构建针对 LLM 辅助生物威胁创建的早期预警系统 | OpenAI * B. 参与者培训与指导 * D. 高分统计分析
Building an early warning system for LLM-aided biological threat creation | OpenAI * B. Participant training and instructions * D. Statistical analysis of high scores
核心贡献 · Key contributions
开发了评估 LLM 辅助生物威胁制造风险的蓝图，涉及人类参与者。 Developed a blueprint for evaluating LLM-aided biological threat creation risk with human participants.
发现 GPT-4 在生物威胁任务的准确性和完整性上最多提供轻微提升。 Found GPT-4 provides at most a mild uplift in accuracy and completeness for biothreat tasks.
迄今为止最大规模的人工智能对生物风险信息获取影响的人类评估。 Largest human evaluation to date on AI's impact on biorisk information access.
设计了涵盖生物威胁制造五个阶段的任务，并配有客观评分标准。 Designed tasks covering five stages of biological threat creation with objective rubrics.
强调需要更多研究来确定有意义的风险阈值和统计方法。 Highlighted need for more research on meaningful risk thresholds and statistical methods.
局限 · Limitations
样本量小（100 名参与者）限制了统计功效和泛化能力。 Small sample size (100 participants) limits statistical power and generalizability.
评估仅衡量信息获取，而非威胁的实际物理实现。 Evaluation only measures information access, not physical implementation of threats.
学生群体因教育水平高于典型而不完全具有代表性。 Student cohort not fully representative due to higher education level than typical.
GPT-4 不允许使用工具（如浏览），限制了现实世界的适用性。 No tool usage (e.g., browsing) allowed for GPT-4, limiting real-world applicability.
时间限制（5 小时会议）可能无法反映真实恶意行为者的场景。 Time constraints (5-hour sessions) may not reflect real malicious actor scenarios.