At a Hong Kong science symposium on November 7, 2026, Tencent chief AI scientist Yao Shunyu warned that AI models’ biological capabilities could reach a threshold where harm becomes possible, according to a report by the South China Morning Post published the next day. His warning described a possible future risk.
The concern has both beneficial and harmful sides
The South China Morning Post placed Yao’s warning in the context of AI’s potential to help accelerate drug discovery, alongside concerns about malicious biological uses. That dual-use tension matters: the same broad field can support research with medical aims while raising safety questions.
Treat a biological AI prediction as a research output, not as a real-world outcome; physical evaluation is needed to assess its effects.
Chatbots and biological design tools play different roles
A Fall 2023 framework from Georgetown University’s Center for Security and Emerging Technology (CSET) distinguishes general chatbots from biological design tools, or BDTs. Chatbots may make it easier to gather and organize information. BDTs are specialized tools that can help scientifically knowledgeable users generate or prioritize predictions about biological molecules or systems. CSET’s explainer describes these as different pathways, with the possible contribution depending on the tool and the user’s expertise.
CSET’s framework says successful use of BDTs at the time it assessed them required scientific and programming expertise. It also describes biological misuse as possible without AI, so AI is one potential factor in a risk pathway rather than a necessary ingredient in every case.
A digital prediction alone does not establish what will happen in a physical setting. CSET says predictions need to be produced and evaluated physically before their real-world effects are known, and that results can differ across biological contexts. Its framework reflects the state of biological AI tools in Fall 2023.
What the GPT-4 comparison says
In a 2025 analysis, the Centre for International Governance Innovation (CIGI) reported that a team assembled by OpenAI found GPT-4 offered only a marginal advantage over ordinary internet searches for the bioweapon-building tasks it assessed. That finding is specific to the tasks in that assessment; it does not provide a general measure across models or uses.
