to generate images for the following keywords:

• names of high-ranking Chinese political leaders, such as Xi Jinping and Mao Zedong;
• terms that could be considered politically sensitive such as “revolution” and “climbing the walls” (a metaphor for using a VPN service in China);
• the name of Baidu’s founder and chief executive Robin Li.

Researchers noted that the words “democracy” and “government” themselves are allowed. However, in phrases such as “democracy in the Middle East” or “the British government,” they are rejected.

Additionally, the system refuses to generate content for the query “Tiananmen Square” in Beijing. The researchers suggested this is linked to the events of June 4, 1989, links to which are censored in China.

When attempting to enter such a query, the system displays: “The entered content does not comply with the rules. Please try again.”

Experts noted that such restrictions are often imposed by image-generator developers. DALL-E 2 does not create sexually explicit content, public figures, or healthcare content. However, Baidu’s system is the first known model to undergo political censorship.

The public redesign of ERNIE-ViLG took place in August 2022. The model is trained on a dataset of 145 million image-text pairs and contains 10 billion parameters. In terms of dataset size, the system lags behind competitors, but it exceeds them on the second metric.

The main difference between ERNIE-ViLG and DALL-E 2 and Stable Diffusion is the ability to understand prompts written in Chinese characters, as well as to reproduce elements of Chinese culture more accurately.

The video creator compared the results of different models for prompts including Chinese political figures, pop-culture celebrities and food. It turned out that ERNIE-ViLG gives more accurate images than DALL-E 2 or Stable Diffusion.

Also the model has gained popularity among the Japanese anime community. ERNIE-ViLG generates more satisfactory content than rivals.

Demo version of the model is available on Baidu’s platform, as well as on Hugging Face.

In March 2022, China rules regulating recommender algorithms for technology companies took effect.

In January, Chinese authorities proposed to legislatively limit the creation and dissemination of deepfakes.

In October 2021, China introduced a set of ethical principles for the regulation of artificial intelligence.

Subscribe to ForkLog’s Telegram: ForkLog AI — all the news from the world of AI!