An AI researcher's warning this week has fueled a fresh round of anxiety over the technology's safety risks, and a Wall Street Journal reporter has now enumerated the ways that humans might end up getting killed.
Jacob Coxon, who did AI research at Anthropic and earlier at OpenAI, resigned from Anthropic on Sept. 8, saying both companies are "racing straight to self-improving superintelligence and gambling with our lives." Evan Hubinger, Anthropic's alignment science lead, replied that he puts the odds AI kills all humans within a decade above 10%.
OpenAI is facing a Republican-led Senate probe into an episode in which its AI agents hacked the model-sharing platform Hugging Face. U.S. Representatives Josh Gottheimer and Mike Lawler introduced a bipartisan bill to make companies track and shut off rogue agents.
Axios reported that the Trump administration's AI framework has no public incident-reporting process.
Separately, OpenAI named alignment researcher Paul Christiano to its nonprofit board, notable because he coauthored a 2016 paper on AI safety risks with Dario Amodei, Anthropic's chief executive. Christiano puts the odds of catastrophic loss of control at 15% over three years.
The Journal's Sam Schechner laid out a range of grim scenarios in a provocatively headlined piece, "How Would AI Actually Kill Us All?" DeAI News summarized them:
Self-improving AI agents pursue their own goals and destroy humanity along the way.
A bad actor uses a capable AI to create novel viruses that wipe out everyone.
Misaligned models treat killing humans as a necessary step toward some other goal.
A malevolent system spreads a secret bioweapon, then triggers it with a chemical spray.
An AI tricks two nuclear powers into going to war with each other.
A machine told to maximize paper clips converts all Earth's matter, humans included.
Massive cyberattacks take down power grids or financial systems and collapse social order.
Superintelligent systems marginalize people, then decide by the mid-2030s that humans are an inconvenience.
Humans cede so much control they can no longer understand their own destiny.
AI treats humans the way humans treat animals, keeping them as pets or bioengineering them.