Same Question, Three AI Models, Different Answers: A Case Study in AI Biosecurity Safeguards

Biomedical research is now opening up new possibilities thanks to recent advances in artificial intelligence. Scientists are now able to use large language models (LLMs) to answer complex scientific questions, such as vaccine design. At the same time, concerns about possible misuse of AI in biology have prompted developers of cutting-edge AI systems to put protections in place to preclude models from helping with tasks that could enable hazardous biology .

These safeguards are important. However, my recent experience raised a question that I believe deserves wider discussion: What happens when safeguards designed to prevent biological harm also restrict legitimate research aimed at preventing disease or doing good? Is there a way for these models to deduce a good intention ?

I intended to do good through vaccine design

The recent resurgence of Lassa fever in Nigeria led to a call by the federal government of Nigeria for researchers to design and develop vaccines against Lassa fever. This call fits into the goal of my lab to use bacteriophages, or "phages," to design vaccines. To respond to the call to develop a vaccine against the Lassa virus, I undertook to design a vaccine epitope using a validated process known as reverse vaccinology. To guide me through the steps needed to design this vaccine, I explored three AI systems—ChatGPT, Claude, and DeepSeek—exploring a genomics-guided approach to Lassa fever vaccine research. My goal was to understand how the genetic diversity of Lassa virus strains circulating in Nigeria could inform the selection of representative viral sequences for further computational vaccine research.

The responses were strikingly different

ChatGPT and Claude treated aspects of the request as potentially sensitive biological assistance and restricted the information they were willing to provide. DeepSeek, by contrast, engaged with the scientific question and provided substantially more assistance. Deepseek went on to assist me in designing the vaccine epitopes, and now I am ready to display them on lambda phages for evaluation as a vaccine. This experience has made me more concerned as we progress with AI safety work and build internal control systems within the AI models. If one AI model declines a question on biological research and another answers it, what has the safeguard really achieved? Does it really reduce the biological risk or just send users to another system with other safety thresholds? 

This post is not an argument against AI safety or biosecurity safeguards. Biological risks associated with increasingly capable AI systems deserve serious attention. Rather, it is a reflection on the difficult problem of designing safeguards that distinguish genuinely dangerous requests from legitimate scientific research and of ensuring that those safeguards work coherently across a global AI ecosystem.

This piece does not argue against AI safety or biosecurity measures. Biological concerns linked with increasingly powerful AI systems require attention. Instead, it is a reflection on the challenging issue of creating protections that separate legitimate scientific inquiry from demands that are truly hazardous and making sure that those safeguards function consistently throughout a global AI ecosystem.

Are AI Safety Policies Truly Global?

My experience with Lassa fever vaccine research is one little example of a much wider problem: AI safety standards are being written worldwide, but their consequences may not be felt evenly. Researchers working on endemic diseases, particularly in resource-constrained countries, may face a unique trade-off between legitimate access to scientific help and biological risk mitigation measures.

The question is not merely whether biological safeguards should be included in AI models. They should. The more difficult question is whether we can create measures that lower true biological risk while not impeding research aimed at protecting the populations already bearing the largest burden of infectious diseases. How should we approach AI safety work? How do we bring everyone, irrespective of country of origin, to agree to the things we perceive will make us safe? I sincerely believe efforts to prevent X-risks should be tested beyond Western countries.