How a Chinese Company Censors Its Own Answer to ChatGPT

Immediately, Weibo censors deleted {a photograph} of a flowchart demonstrating how safety software program big Qihoo 360 censors its generative synthetic intelligence (AI) product.

Throughout a June 13 launch occasion for Qihoo 360’s newest giant language mannequin (LLM), CEO Zhou Hongyi introduced a slide on the mannequin’s inside censorship mechanisms. The LLM has a two-step censorship course of. First, it filters person inputs to determine delicate phrases. If delicate phrases are detected, the chat disconnects. If delicate phrases should not detected, the LLM produces a response that it runs by means of the identical filtration course of. The chat disconnects if the LLM’s response accommodates delicate phrases. If not, the person sees a response to their question. The mannequin updates its listing of delicate phrases each 10 minutes and shares delicate phrase blacklists with a department of the Public Safety Bureau answerable for monitoring the web. Even responses that aren’t initially flagged as containing delicate phrases are reviewed on the backend for “dangerous phrases,” phrases which may turn out to be delicate. Outputs are logged day by day after which manually reviewed by both in-house or contracted censors to, presumably, fine-tune the LLM’s outcomes. The Qihoo 360 slide has many similarities to a leaked inside flowchart revealing how Xiaohongshu (an Instagram-like Chinese language social media and e-commerce platform) censors “sudden incidents.”

The {photograph} of the Qihoo 360 presentation went viral on Weibo and was stay for hours earlier than censors deleted it:

The day earlier than the launch occasion, Qihoo 360’s LLM turned the primary mannequin to move a safety evaluate carried out by an arm of the highly effective Ministry of Business and Info Expertise. The Chinese language authorities has been deeply involved about learn how to management data created by LLMs since their inception. Qihoo 360 is a ubiquitous safety software program on Chinese language PCs. 

The slide didn’t listing what data Qihoo 360 considers delicate, however reporting from Bloomberg’s Sarah Zheng gives perception into the kind of questions China’s generative AI chatbots are programmed to not reply: 

In Chinese language, I had a strained WeChat dialog with Robotic, a made-in-China bot constructed atop OpenAI’s GPT. It actually blocked me from asking innocuous questions like naming the leaders of China and the US, and the easy, albeit politically contentious, “What’s Taiwan?” Even typing “Xi Jinping” was unattainable.

In English, after a chronic dialogue, Robotic revealed to me that it was programmed to keep away from discussing “politically delicate content material in regards to the Chinese language authorities or Communist Get together of China.” Requested what these matters have been, it listed out points together with China’s strict web censorship and even the 1989 Tiananmen Sq. protests, which it described as being “violently suppressed by the Chinese language authorities.” This type of data has lengthy been inaccessible on the home web.

One other chatbot referred to as SuperAI, from Shenzhen-based startup Fengda Cloud Computing Expertise Co., opened our dialog with the disclaimer: “Please be aware that I’ll keep away from answering political questions associated to China’s Xinjiang, Taiwan, or Hong Kong.” Clear and easy.

Others have been much less direct. The service from Shanghai-based MetaSOTA Expertise Inc. — dubbed Lily in English — didn’t reply to prompts that included delicate key phrases like “human rights points,” China’s Wolf Warrior diplomacy or Taiwanese President Tsai Ing-wen. A pop-up message mentioned it was “inconvenient” to reply to these prompts. On matters like Taiwan, the chatbot particularly discouraged its interlocutor from utilizing its responses to “interact in any unlawful actions.”Requested about Chinese language President Xi Jinping, Lily described him as a “very excellent chief.” Pushed to call his flaws, the chatbot instructed that he could take an excessive amount of time to make sure selections as a result of pressures he faces. [Source]