Reuters/Ipsos Poll Finds Broad Concern About AI Safety and Company Oversight

Reuters/Ipsos Poll Finds Broad Concern About AI Safety and Company Oversight

Updated September 22, 2026

A new Reuters/Ipsos poll suggests that concern about artificial intelligence is no longer limited to technical researchers or policy specialists. Seventy-three percent of respondents said AI companies have not done enough to prevent serious harm to society, while 39% said AI is having a negative impact—a record high in Reuters/Ipsos polling on that question. Only 11% described AI’s impact as positive, according to Reuters’ September 22 report.

The four-day online poll concluded on Sunday and collected responses from 1,277 U.S. adults. Reuters reported a margin of error of three percentage points in either direction. The findings are a snapshot of public opinion, not a technical assessment of AI safety or a direct forecast of future regulation. Still, the results matter because they show the public debate moving beyond enthusiasm about convenience and productivity toward demands for accountability.

A majority, 55%, said slowing AI development would be a good thing, compared with 13% who said it would be bad. Seventy-three percent said safe and responsible development mattered more than keeping the United States ahead of other countries in AI, while 23% prioritized global leadership. Sixty-nine percent said they had followed news of major AI companies calling for slower development and stronger oversight.

What the poll measures—and what it does not

The poll records people’s views about AI, risk, and the responsibilities of companies. It does not measure the probability of an AI catastrophe, the safety of a specific model, or whether any particular regulation would work. The word “harm” can mean different things to different respondents, from job disruption and privacy loss to cybersecurity threats or extreme long-term risks. The survey should be read as a public-attitudes indicator rather than a technical verdict.

That distinction is important because headlines about AI often combine highly different concerns. A user’s worry about automated job decisions is not the same as a researcher’s concern about a powerful system pursuing an unintended goal. Both can be legitimate, but they call for different evidence and responses. Poll questions cannot substitute for careful definitions of the risks being discussed.

The sample size and margin of error also matter. A 1,277-person national poll can provide a useful estimate of broad opinion, but small percentage differences should not be overinterpreted. The poll is online, and any survey relies on its sampling and weighting methods. Its strongest signal is the large share of respondents expressing concern about corporate safeguards, not a fine distinction between neighboring percentages.

Reuters reported that 39% see AI’s impact on society as negative, up from 36% the previous month and the highest level since the question began in March. That month-to-month change is modest relative to the poll’s uncertainty; the longer pattern and the broader distribution of answers offer more context than a single movement. Only 11% said the impact was positive, while others were uncertain or did not answer.

The accountability question behind the numbers

The sharpest result may be the 73% who say companies are not doing enough to prevent serious harm. That is a judgment about trust and responsibility. People are being asked to use AI products while the systems change rapidly, and they may have little visibility into training data, testing, monitoring, or what happens when an automated decision goes wrong.

Companies often publish safety commitments, usage policies, and model evaluations. Those can be useful, but the public may struggle to compare them. A benchmark score does not tell a user how a system will behave in a workplace, school, medical setting, or customer-service dispute. Nor does a company’s own evaluation independently establish that its safeguards are effective.

A practical response would make claims easier to verify. Developers can publish clear summaries of how systems were tested, what the evaluations do not cover, how incidents are handled, and when a system’s intended use changes. Independent reviewers can scrutinize high-impact claims, subject to appropriate protections for private data and security-sensitive details. Regulators can specify the information organizations must retain and provide when systems affect people’s rights or safety.

Accountability also needs a path to remedy. If an AI system produces a harmful decision, affected people need to know who is responsible, how to contest the outcome, and whether a human can review it. Safety language means little if users have no meaningful way to report failures or correct records.

The slowdown question is more complicated than a yes-or-no poll

The poll’s finding that 55% favor slowing AI development can be read as a call for restraint, but it does not define what should slow down. Respondents may have different ideas: pausing the largest models, delaying deployment in sensitive domains, limiting particular capabilities, increasing testing time, or changing how quickly firms release products. Those approaches have different costs and benefits.

A broad pause could delay beneficial tools in research, accessibility, education, and productivity. At the same time, speed without adequate evaluation can expose users to failures that were foreseeable and preventable. A policy discussion should distinguish slowing the most consequential releases from stopping every AI project. Risk-based requirements can focus additional scrutiny where systems have greater capabilities, reach, autonomy, or potential impact.

Development and deployment are also different stages. A model can be researched without being released widely. A product can be tested in a limited setting before it is offered to the public. Organizations can require additional review when a system is used for high-stakes decisions. This makes it possible to manage exposure without treating all AI development as equivalent.

The figure indicating that 73% prioritize safe development over international competition is politically significant because it challenges a familiar argument: that the country must move as fast as possible to stay ahead. The result does not mean respondents reject investment or scientific progress. It suggests that many want leadership to be compatible with safeguards rather than defined only by speed.

Why public trust is an operational issue

Trust affects whether people adopt a tool, share information with it, or accept decisions it helps produce. For businesses, low trust can mean employees avoid approved systems, customers withhold data, or users abandon products after a failure. For public agencies, it can undermine confidence in services that rely on automated processing. Trust is not a substitute for performance, but it determines whether performance is accepted and scrutinized.

Organizations deploying AI should make the boundaries visible. Users need to know when a model is involved, what information it can access, whether a person reviews its output, and how to correct mistakes. Employees should receive practical guidance on when an AI output needs independent verification. These measures can address everyday concerns even while governments debate larger risks.

Companies should also be candid about uncertainty. Models can produce inaccurate information, reproduce biased patterns, or behave differently after updates. A responsible deployment explains the appropriate use and known limits rather than promising that errors have been eliminated. When a serious failure occurs, prompt disclosure and remediation can preserve more trust than a vague assurance that the system is being improved.

What regulators and policymakers can take from the survey

The survey does not itself tell lawmakers which statute to pass. It does establish that demands for oversight have a public constituency. Policymakers can respond by identifying clear obligations for systems used in consequential settings: risk assessment before deployment, human review where needed, reporting of serious incidents, privacy protections, and avenues to appeal decisions.

Those rules should be specific enough to enforce and flexible enough to account for changing technology. A requirement that merely says “use AI responsibly” gives little guidance to a company or regulator. A rule that specifies documentation, testing, user notice, and accountability can be evaluated and improved over time.

Policymakers should also avoid interpreting public concern as a mandate for any single regulatory approach. People may support oversight while disagreeing about whether it should be federal, state, industry-led, or international. They may also weigh safety against affordability, access, innovation, and national security differently. Consultation and transparent evidence help turn broad sentiment into workable policy.

The survey was conducted in the United States. Its results should not be generalized automatically to other countries, where public experiences, regulation, and economic expectations may differ. International comparisons require comparable questions, samples, and methods.

Questions to watch next

Future polling can show whether the concern reflects a stable shift or a response to a high-attention news cycle. Researchers should compare attitudes across specific use cases—health, employment, education, creative work, and public services—rather than asking only whether AI is good or bad. More detailed questions can identify which safeguards people actually expect and which kinds of harm they consider most urgent.

For companies, the signal is clear enough: claims of self-regulation need evidence people can understand. For government, public concern is a reason to hold hearings and evaluate policy options, not a shortcut around careful legislative work. For users, the figures describe an ongoing debate, not a conclusion that every AI system is unsafe.

The strongest interpretation is also the narrowest: in this Reuters/Ipsos survey, a large majority of U.S. adults expressed concern that AI companies are not doing enough to prevent serious harm, and more respondents favored safety over winning the global race. Whether that sentiment leads to durable policy will depend on the quality of proposed safeguards, the evidence behind them, and whether institutions can make responsibility visible when systems fail.

Sources and further reading

• Reuters: Three out of four Americans say AI firms not doing enough to prevent disaster
• Reuters/Ipsos Polls
• Ipsos U.S. public opinion research
• NIST AI Risk Management Framework
• OECD AI Principles
• U.S. Government Accountability Office: Artificial intelligence
• Federal Trade Commission: AI and consumer protection

How leaders should interpret the trend

One poll cannot establish why sentiment changed. The movement may reflect recent coverage of AI safety, personal experiences with inaccurate systems, concern about jobs, or unease about the scale of investment. The poll asks respondents how they view AI’s impact and corporate precautions; it does not isolate the cause of those views. Decision-makers should avoid assigning a single explanation without additional research.

Repeated polling can help track whether attitudes move after specific events or product releases. Good comparisons require consistent wording and methods. Policymakers should also pair public surveys with evidence about actual system failures, adoption, economic effects, and the effectiveness of safeguards. Public opinion identifies priorities, while technical and social research helps determine which interventions are likely to work.

Companies can make trust measurable inside their own operations. They can track complaints, reversals, appeals, and the time required to correct harmful outputs. They can publish how often human review changes a consequential recommendation and how they handle repeat failure patterns. These indicators are more informative than a general pledge to be responsible because they show how governance operates in practice.

The risk of treating trust as a communications problem

A public concern cannot be answered only with better messaging. If an organization says a product is safe while users cannot see its limits or challenge a decision, the gap is substantive. Clear communication is important, but it has to match the system’s actual behavior and the remedies available when it causes harm.

The poll also does not imply that the public opposes AI altogether. People may value assistance in one setting and reject automation in another. They may want faster scientific research while demanding human review for decisions about employment, benefits, or healthcare. Product and policy choices should reflect those differences rather than assuming one universal level of acceptance.

A responsible response therefore combines transparency with control: understandable notices, careful deployment boundaries, meaningful human recourse, and independent assessment where consequences are substantial. Those measures make it possible to earn trust through observable practice instead of asking the public to take corporate assurances on faith.