OpenAI says there's only a small chance ChatGPT will help create bioweapons

ai-safetybioweaponsopenairisk-assessmentgpt-4

Abstraction: OpenAI self-study finds GPT-4 gives marginal bioweapon research uplift

Key points:

OpenAI preparedness team study: 100 participants (50 biology experts, 50 students) tested GPT-4 vs. regular internet on 5 bioweapon research tasks including synthesizing Ebola
GPT-4 group had slightly higher accuracy scores but the increase was deemed "not statistically significant" — though a footnote admits aggregate total accuracy advantage was statistically significant
Students using GPT-4 reached expert baseline on 2 specific tasks (magnification and formulation); task details withheld due to "information hazard concerns"
GPT-4 responses were consistently longer with more task-relevant details
Study criticized for conflict of interest (self-conducted by OpenAI) and for potentially understating capabilities
Preparedness team also studying AI risks for cybersecurity and influence/belief-change operations

Connections: Openai · GPT-4 · AI Safety · AI Risk Assessment