Americans increasingly say AI models are behaving in unintended ways
Following reporting two weeks ago on the hacking of HuggingFace by OpenAI models, concern about AI’s impact on society jumped, and more people than ever are saying AI models have gone rogue.
In a survey that went into the field two days after reporting on the incident broke:
72% of Americans– nearly three quarters – say they’re somewhat or very concerned about AI’s growing impact on society, an increase from 66% in late June and early July. Within that,people who say they are very concerned increased from 28% to 35%. The number of people saying that AI systems are behaving in ways their developers didn’t intend increased from
. The shift was mostly attributable to people moving from saying this was28% in late June to 35% earlier this weeklikely to happento saying it’salready happening.
These shifts are both outside the cross-wave margin of error, and we believe this shows that the spike in concern is correlated with news of the hacking incident.
Our team is assembling briefing materials on these findings, and if you or your team work on AI policy issues, feel free to reach out to us for more information.
The survey of 1,815 American adults was conducted from July 23 to 28 with NORC at the University of Chicago’s AmeriSpeak panel.
Other Results
Other highlights from this week’s survey, including movements within the margin of error, but that we’re watching:
↑
Resignation: 72% say that AI is coming into their lives whether they want it or not, compared to 13% who say people like them can shape how much AI is a part of their lives. This increased 1 percentage point from the previous survey.↓ Productivity: 44% say people are already seeing people save time on everyday tasks due to AI, or expect to soon. This decreased 4 percentage points from the previous survey.↓ Job insecurity: 76% say AI is making job security worse. This decreased 3 percentage points from the previous survey.↑
Action on data centers: 46% want government to do more about AI’s environmental footprint, including from data centers, compared to 24% who say government is doing too much. This increased 3 percentage points from the previous survey.↓ Likelihood of WMDs: 67% say AI being used to develop weapons of mass destruction is likely to happen or already happening. This decreased 2 percentage points from the previous survey.
Other News
Five interesting insights from recent polls:
Americans want the US to lead: In a June survey,Pew Research Centerfound that 77% of Americans feel that it’s extremely or somewhat important for the U.S. to be the leader in AI development, but only 12% believe that the U.S. is currently beating China in the AI race.
Increasing opposition to data centers: New studies fromPOLITICO/Public First,Fox News, andEmerson College Pollingall found majorities or pluralities of Americans across party lines opposing the construction of data centers in their communities by large margins (24%-41%, 30%-70%, 27%-63% support-oppose). Earlier in January, POLITICO/Public First polled the same question and found 37%-28% support-oppose.
Digging deep on data centers: A poll fromKSTP/SurveyUSAfound that the data center opposition in Minnesota was driven largely by concerns about resources as well as general concerns about AI, more than costs or pollution.
Young adults increasingly negative on AI: A recently published survey fromBentley University-Gallup Business in Society(fielded mid-May) found young adults becoming the most negative on AI, a finding that was echoed byFox News’s survey.
Concern about AI’s impact on jobs remains high: TheBentley University-Gallup Business in Societyalso found that 79% of Americans see AI negatively impacting jobs in the U.S., nearly the same number we found in our survey (76%), which has not moved significantly in the past 6 weeks.
Experts and public concerned about catastrophic risks: Experts surveyed about various catastrophic harms by AI byMITsaid the likelihood of catastrophic harms by 2030 from AI possessing dangerous capabilities is 21.5%, and from cyberattacks or weapons development is 21%. The public is also concerned: in the most recentAmericans on AIsurvey, 67% of Americans said they believe AI is already being used to create weapons of mass destruction–or will be in the future.