When AI Accelerates Research, Who Governs the Risk?
Artificial intelligence is becoming a supporting tool in biological research. It can help make some research and analysis activities more accessible, with potential benefits for medicine and public health. The same capabilities, however, can have different uses and raise security concerns.
This balance is at the center of part of the debate among model developers. Dario Amodei, CEO of Anthropic, has discussed the risks of misuse and the need to combine technical safeguards with public measures. OpenAI, in turn, describes assessments and controls for the biological capabilities of its models. These are relevant positions, but they come from companies that build and distribute these systems: they should be read as statements by the respective developers, not as independent assessments.
A scientific opportunity with potential dual uses
In biology, AI can help organize information and support research activities. OpenAI cites drug, vaccine, and new therapy research among its potential uses. At the same time, it acknowledges that similar capabilities could be misused and that current assessments cannot predict with certainty whether a real-world event will occur.
This dual nature makes the issue more complex than a choice between “good AI” and “dangerous AI.” What matters includes a system’s level of capability, who can access it, the context in which it is used, and the controls that govern access. Even legitimate research may require safeguards when knowledge has more than one possible application.
Safeguards need to work on multiple levels
Providers can train models to refuse harmful requests, monitor suspicious use, and assess systems before deployment. OpenAI describes red-teaming activities, security checks, and collaboration with external experts. Anthropic has published examples of activity detected on its platform and notes that automated classifiers alone cannot always distinguish legitimate research from risky use.
These measures can reduce some risks, but they do not resolve the issue on their own. A control may block an explicit request without capturing the full context; an assessment in a controlled environment is not a certain prediction of what will happen outside the lab. Safeguards should therefore be treated as a set of measures: capability assessment, access rules, monitoring, incident response, and periodic review.
The question of responsibility
A central question is who should define and verify these controls. Companies know their models and can intervene in how they behave; independent researchers and public institutions can provide checks and different perspectives. Transparency about assessment methods and limitations helps make the debate more concrete, while protecting details that could make safeguards easier to circumvent.
There is also a question of incentives. If adopting safeguards involves costs while competition rewards speed, relying only on one company’s voluntary decisions may leave gaps. Shared standards, independent oversight, and proportionate rules can help distribute responsibility more effectively. Which measures are appropriate depends on the system’s capabilities and the available evidence: there is no formula that works for every model and context.
What businesses can learn
The debate about biological security concerns specific cases. It does not mean that every business use of AI carries the same level of risk. It does, however, offer a practical lesson in governance:
– define which tasks can be delegated to AI and which require human assessment;
– assign clear responsibility for checking outputs and handling errors;
– review outputs before using them in decisions or customer-facing content;
– choose tools and access levels based on the data involved and the consequences of an error;
– revisit rules and procedures as systems and use cases change.
These are general criteria, not a security certification or a guarantee that errors will not occur. Their value lies in making clear who oversees the system, for what purpose, and within which limits.
Governing AI is part of the design
The race for AI is often described as a race to build greater capabilities. For businesses, institutions, and individuals, it also matters how those capabilities are distributed, assessed, and governed. In biological research, developers’ sources describe potential applications and risks that call for caution; on their own, they do not establish how likely any specific scenario is.
A responsible discussion therefore starts with questions that can be checked: What uses are intended? What limits are in place? Who reviews the outputs, and what happens when a problem emerges? The answer cannot depend solely on a model’s promise or technical capability. Processes, controls, and responsibilities need to match the impact of the use.
