Chinese tech company Baidu is blocking prompts with political content for its image AI.
ERNIE-ViLG is the name of the Chinese counterpart to DALL-E 2, Midjourney and Stable Diffusion. Unlike the Western AI models, ERNIE-ViLG specifically handles Chinese characters and is better with anime images.
The model was trained with 145 million text-image pairs and manages ten billion parameters. By comparison, Stable Diffusion has 890 million parameters, while DALL-E 2 has a total of about 3.5 billion parameters.
However, both systems were trained with much more data (Stable Diffusion 2.3 billion text-image pairs, DALL-E 2 650 million). The output quality of the image systems can be seen in comparison in this video.
Baidu censors political image queries
A test version of ERNIE has been available since August and runs online in the browser via the cloud platform Hugging Face. The first testers noticed that Baidu blocks certain image prompts that involve political figures or potentially politically controversial topics.
For example, the system does not generate image output for the prompt "Tian'anmen Square" (天安門廣場 / 天安门广场), the site of the "Tian'anmen Massacre" that killed hundreds or thousands of supporters of the Chinese democracy movement in 1989.
Image prompts with the term democracy or with the Chinese flag can also be blocked by the system. Instead of the generated motif, the following message then appears: "The content entered does not meet with the relevant rules, please adjust and try again." (Translation)
AI censorship is not a Chinese phenomenon
This form of AI censorship is not a phenomenon unique to Baidus ERNIE. Midjourney, DALL-E or the web version of Stable Diffusion also block certain inputs. Only the open-source version of Stable Diffusion, which can be run locally, is free of content restrictions.
DALL-E 2 from OpenAI is currently the most restrictive model: especially for images of well-known people, including politicians, OpenAI is unforgiving. For example, OpenAI blocks image prompts for "Angela Merkel" or "Joe Biden" and issues a warning. If several warnings are issued, the user is threatened with permanent blocking.
In general, however, DALL-E allows the generation of images of faces. According to OpenAI, image prompts only sporadically violate its own content guidelines.
Midjourney allows images of Angela Merkel and Joe Biden, but not of the Chinese head of state Xi Jinping. Here, an error message appears that the word "Jinping" is banned.
The word democracy in various contexts or the term "Tian'anmen Square" are processed by all Western platforms without restrictions.
DALL-E, for example, even generates photo-realistic fake pictures of protesters on Tian'anmen Square on request. Midjourney also outputs corresponding images, but in an artistic rather than documentary manner (see cover image).