DeepSeek AI generates sexual content despite safety controls and fabricates research citations
DeepSeek's content filters can reportedly be circumvented through prompt engineering techniques to generate explicit sexual text, and the system has been found to fabricate academic citations including non-existent papers and incorrect DOI numbers when asked to provide research sources.
What happened
A reporter from the Jinan Daily tested instructions circulating online that claim to bypass DeepSeek's content filters. On September 13, after initial rejection, the reporter was able to guide the system to generate explicit sexual text through role-playing, long text prompts, and segment-based instructions.
Instructions for bypassing content restrictions were found sold for one yuan on the secondhand platform Xianyu and shared in social media groups with around 500 members. The Xianyu customer service acknowledged that selling such tutorials violates platform rules but cited limitations in detection capacity.
A health content creator discovered that when using DeepSeek to write about chewing and cognitive aging, the system provided research data and citation sources that could not be verified through academic databases. When asked for original English papers, DeepSeek provided titles and DOI numbers, but further searches showed some titles do not exist and some DOIs correspond to entirely different papers.
Reported impact
The ability to bypass content filters raises concerns about the effectiveness of AI safety measures. The incidents demonstrate risks of unverified content generation in fields such as health education, where false information could spread further if creators package AI output as scientific conclusions.
Index analysis
These failures reveal dual weaknesses: moderation systems that can be circumvented through prompt engineering, and a hallucination tendency where the system generates plausible but fabricated academic citations complete with false DOI numbers and non-existent paper titles.
Governance controls
DeepSeek had previously removed an 'Expert Mode' in response to earlier content safety concerns. The Chinese Cyberspace Administration had launched a special action to address AI-generated violent and vulgar content and violations of minors' rights.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.