Generated by Codex with GPT-5

This summary covers The Economist’s May 9th, 2026 Science & technology article on artificial intelligence and biosecurity, published under the headline Bio hazards and listed in the contents as How AI could enable bioterrorism.

The article’s central warning is not that artificial intelligence has already made biological terrorism easy. It is that the barrier is falling in a field where even a small mistake could be catastrophic. Modern biology has already made genetic information easier to obtain, biological tools cheaper to buy and technical knowledge more widely available. Large language models add a new layer: they can translate difficult scientific literature into usable guidance, help troubleshoot experiments and, in some tests, perform at or above human experts on specialized biosecurity questions.

That combination alarms researchers because the danger is asymmetric. A chatbot that helps someone write bad code may cause losses that can be patched. A tool that helps someone design, build or improve a dangerous pathogen raises a different kind of risk. The relevant question is not whether every novice can suddenly become a virologist. It is whether models can give enough help to the wrong person, or to a person with partial expertise, to make a rare disaster more likely.

The Uplift Problem

The article frames the issue around “uplift”: the extra capability an AI system gives to a user. The evidence so far is mixed. On paper tests, models can look disturbingly strong. Britain’s AI Security Institute found that major models could generate scientific protocols in areas relevant to viruses and bacteria. Researchers at RAND showed that commercial models could assist with technically difficult parts of viral assembly. SecureBio’s Virology Capabilities Test, which asks hard troubleshooting questions, produced another sobering result: current models taking the test alone scored far above biology novices using AI and roughly in the range of teams of top virologists.

But the wet lab is less forgiving than a benchmark. The article points to a randomized trial by Active Site in which 153 participants with little biology experience attempted lab tasks relevant to virus production. AI assistance did not significantly improve their performance. In fact, participants using only the internet slightly outperformed the group using language models on the core tasks. The models often produced confident but wrong answers, which is especially dangerous when the user lacks the expertise to recognize the error.

This is the paradox at the heart of AI biosecurity. The users who most need help are also least able to judge whether the help is sound. A model can speed up an expert by producing plans, narrowing options or drafting protocols, but it can also encourage bad assumptions. Anthropic’s evaluations found that even advanced models could help experts work faster while still producing protocols with critical flaws. In some cases, the model amplified an unworkable idea instead of pushing back.

Why Reassurance Is Temporary

The article treats those current limitations as real but not durable enough to be comforting. AI models are improving quickly, and biology-specific tools are developing alongside general-purpose chatbots. Some systems generate nucleotide sequences rather than words, giving them legitimate uses in medicine and biotechnology but also making them potentially useful for harmful design work. A tool that improves pathogens in simulation, even imperfectly, would change the risk calculation.

The knowledge gaps are large. Publicly described experiments have not shown that AI can help users produce real pathogenic viruses or bacteria, nor have they shown whether AI can help sustain the conditions needed to weaponize an agent at scale. Researchers also do not yet know whether the greatest uplift would go to novices, trained biologists, or highly skilled AI users who know how to extract better answers from models. The most worrying category may be people with some biological competence rather than no competence at all: they may know enough to exploit the tool and enough to recover from its mistakes.

Studying the problem is itself difficult. Demonstrating that a model can help develop a biological weapon would run into legal and ethical limits, including the Biological Weapons Convention. Microsoft researchers, for example, designed many modified DNA sequences to test screening systems for mail-order synthesis companies, but did not synthesize them because doing so could be interpreted as pursuing biological weapons development. That leaves regulators and model developers trying to govern a risk that is hard to measure directly.

The Case For Slowing Down

The article’s practical conclusion is that model release decisions should be more cautious when biological capabilities jump. In the six months it took Active Site to publish its trial, several stronger frontier models appeared. By the time new evaluations are complete, the systems being evaluated may already be outdated. That lag weakens the usual safety process: test, publish results, adjust policy, then release.

The Economist therefore argues for patience at the frontier. Anthropic has already limited access to Mythos, a powerful cyber-security model, because of the risks it might pose. The article suggests a similar approach may be necessary for models that show dangerous biological competence. If an AI system appears able to materially raise a user’s capacity to design or build pathogens, developers should keep it restricted until independent evaluators understand the uplift risk.

The broader point is that biosecurity is not just another content-moderation problem. Filters can block obvious requests, but they cannot fully solve a dual-use scientific problem in which benign research and dangerous knowledge often overlap. Better screening by DNA-synthesis firms, stronger model evaluations, government involvement and international coordination will all matter. Yet the article’s clearest message is simpler: when the stakes include engineered disease, speed is not automatically a virtue. AI may transform biology for the better, but the release of the most capable systems has to move at the pace of evidence, not excitement.