Americans picked AI safety over winning the AI race, 73 to 23

01

WHAT ACTUALLY CHANGED

Reuters published the new Reuters and Ipsos poll on September 22. The survey ran for four days and ended Sunday, September 20, with responses from 1,277 US adults nationwide.

The headline choice produced a wide split. Seventy-three percent said ensuring safe and responsible AI development was more important than keeping the United States ahead of the rest of the world. Twenty-three percent chose staying ahead.

A different question also produced 73 percent. That share said AI companies had not gone far enough to prevent AI from causing serious harm to society. The repeated number is a coincidence across two questions, not one combined measure.

Fifty-five percent said slowing AI development would be a good thing. Thirteen percent said it would be a bad thing. The remaining 32 percent were unsure or did not answer, according to Reuters.

Thirty-nine percent said AI is having a negative effect on society. Reuters says that was up from 36 percent in August and was the highest negative reading since Reuters and Ipsos began asking the question in March. Eleven percent called the effect positive. Half were unsure or did not answer.

Sixty-nine percent said they had followed news about calls by major AI companies for slower development, stronger oversight and safety standards. Following a debate does not establish detailed knowledge of a specific model, incident or proposal.

Most respondents said federal officials should play a major role in setting safety standards, Reuters reported. The published article does not provide the exact percentage or complete response distribution for that item.

Reuters reports a margin of sampling error of plus or minus 3 percentage points for the full sample. That means the safety-over-race gap remains large even at the ends of the stated interval, while smaller changes and subgroup differences require more caution.

The survey was conducted online, but it was not simply open to anyone who found a link. Reuters explains that Ipsos recruits its KnowledgePanel through postal address-based probability sampling and provides internet service and a laptop to selected households that lack access.

Ipsos weights the collected responses to population benchmarks including age, gender, race and ethnicity, education, household income and census region. Political composition may also be adjusted using party identification or reported 2024 vote choice. Weighting can reduce known imbalances. It cannot remove every source of nonresponse, measurement or wording error.

02

WHY THIS MATTERS

The race metaphor asks people to accept speed because somebody else might move faster. This poll suggests that the metaphor loses when safety and responsibility are offered as the competing priority. Public legitimacy may depend less on who claims first place and more on whether the rules can produce evidence before deployment.

A preference is not a probability. Seventy-three percent concern does not mean respondents estimate a 73 percent chance of catastrophe. It says a broad public sample chose one statement or agreed with one judgment under the wording and context of the survey.

The two 73 percent results point in the same general direction but measure different things. One compares safe development with national leadership. The other judges whether companies have done enough to prevent serious harm. Combining them would manufacture precision that the questions did not provide.

The large unsure groups are part of the result. When half of respondents do not classify AI as positive or negative, and nearly a third do not call slowing good or bad, the public is not speaking with one detailed voice. People may be cautious, unfamiliar, conflicted or waiting for better evidence.

Question wording carries political freight. Serious harm can mean fraud, job loss, discrimination, cyberattack, surveillance, unreliable medical advice, military escalation or human extinction. Safe and responsible development can mean voluntary testing, licensing, product liability, sector rules or a temporary pause. A broad phrase can unite people who would divide over the policy underneath it.

The method matters because online does not automatically mean opt-in. KnowledgePanel starts with randomly selected postal addresses and tries to include households without internet access. That supports national inference more credibly than a social-media poll, while the final estimates still depend on participation, weighting and accurate answers.

The stated margin of sampling error addresses uncertainty from observing a sample instead of every adult under the survey design. It does not measure every possible error. Poor wording, misunderstanding, nonresponse, inaccurate recollection, last-minute news and the order of questions can move a result without appearing inside the three-point interval.

The trend from 36 percent negative in August to 39 percent in September is exactly the size of the stated overall margin. That does not prove there was no change, but it is too small to present as a clean one-month shift without the complete estimates, design information and repeated measurements.

The poll covers adults, not only voters. That is appropriate for a question about social policy. It also means the result is not a forecast of a ballot, the composition of Congress or the position of officials who will write the rules.

For laboratories, the useful reading is not that marketing failed. It is that evidence needs to arrive before trust. A release should name the risks tested, the thresholds, the outside evaluators, the failures, the fixes, the remaining limits and the person who had authority to say no.

For lawmakers, broad concern creates room to govern but not a blank check. A credible policy has to state which systems it covers, who measures capability, which incidents must be reported, how trade secrets and public evidence coexist, what happens after failure and how smaller developers can comply.

For journalists and advocates, the temptation will be to treat a large number as a mandate for a favored plan. The honest next step is more specific questioning. Ask separately about incident reporting, independent evaluation, liability, licensing, compute thresholds, open weights, critical infrastructure, employment and consumer protection.

For the public, this is a reminder that concern can be translated into inspectable demands. Safety is not a mood. It can mean a named test, a public report, a protected whistleblower, an appeal route, a preserved log and a regulator capable of stopping a release.

03

WHERE IT COULD HELP

• Publish the complete questionnaire, response options, question order, field dates, sample design, weighting variables, design effect, subgroup bases and full topline before turning one percentage into a policy claim
• Repeat the same core questions on a stable schedule so readers can distinguish a durable trend from movement inside normal survey uncertainty
• Split serious harm into concrete categories such as fraud, cyberattack, discrimination, job displacement, unsafe advice, military misuse and loss of human control
• Ask respondents to evaluate specific controls including incident reporting, independent testing, licensing, liability, model access for regulators and temporary release restrictions
• Show the unsure and no-answer share beside every yes and no result instead of treating uncertainty as empty space
• Report subgroup results only with the subgroup sample size, larger margin of sampling error and enough repeated evidence to avoid building a story on noise
• Test wording effects by rotating race, competition, innovation, safety and responsibility frames across comparable groups
• Give respondents a short neutral description of a proposed rule before asking whether they support it, then publish the description exactly
• Ask what evidence would increase trust, including public evaluations, named regulators, independent audits, incident records, compensation and a right to appeal
• Require AI developers to publish release receipts that connect each major risk to a metric, threshold, result, limitation, decision owner and corrective action
• Create a public incident system that separates reports, confirmed events, causes, severity, remedies and unresolved disputes
• Pair opinion polling with observed behavior, complaint data, adoption, incidents, employment changes and independent technical evaluations so politics is not asked to carry the whole measurement burden