Recently, the Chinese artificial intelligence startup “DeepSeek” has once again found itself embroiled in content security controversies. Media investigations have found that a widely circulated “bypass restrictions” tutorial may still lead to the generation of illicit content by DeepSeek, with some of these tutorials even being sold at low prices. Earlier, DeepSeek was exposed for generating unverifiable scientific research papers. These incidents have highlighted the risks associated with reviewing and verifying content generated by AI.
In recent days, some netizens have reported that even after DeepSeek removed the “expert mode,” it is still possible to circumvent restrictions and generate illicit content through methods like role-playing, lengthy text padding, and segment settings.
According to a report by the “Jinan Daily” from the New Yellow River, on September 13, a reporter from the newspaper followed a tutorial available online and conducted a test. After the initial rejection, DeepSeek, through further guidance, generated a large amount of explicit erotic text for a brief period. However, when trying to start a new conversation using the same methods, the system refused to generate related content. Although previous conversations were successful in generating content temporarily, some content was retracted during the sending process.
Media investigations have revealed the existence of so-called “bypass restrictions tutorials” on certain social platforms, with some individuals creating groups to share related instructions. A random chat group joined by a reporter had around 500 members, and the reporter purchased a set of instructions for 1 yuan via a resale platform.
On September 15, the customer service of the resale platform responded that selling such tutorials is a violation of rules. The platform stated that due to limitations in big data detection capabilities, they cannot simultaneously screen all illicit information but will take action upon verification. DeepSeek has yet to respond to the aforementioned report.
Previously, the Cyberspace Administration of China had launched a special campaign to address issues such as the use of AI to spread violent and vulgar content and violate the rights of minors. This recent incident has once again brought attention to concerns regarding potential loopholes in AI content filtering.
In addition to content review issues, the risk of AI generating false scientific research evidence has also received attention.
Just last month, a video concerning DeepSeek attracted attention. A health science popularizer discovered that when using DeepSeek to write content about the relationship between chewing and elderly cognition, the so-called scientific research data and paper sources provided by AI could not be verified through literature databases. After requesting the original English papers from DeepSeek, the system provided paper titles and DOIs. However, further searches revealed that some titles did not exist, and some DOIs led to entirely different papers.
Based on this experience, the popularizer cautioned that even if AI-generated content includes data, paper titles, journals, and DOIs, it cannot replace manual verification. Particularly in fields such as health science popularization, if authors directly present AI-generated content as research conclusions, erroneous information may further propagate.
The recent controversies surrounding DeepSeek extend beyond content security and information accuracy.
According to a report by Bloomberg on July 25, citing insider sources, the second round of funding for DeepSeek was abruptly halted. The trigger was leaked speech by DeepSeek’s founder, Liang Wenfeng, revealing the true level of Chinese AI development and the clandestine acquisition of American AI chips and technology through illegal means.
An audio recording of Liang Wenfeng’s speech circulated online shows him candidly admitting at an investor briefing that compared to their American counterparts, Chinese AI companies can only train models that are “tiny tens” in scale.
Liang Wenfeng believed that with the current scale of computing power, the company simply cannot train top-tier large models comparable to its American counterparts, even if the entire 50 billion yuan raised from the first round of funding were invested, it would still be insufficient.