We really hate to do this on a Friday morning, but regrettably we need to inform you about some more troubling developments in AI land.
What happened: Anthropic reported that it had blocked people — including state actors — using its AI tools for a wide range of nefarious purposes, including developing conventional weapons, spying on political dissidents, and conducting research that could be used to develop biological weapons.
In one instance, Anthropic said it blocked efforts by users in Russia to build the software to power an autonomous “kamikaze drone swarm.”
Perhaps more disturbing were several reported cases related to bio-weapons, including one in which a scientist outside the U.S. used Claude to research ways that the bird flu could be made more transmissible and dangerous to humans. In this case, the researcher also took steps to hide their identity and circumvent geographic restrictions.
Catch up: The revelations add fuel to fears about the risks posed by AI, long held by many in the AI world but which broke into the mainstream this week with the resignation from Anthropic of Jacob Coxon, who warned that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
That was followed yesterday by similar apocalyptic warnings from Paul Christiano, an ex-OpenAI researcher and current member of its foundation’s board, as well as Nobel Prize-winning computer scientist Geoffrey Hinton.
Why it matters: While these predictions that superintelligent AI could soon turn against and wipe out its creators have triggered alarm, this report calls attention to the more familiar threat of humans using new technology against each other.
Bottom line: Unlike the risks presented by yet-to-be-built superintelligent AI, it appears that today’s actually existing AI models are — in the wrong hands — already capable enough to inflict real damage.—TS
