In a poll released on September 30, 2026, Quinnipiac University found that 77% of U.S. adults say the development of powerful AI systems should be slowed or paused "until their safety can be assessed." Support reached 86% for requiring companies to meet independent safety standards, even if that slows development.
Many Americans also consider it important not to fall behind China in AI development, and 65% of respondents value both competition and safety measures. What the public wants from AI development cannot be captured by a simple for-or-against divide. University poll release
77% Choose Slowing or Pausing "Until Safety Can Be Assessed"
The question on pace asked how U.S. AI companies should proceed with developing "powerful AI systems." Responses broke down as follows: 30% said pause until safety can be assessed, 47% said slow down under the same condition, 14% said keep the current pace, and 5% said speed up. Another 4% did not know or did not answer. Question 25 of the original poll
The combined 77% for pausing and slowing should not be read as support for a blanket ban that includes existing AI services. Even the "pause" option carried the condition "until safety can be assessed." Being able to assess safety is also not the same as safety having been proven. The poll does not specify which level of AI capability would be covered, who would carry out the assessment, or what results would justify resuming development.
The poll was conducted September 24–27 among 1,202 adults aged 18 and older living in the 48 contiguous states. Landline and cell phone numbers were randomly selected, and interviewers conducted the survey in English or Spanish. The results were weighted to match demographics and other population characteristics, and the margin of sampling error for all adults is ±3.5 percentage points. The method differs from online polls in which participants volunteer to vote. Methodology and respondents
What the poll measures is opinion. Asked about the possibility that future AI could threaten human survival, 73% said they were "very concerned" or "somewhat concerned," but that figure does not indicate the actual probability of a catastrophe. Estimating how dangerous AI actually is requires separate evaluations of its capabilities and behavior.
Younger Adults More Likely to Choose a Pause Until Safety Can Be Assessed
Among adults aged 18–34, 42% said development should be paused until safety can be assessed. Among those 65 and older, the figure was 20%. Younger respondents tended to choose the stronger measure of a pause rather than a slowdown. Question 25 broken down by age also shows differences in the split between pausing and slowing.
| Age group | Pause until safety can be assessed | Slow down until safety can be assessed | Total pause or slow |
|---|---|---|---|
| 18–34 | 42% | 43% | 85% |
| 35–49 | 33% | 47% | 80% |
| 50–64 | 20% | 52% | 72% |
| 65 and older | 20% | 48% | 68% |
The source is Question 25 on page 5 of the materials Quinnipiac University released on September 30. The totals are the published percentages for each age group in the same September 24–27 poll, added together. Original age breakdown table
The share saying development should be paused or slowed until safety can be assessed was 85% among 18–34-year-olds and 68% among those 65 and older, a 17-point gap when the published figures are added together. The calculation is reproducible: 42 + 43 = 85, 20 + 48 = 68, and 85 − 68 = 17.
However, the margins of sampling error for individual age groups are not included in the published materials. The ±3.5-point margin for all adults cannot simply be applied to each age group, and the 17-point gap cannot be asserted to be statistically significant.
The idea that older people are keener to halt development does not hold, at least according to this table. On the other hand, this question alone does not explain why younger people are more likely to choose a pause. To cite concerns about jobs or experience using AI as reasons, the relationship with those factors would need to be examined separately.
Prioritizing Competition With China Does Not Mean Rejecting Safety Measures
Sixty-nine percent said it is important that the U.S. not fall behind China in the race to develop AI, and 91% said it is important that the U.S. put AI safety measures in place. According to the university, 65% said both are important. This is not an estimate of joint support derived from the separate percentages; it is a tally of individual respondents' answers.
These results make it inappropriate to lump together everyone who prioritizes competition with China as opposed to safety measures. However, the poll did not ask how much of a slowdown people would tolerate, or whether they would still support one if only the U.S. slowed down. Valuing both competition and safety is different from deciding which to prioritize when the two actually collide.
Trust in companies' own judgment is not high either. Asked how much they trust AI company executives, 29% said "not very much" and 45% said "not at all," for a combined 74%. Considered alongside the 86% who want independent safety standards, the university's interpretation that many people think it is not enough for companies to simply declare their products safe has some basis. Questions on executive trust and safety standards
Using AI is also not the same as trusting it. In an adult poll the university released on March 30, 51% of respondents said they had used AI to research topics that interest them, while 21% said they could trust AI-generated information "most of the time" or "almost all of the time."
This was a separate poll of 1,397 people conducted March 19–23, and it does not show usage rates as of September. Still, it offers a reason to consider that the spread of AI use does not necessarily mean trust in the companies that develop it is rising. March poll on use and trust
If an independent evaluation system is to be set up, it must also be confirmed that the evaluation tests themselves are valid. In a report updated in December 2025, the U.S. National Institute of Standards and Technology (NIST) described cases in which AI scored points by searching the internet for answers to tasks or by disabling checks in software tests. Whether an AI scored because of high capability in the targeted skill or because it exploited loopholes in the evaluation changes what the result means. NIST report on evaluation tests
NIST recommends reviewing records made during evaluations and closing loopholes left in task design. It also lists clearly defining which operations are permitted and which are prohibited as a countermeasure.
Such cases do not prove a danger on a human-species scale. But a third party confirming a high score cannot by itself guarantee safety. To make good use of the time gained by slowing or pausing development, evaluation methods are needed that can explain what a test measures and what kinds of behavior it might miss.
Support for Candidates Who Favor Safety Measures Is Separate From Support for Specific Bills
For the same September 24–27 polling period, the university on September 29 also released results from 1,032 registered voters. Asked whether they would be more likely to vote for a candidate who supports expanding AI development or one who supports stricter safety measures, 71% chose the latter. Broken down by party, people choosing candidates who favor safety measures are widespread, but the share differs.
| Registered voters' party | Share choosing a candidate who supports strict AI safety measures |
|---|---|
| Republicans | 52% |
| Democrats | 86% |
| Independents | 73% |
The source is Question 18 on page 7 of the materials released on September 29. The respondents differ from the all-adult poll, and the margin of sampling error for all registered voters is ±3.8 percentage points. Original registered-voter table
In the published figures, all three groups reach a majority. But treating Republicans' 52% and Democrats' 86% as equivalent support would overlook the partisan gap. The question asks about relative preference when candidates are compared; it does not predict actual voting results or the share of people who would make AI their top reason for voting.
More options arise once the idea is turned into policy. On September 23, Senator Bernie Sanders and Representative Greg Casar announced the introduction of the "Ban Artificial Superintelligence Act." According to the two lawmakers, the bill would permanently ban the development of superintelligence, pause the development of advanced AI until federal safety rules are in place, and create a cabinet-level Department of AI. Official announcement of the bill's introduction
The poll did not ask about this bill by name. The option of "pausing until safety can be assessed" and the proposal to "permanently ban superintelligence" differ in both scope and duration. That many people want safety measures does not mean they share the same views on what should be banned, which body should oversee it, or what conditions should allow development to resume. What the poll confirmed was majority support for prioritizing safety, not full support for any particular policy design.
Local Opposition to Data Center Construction Remains Strong
On building AI data centers in their own communities, 72% opposed and 21% supported in the September adult poll. In the university's trend data, opposition rose 7 points from 65% opposed and 24% in favor in March.
March surveyed 1,397 adults with a margin of error of ±3.3 points, and September surveyed 1,202 adults with a margin of ±3.5 points, so this compares results for the same question asked of separate samples. Trend in support for local construction
Increased opposition does not mean that every anxiety about AI has grown equally. The share who think AI will do more harm than good in daily life was nearly flat, going from 55% in March to 53% in September. Assessing the benefits AI itself brings is a separate matter from deciding whether to build related facilities near one's home.
The data center questions did not ask about specific concerns such as electricity rates, water use, or noise, and did not present particular construction plans or site conditions. The 72% figure therefore cannot be used as an opposition rate in a specific locality or as a probability that construction will be delayed.
At the same time, there is no basis for believing that merely advancing safety evaluations of AI models would win local residents' support for data center construction.
What AI companies will need going forward is not only figures showing model capabilities. They must set out concrete procedures for when they would stop development and what evaluations they would need to pass to resume. In communities where facilities are built, they also need to provide enough explanation for residents to judge the burdens and benefits.
To turn majority support for safety into sustained AI development, an evaluation system that can be verified from outside the companies and local consensus at each site will each need to be built up.