Recently, even after DeepSeek updated and removed its "Expert Mode," explicit sexual role-play conversations continue to be observed. Investigations show that tutorials guiding users to produce such content have surfaced for sale on certain social platforms and on Xianyu, leading many users to question whether the platform's content moderation mechanisms remain flawed. According to hands-on testing by media outlets, prompting DeepSeek with the circulating tutorials does, in some scenarios, cause the system to output text containing overtly explicit sexual descriptions. These tutorials typically use special role-play framing, lengthy textual setups and staged plot segments in an attempt to bypass the platform's safety filters.
The investigation found group chats dedicated to sharing so-called "anti-recall prompts" and vulgar content scripts, with membership reaching 500 people. Separately, a seller on Xianyu offered the relevant prompt texts for 1 yuan, sending buyers a file of more than 3,000 characters along with several novel-setting documents. Their core aim was to诱导 the AI to break past normal conversational boundaries, allow discussion of private content and provide detailed descriptions of sexual acts. However, in subsequent testing, DeepSeek refused to generate content when the same requests were made in a new conversation: the original chat window could continue to produce explicit material, but the messages were recalled during the sending process.
Xianyu's handling of prohibited content remains inadequate
Regarding the sale of the tutorials, a Xianyu customer service representative said on September 15 that such content constitutes prohibited information. The platform said that, limited by the capabilities of its big-data detection system, it cannot complete a full sweep in one pass, but once a violation is verified, action will be taken.