A fresh warning that artificial intelligence can speed up the design of biological and chemical weapons has reopened the question of whether AI development should be slowed, though the experts closest to the risk do not agree on the answer.
The evidence
MIT Technology Review, the magazine published by the Massachusetts Institute of Technology, argued this week that AI tools now make it easier to design and troubleshoot dangerous agents, opening new routes to misuse.
Its examples are concrete rather than hypothetical.
In 2022 a drug-discovery model built to find medicines was rerun to hunt for harmful compounds, and it produced 40,000 toxic molecules in under six hours.
Anthropic, the AI company behind the Claude assistant, has since disclosed that people tried to use its models to explore modifying the chikungunya virus and other biological targets, and it now calls biological misuse one of the most serious risks of advanced AI.
The case for caution
Some researchers want the industry to pull back.
Kevin Esvelt, a biologist at MIT who studies both gene-editing technology and ways to contain it, urged companies on X to "err on the side of caution", saying a model had described a form of bioweapon he had not realised was possible.
His call echoes Dario Amodei, Anthropic's chief executive, who has argued that AI progress should be paced, a view Sam Altman of OpenAI has publicly shared.
The case against panic
Others say the headline fear is overblown.
Esvelt himself rates the chance of an AI-driven pandemic as low, while arguing that the potential scale still justifies serious effort to reduce it further.
Wendy Barclay, a virologist at Imperial College London, notes that the biggest pandemic threat today comes not from a bioweapon but from pathogens already circulating, such as H5N1 bird flu.
Canvasses of biosecurity specialists find broad agreement that the near-term risk is real but narrow, and little support for the scenario of a rogue system engineering a species-ending virus.
Do the safeguards work?
The defences that exist are only partial.
Companies screen orders for synthetic DNA, probe their own models by attacking them, a practice known as red-teaming, and build filters to block dangerous requests, yet experts warn these can be evaded and that AI could itself help find the gaps.
That is why much of the argument has moved from whether to slow AI towards what to build around it.
The answer emerging from the coverage is less a halt to development than stronger screening, tighter access to biology-capable models and better public-health preparedness.
So the honest response to whether this is a reason to slow down is that most of those raising the alarm are asking for something narrower: controls and monitoring that keep pace with the technology, rather than a brake on it.
