AI-Driven Biology: The Promise and Peril of Automated Science

Recent developments show artificial intelligence (AI) systems capable of autonomously designing and executing biological experiments, while regulatory frameworks remain underdeveloped.

Key Developments

In February 2026, OpenAI and Ginkgo Bioworks announced that OpenAI's GPT-5 model autonomously designed and executed 36,000 biological experiments using a robotic cloud laboratory. The process involved AI proposing study designs, robots executing them, and data being returned to the model for iterative refinement. Humans set the initial research goal, and machines performed the majority of laboratory work. This resulted in a 40% cost reduction for producing a specific protein.

Ginkgo Bioworks, founded by former Massachusetts Institute of Technology graduate students, operates an autonomous laboratory where robots carry out scientific experiments. The company uses AI to translate experimental designs into instructions for robotic systems. Co-founders reported the system ran over 30,000 experiments over six months.

Technological Framework

This approach, referred to by some as "programmable biology," involves designing biological components through digital means and constructing them physically. AI completes an experimental loop that traditionally required direct human intervention at each stage. The method follows a design-build-test-learn cycle, allowing parallel exploration of thousands of design variations.

Protein design represents a clear application. Protein language models, trained on millions of natural protein sequences, can predict mutation effects or design new proteins. When combined with automated laboratories, these models create rapid experimental and revision cycles, completing tasks in days that previously required months or years.

Risk Assessment

The technology presents a dual-use problem, where tools developed for beneficial purposes could also be used for harm. Researchers have found that AI models integrated with automated labs can optimize viral spread without specialized training. Risk-scoring tools evaluate how AI could modify viral capabilities, such as altering host species or evading immune systems. AI models can guide users through technical steps to recover live viruses from synthetic DNA.

Research on whether AI can assist individuals with limited biology training in conducting dangerous laboratory work has produced mixed findings:

• A study by Scale AI and SecureBio indicated that novices using large language models completed biosecurity-related tasks with four times greater accuracy and sometimes outperformed trained experts. Approximately 90% of these novices reported minimal difficulty obtaining risky biological information despite safety filters.

• A study led by Active Site found that AI assistance did not significantly improve novices' ability to complete complex workflows to produce a virus in a biosafety laboratory. However, the AI-assisted group had higher success rates and faster completion times for certain steps, such as cell growth.

Regulatory Landscape

Existing regulations for biological research do not account for AI-driven automation, and AI regulations do not specifically address biological applications.

In the United States, a 2023 executive order on AI security that included biosecurity provisions was revoked. Screening of synthetic DNA by commercial providers remains largely voluntary. A 2026 bipartisan bill to mandate DNA screening does not address AI-designed sequences that might evade detection. The 1975 Biological Weapons Convention lacks AI provisions.

Safety evaluations conducted by AI labs are described by some researchers as opaque and insufficient for real-world risk assessment. Researchers estimate that improvements in AI's ability to plan pathogen-related experiments could lead to thousands of additional deaths from bioterrorism annually.

Proposed Responses

Proposals to address these risks include:

• A managed access framework for biological AI tools
• Improved DNA synthesis screening
• Model evaluations before release
• Governance of biological data, particularly genomic data

Some AI companies have voluntarily implemented safety measures. Industry leaders acknowledge that AI development pace may soon exceed individual company assessment capabilities.

Statements from Principals

"I saw a lab notebook entry written by the model and it was really, really wild. We're now spending more time designing experiments for robots to execute overnight."
— Reshma Shetty, Ginkgo Bioworks co-founder

"I foresee a day when scientific practice becomes democratized."
— Jason Kelly, Ginkgo Bioworks co-founder

"AI could enable people with little training to run experiments with questionable goals, including bioweapons. We must prioritize regulations to mitigate the risks of potential mass production of viruses."
— Drew Endy, Stanford University

Unanswered Questions

AI can facilitate research goals in controlled environments. The implications of these capabilities operating outside controls remain an open policy question. Overreaction risks diverting talent and investment, while under-reaction could allow exploitation of the technology for harm.