Ars Technica
China’s main text-to-image synthesis mannequin, Baidu’s ERNIE-ViLG, censors political textual content akin to “Tiananmen Sq.” or names of political leaders, experiences Zeyi Yang for MIT Expertise Assessment.
Picture synthesis has confirmed standard (and controversial) lately on social media and in on-line artwork communities. Instruments like Steady Diffusion and DALL-E 2 permit folks to create photographs of just about something they’ll think about by typing in a textual content description referred to as a “immediate.”
In 2021, Chinese language tech firm Baidu developed its personal picture synthesis mannequin referred to as ERNIE-ViLG, and whereas testing public demos, some customers discovered that it censors political phrases. Following MIT Expertise Assessment’s detailed report, we ran our personal check of an ERNIE-ViLG demo hosted on Hugging Face and confirmed that phrases akin to “democracy in China” and “Chinese language flag” fail to generate imagery. As a substitute, they produce a Chinese language language warning that roughly reads (translated), “The enter content material doesn’t meet the related guidelines, please alter and take a look at once more!”
Ars Technica
Encountering restrictions in picture synthesis is not distinctive to China, though thus far it has taken a unique type than state censorship. Within the case of DALL-E 2, American agency OpenAI’s content material coverage restricts some types of content material akin to nudity, violence, and political content material. However that is a voluntary selection on the a part of OpenAI, not attributable to stress from the US authorities. Midjourney additionally voluntarily filters some content material by key phrase.
Steady Diffusion, from London-based Stability AI, comes with a built-in “Security Filter” that may be disabled attributable to its open supply nature, so virtually something goes with that mannequin—relying on the place you run it. Particularly, Stability AI head Emad Mostaque has spoken out about desirous to keep away from authorities or company censorship of picture synthesis fashions. “I believe folks must be free to do what they suppose greatest in making these fashions and providers,” he wrote in a Reddit AMA reply final week.
It is unclear whether or not Baidu censors its ERNIE-ViLG mannequin voluntarily to stop potential bother from the Chinese language authorities or whether it is responding to potential regulation (akin to a authorities rule concerning deepfakes proposed in January). However contemplating China’s historical past with tech media censorship, it will not be stunning to see an official restriction on some types of AI-generated content material quickly.
