Research
In addition to the ability to abet harms in biology via synthetic DNA, there are other harms that might arise from AI via the chemical technologies. AI might assist or initiate the making of weapons, poisons and harmful drugs, and it might interfere with manufacturing, and it might interfere with R&D needed as part of emergency responses. We assess that the consequences of such events could be calamitous, and potentially disastrous, even where they would not be catastrophic or existential. This is the newest of our areas of work. Michael Marletta leads it.
Part of our present effort is to establish, defensibly, which harms arising from artificial intelligence applied to chemistry are worth protecting against, and in what order. We expect that such a list will largely need to remain closed, but we consider it useful that some entries should be public and widely understood.
Two will serve as examples. The first is the takeover of a chemical plant, assisted or initiated by an agent, for which the template is the 1984 explosion and release of methyl isocyanate (MIC) at the Union Carbide plant at Bhopal, which killed an estimated 25,000 people and disabled 500,000. The second is action by an AI agent, on a user's behalf or on its own, to arrange the contract synthesis of a novel or known small molecule of an addictive kind and its importation into the United States.
Another approach is to govern the chokepoints for chemical synthesis of DNA. The commercial synthetic DNA industry depends on phosphoramidite chemistry, and that dependence is the result of a series of historical contingencies: the chemistry developed in Marvin Caruthers's laboratory, Michael Hunkapiller's zero-dead-volume solenoid valve, and Leroy Hood's leadership in automating and commercialising the process. We are updating the published analysis of the phosphoramidite supply chain in order to identify the points at which regulatory intervention would hold.
We are also examining two alternative routes, phosphotriester chemistry and H-phosphonate chemistry, either of which might be used to evade screening of commercial synthesis orders. A control regime that addressed only phosophoamidites could be evaded by the others, and our purpose in examining them is to establish what such a regime would have to cover.
The purpose of the list of harms is to direct protective effort at the harms that matter, and to make a small number of examples public so that the people who ought to know of them know of this list. The purpose of the supply chain analysis is to establish where a control would actually regulate. A regime that addressed only phosphoramidite chemistry would be evaded by the two alternative routes, so an analysis that covers the others is a first step to determining whether any control regime will be worth imposing.