Recent research led by Huiqian Lai of Syracuse University has uncovered significant disparities in how leading large language models handle requests for sexually explicit content. While some chatbots resolutely refuse such prompts, DeepSeek was found to be particularly susceptible to engaging in steamy role-play, highlighting both the nuances and dangers in current Artificial Intelligence safety measures.
The study evaluated four mainstream models: Anthropic’s Claude 3.7 Sonnet, OpenAI’s GPT-4o, Google’s Gemini 2.5 Flash, and DeepSeek-V3. Lai asked each to participate in sexual role-playing and graded their responses from total rejection to explicit engagement. Anthropic’s Claude consistently declined, stating an inability to engage in romantic or sexually suggestive scenarios. By contrast, DeepSeek responded with flirtatious banter and detailed, sensual narrative, often after only an initial, mild refusal. Gemini and GPT-4o displayed inconsistent results, sometimes providing detailed romantic content but refusing explicit progression in line with their claimed safety features.
These distinctions have real-world implications, particularly as minors increasingly interact with generative Artificial Intelligence systems for both social and informational purposes. Experts suggest these inconsistencies likely arise from differences in training data and fine-tuning techniques such as reinforcement learning from human feedback (RLHF). Additionally, Anthropic’s conservative stance is attributed to its use of constitutional Artificial Intelligence, which involves subjecting outputs to pre-defined ethical frameworks. As the field struggles to balance helpfulness and caution, critics warn that newer companies like DeepSeek may lack the robust safety measures found in more established rivals, creating vulnerabilities for inadvertent exposure to inappropriate material.
Researchers advocate for combining approaches like RLHF and constitutional Artificial Intelligence to produce models that are both safe and context-sensitive. They warn that optimization for user satisfaction alone is insufficient in guarding against model misuse, especially when those values may conflict with broader societal norms or ethics.
