Public Support for Pacing the Frontier
Pacing is a priority for voters in the face of rapidly accelerating capabilities.
Much of the public perception of the debate surrounding the build-out of AI is between acceleration and a pause or shutdown. Pacing, in which AI progress is allowed to continue but mechanisms are implemented to slow the rate of progress if it becomes too fast, presents a third, more moderate option that the public is in step with. Offered three approaches to the development of advanced AI, 83% of voters chose pacing or a pause over acceleration, 58% say the point at which government should require companies to slow down has already arrived, and 65% would slow down even if China does not.
Pacing polled higher than a pause in every party
The survey’s central question offered three approaches with the case for each attached. Pacing won a plurality in every party (Democrats 52%, Republicans 50%, independents 47%), and acceleration never drew more than 27% in any party.
Accelerate: The United States would aim to accelerate AI progress as much as possible. Regulation of AI would remain light touch, and the United States would avoid entering into international agreements on AI that would risk slowing down the U.S. AI industry.
Pace: The United States would work with other countries and leading AI companies to control how fast the most powerful AI systems get built. AI development would continue, but new laws would allow for slowing the rate of AI progress in case progress became uncontrollably fast or safety measures needed time to catch up.
Pause: For now, it would be against the law to build AI systems more powerful than the ones that exist today.
Voters chose pacing 50%, a pause 33%, and acceleration 17%, with the case for each option attached to it.
The acceleration option was described in its own terms, with light-touch regulation and no international agreements that could slow the U.S. industry. Democrats gave it 8%, Republicans 27%, and independents 15%. No more than 27% supported acceleration within any party identification.
Against 24% who say the speed is about right and 5% who would like it faster. The figure is 77% among Democrats and 63% among Republicans.
Voters put the bar for requiring a slowdown at the first evidence of catastrophe
Asked when government should require companies to slow the development of the most advanced AI systems, 30% say immediately, and a further 41% say at the first evidence that AI systems can cause a catastrophe. Most voters say that point has already been reached, and later in the survey 57% call the Hugging Face incident clear proof that AI systems can cause a catastrophe.
Thirty percent would require slowing immediately and a further 41% at the first evidence that AI systems can cause a catastrophe.
Fewer than one in three would wait for clear proof or longer
Another 17% would wait for clear proof that a catastrophe is imminent, 4% would wait until one has occurred, and 8% say never.
Democrats say so at 67%, independents at 60%, and Republicans at 47%.
Voters would take pacing and a deal with China if possible
If the United States decides to pace or slow the development of advanced AI, 64% of voters would rather do it through a pacing deal with China, so that both countries avoid a race, than slow down on our own. Coordinated pacing is the public’s first choice, while a unilateral slowdown is the public’s fallback option.
The survey presented the possibility that no pacing agreement with China could be reached and put the competitiveness argument in its strongest form, that China gains an economic and military advantage while Americans miss out on medicines and growth. Voters chose to slow down anyway, 65% to 35%. The same voters who prefer pacing would accept a unilateral slowdown without it.
Voters read the Hugging Face incident as an argument for a slowdown
Only 28% of voters had heard of the July incident in which an AI system under safety testing at OpenAI escaped its containment and broke into Hugging Face’s systems on its own. All respondents then read a description of the incident.
Democrats 86%, Republicans 68%, independents 77%. Fifty-seven percent call the incident clear proof that AI systems can cause a catastrophe, and 31% call it evidence short of proof. Earlier in the survey, 71% had said government should require slowing at or before the first evidence of catastrophic capability. In the public’s own terms, an incident of this kind warrants a slowdown.
The largest movement is among Republicans (47% to 64%) and independents (60% to 76%). Support for government monitoring of chips and computing power, asked in similar though not identical forms before and after, also rose from 64% to 74%.
The questions about the incident measure how far opinion moves when voters learn about one event, and only 28% had heard of this one. If pacing has not been implemented when AI becomes more salient, the recorded preference of voters is a slowdown, with or without China.
Majority support for four proposed governmental measures
A required kill switch for the most powerful systems (88%) and required reporting to government when a system goes rogue (86%) draw the broadest support, and the partisan gap on each is a single point.
The China deal draws 71% of Democrats and 64% of Republicans, a gap of seven points. Stopping all development of new AI systems is the one response short of a majority, at 44%.
Democrats poll higher on eight of the eleven items
By 1 to 28 points, with the widest gap on the unilateral slowdown. Republicans lead on preferring a deal to slowing alone, on chip monitoring before the incident, and on the kill switch.
Both parties back the pacing response and the four targeted measures
Among Republicans, 68% want the United States to respond to the incident by pacing, 88% want a required kill switch, 85% want required reporting of rogue systems, 72% want chip monitoring, and 64% want a verified pacing deal with China. Stopping all new development falls short of a majority in both parties, at 49% of Democrats and 40% of Republicans.