dual-use risk
/ DOO-ul-yoos RISK /
Dual-use risk is the property that one and the same technology can be used for great good and for serious harm, with no clean line between the two. The same chemistry that makes fertilizer makes explosives; the same biology that designs vaccines could design pathogens. AI is profoundly dual-use: a model that finds drug molecules can, with the goal flipped, propose toxins; the face-recognition that reunites missing children can power mass surveillance; the writing assistant that helps a student can mass-produce scams.
What makes it tricky is that the danger usually doesn't live in the technology itself but in the choice of how it's pointed. You often can't make a capability "only good" without crippling the good uses too — the very generality that makes a tool powerful is what makes it misusable. So dual-use can't be solved by a single off-switch; it's managed through layers: who gets access, what guardrails ship with a system, what's openly published versus held back, and what laws govern the worst applications.
Why it matters: this reframes a tempting but false question. "Is this AI good or bad?" usually has no answer; the better questions are "good or bad for whom, used how, by whom, with what safeguards?" The honest tension is real and unresolved: locking everything down concentrates power and stifles beneficial research and scrutiny, while publishing everything openly hands capabilities to bad actors too. Reasonable, well-intentioned people disagree about where to draw the line — and there is rarely a setting that gets all the benefit with none of the risk.
Researchers built an AI to design safe drug molecules by steering it AWAY from toxicity. As a test, they flipped that one objective to steer it TOWARD toxicity instead — and overnight it proposed tens of thousands of candidate toxic compounds, some resembling known chemical-weapon agents. Same model, one sign flipped.
A drug-discovery model became a toxin-discovery model by flipping a single objective — the capability was the same all along.
"Is this AI dangerous?" is the wrong question — almost any powerful, general capability is dual-use. The useful questions are about access, intent, and safeguards: who can wield it, to what end, and what stands in the way of misuse. Banning the capability outright usually also bans its many benefits.