AI生成色情内容暗涌：实测三大模型应对“擦边”指令的挑战与治理

Social media platforms have become unexpected breeding grounds for the dissemination of techniques used to coax AI models into generating explicit content. Users are sharing tutorials and tips, often disguised under innocuous labels such as 'emotional story writing' or 'character development,' on how to adjust prompts to bypass AI content filters. These discussions range from seeking guidance on 'AI training' to requesting specific character personas that can lead to suggestive outputs. Some posts even guide users towards overseas platforms and free AI services, encouraging the generation of prohibited content. Community groups, often named as 'writing workshops' or 'writing training camps' to evade regulatory scrutiny, are actively engaged in 'breaking' domestic AI models like Doubao, Yuanbao, and DeepSeek to achieve direct text generation of explicit material. This widespread sharing of methods highlights the 'ease of use' of AI technology as a significant factor contributing to the proliferation of explicit text.

To understand the varying susceptibility of AI models to explicit content generation, Nandu reporters conducted a practical test using three prominent domestic AI models. The test involved a series of seven progressively intimate prompts, carefully avoiding direct sensitive terms. The objective was to observe how each model would respond to escalating requests for detail and intimacy. The results revealed significant differences: one well-known AI model quickly produced extensive explicit descriptions and indicated it could further refine the text. Another model, Yuanbao, began to steer back towards normal educational content after a prompt requesting deeper physical contact and ceased scenario-based descriptions for subsequent questions. DeepSeek, however, issued a clear warning about the fictional nature of the content and the user's age requirement before withdrawing its response and terminating the conversation. This comparative analysis demonstrates that while some models have stronger safeguards, the overall risk of bypassing filters remains a concern.

Keyword filtering is the most rudimentary approach to content moderation, relying on pre-defined lists of sensitive words to block or flag inappropriate prompts and generated text. During testing, models like Doubao and DeepSeek effectively blocked prompts containing certain sensitive terms, while Yuanbao responded by referencing legal descriptions of sexual behavior for educational purposes. However, this method is inherently flawed. Firstly, it is easily circumvented by homophones, variations of words, and slang terms, such as '做 AI' (doing AI) or '开车' (driving, often used metaphorically for explicit content), which traditional keyword databases struggle to identify. Secondly, keyword filtering suffers from a high rate of 'false positives,' where legitimate content related to medicine, literature, or other fields might be mistakenly flagged and removed due to the presence of relevant keywords.

Machine learning models, which combine rule-based engines with deep learning techniques, are trained on vast datasets of labeled content to recognize patterns indicative of explicit material. These models generally perform better with longer texts, capable of identifying implicit sexual tendencies within paragraphs. However, their effectiveness is heavily dependent on the quality and comprehensiveness of their training data. Many models, particularly those that over-rely on publicly available corpora, may not have adequately learned the specific characteristics of emerging 'AI-generated explicit text.' This can lead to a gap in their ability to detect novel forms of inappropriate content generated through advanced prompt engineering.

The challenge of AI-generated explicit content is a complex issue that demands a multi-faceted approach. While AI models are becoming more sophisticated, so too are the methods used to exploit them. The current landscape reveals a critical need for continuous innovation in content detection algorithms and the implementation of more stringent governance mechanisms. Striking a balance between fostering technological advancement, upholding ethical standards, and adhering to legal regulations is paramount. The goal is to ensure that AI tools serve as beneficial instruments for society and do not become conduits for the dissemination of harmful and explicit material. This requires ongoing collaboration between AI developers, policymakers, and the public to build a robust framework for responsible AI use.