AGI Capability and Safety Should Matter Equally, Say 45% in Consumer Research

A small US survey points to a clear public preference: advanced AI should be built with safety and control treated as core requirements. Only a small minority wants capability to take priority on its own.

A survey of 100 Americans found that 44.6% consider AI capability and safety equally important when companies develop advanced systems. Another 42% place safety above capability, while only 7.9% choose capability alone. Rounded to the nearest whole number, about 45% want both goals to carry equal weight.

That result matters because the artificial intelligence industry is increasingly defined by competition to build systems with broader reasoning, greater autonomy, and potentially human-level performance across many tasks. Companies may view capability gains as evidence of technical progress. The public appears to be judging progress by a stricter standard: whether those systems can be controlled, evaluated, and governed before they are widely deployed.

The survey also found that 87.6% of respondents reject an unrestricted race toward artificial general intelligence, or AGI. They favor slower development, independent safety oversight, or a complete pause. Only 12.4% support moving as quickly as possible.

Together, the findings show why a company reaching AGI first would gain limited public legitimacy unless it can demonstrate that safety decisions are independent, transparent, and enforceable.

Nolla Health began a one-year Utah pilot focused on adults with mild to moderate acne. Patients pay $4.99 for the subscription, submit health information and facial photos, and receive an AI-generated treatment recommendation that can be sent to a pharmacy. Licensed clinicians remain available through telehealth, and cases involving drugs outside the program’s topical treatment list are routed to them.

The arrangement gives Utah a useful testing ground for a question that reaches well beyond dermatology: how much medical decision-making can be automated safely when the system has a limited scope, a low price, regular follow-up, and a clear path to human care?

The survey question forced respondents to choose among three priorities: building the most capable AI system, building the safest and most controllable system, or treating both goals as equally important.

The answers were heavily weighted toward safety:

The combined figure is the most useful way to read the result. A total of 86.6% placed safety at least on par with capability. That leaves a narrow constituency for a development strategy centered mainly on performance, speed, or technical lead.

This preference reflects a practical concern. A more capable model can create more value, but it can also affect more areas of life, make more consequential decisions, and produce more serious failures. If an advanced system is used in health care, finance, education, public services, software development, or critical infrastructure, a failure can spread beyond a single user or product session.

Safety and controllability therefore influence the value of the system itself. A model that performs impressive tasks but cannot be reliably monitored, constrained, or shut down may be difficult to use in high-stakes settings. Capability without dependable control can limit adoption, invite regulation, and damage the reputation of the company behind it.

For the public, the issue is easy to understand. A system that can do more should also come with stronger evidence that it can be managed when something goes wrong.

A survey of 100 US adults highlights a widening gap between perceived AI capability and public confidence in responsible development. OpenAI was viewed as the most likely company to develop AGI first, yet only 11.4% trusted it most to do so responsibly. Respondents generally supported continued research, but preferred slower progress, independent safety oversight, incident disclosure, shared standards, and the authority to pause risky releases. The findings are directional rather than nationally representative, but they point to a clear challenge: technical leadership will not produce durable public support without visible and enforceable accountability.

The survey presented respondents with a scenario in which one company appeared close to developing AGI. The question asked what should happen next.

The responses favored conditions on development:

These answers show strong opposition to an unrestricted race. They also show that the public does not hold a single view about the proper speed of AI development. Some respondents favor continued research under external oversight. Others want a delay while safeguards are built. A smaller group supports stopping development altogether.

The common thread is the demand for conditions that companies cannot define for themselves. When the responses are grouped by whether development should continue, 48.6% favor continued development either quickly or with oversight, while 51.4% prefer slowing or stopping it.

That near-even split on continuation makes the 87.6% figure especially important. The public may be divided about whether AGI development should continue, but there is much wider agreement that a company should not be allowed to set the pace without meaningful checks.

A winner-take-all structure creates a direct incentive to move before competitors do. Each company may fear that caution will cost it technical leadership, investment, talent, or market share. From the public’s perspective, that incentive is precisely why safety cannot remain an internal promise.

The survey found a sharp gap between perceived ability to develop AGI and confidence in responsible development.

OpenAI was viewed by 31.7% of respondents as the company most likely to develop AGI first. Only 11.4% trusted OpenAI most to develop it responsibly. Among respondents who expected OpenAI to reach AGI first, only 36% also trusted the company most with responsible development.

Google produced a closer match between capability and trust. Half of respondents viewed Google as the most likely company to develop AGI first, and 54.6% trusted Google most to develop it responsibly. Google was the only company to receive majority support on responsible development in the survey.

Those figures do not prove that one company is safer than another. They show how the public separates technical reputation from institutional credibility. A company may be seen as highly capable while still facing doubts about its judgment, incentives, communication, or willingness to accept outside review.

That distinction has direct consequences for AGI deployment. A technical lead may attract attention, talent, and capital. Public trust determines whether institutions, regulators, customers, and employees are willing to accept the company’s claims about safety.

Technical achievement can open the door to deployment. Credible oversight helps determine whether the door stays open.

The survey’s most important finding is not that the public rejects AGI; it is that people reject the idea that competitive speed should determine when powerful systems reach the world. The sample is too small to serve as a precise measure of national opinion, but its pattern is consistent with broader skepticism toward companies’ ability to govern AI responsibly. Developers should treat trust as an operational requirement, not a communications problem. Independent evaluation, transparent incident reporting, staged deployment, and real authority to delay a launch would provide stronger evidence than assurances issued by the companies themselves.

Respondents were divided about who should have the final say on whether an advanced AI system is safe enough to release. Government regulators, the developer itself, and independent scientific experts each received about 29% support. An international regulatory organization received 12.7%.

The most important result appears in the combined numbers. About 70.5% assigned primary responsibility to an entity outside the company that built the system. Only 29.5% believed the developer should have final authority.

That finding points to a basic governance principle: the party with the strongest commercial incentive to release a system should not be the sole judge of whether the system is ready.

External oversight could take several forms. The survey does not identify one preferred structure, and the lack of consensus should be taken seriously. Possible models include:

Each model has tradeoffs. Government review can become slow or politically contested. Independent experts need access to enough information to conduct meaningful evaluations. Company safety teams need authority and resources rather than a symbolic role. International coordination can improve consistency but may be difficult to implement.

A workable system may use several layers instead of relying on one gatekeeper. Internal testing can catch problems early. Independent evaluators can challenge company assumptions. Regulators can establish minimum requirements and intervene when risks exceed accepted limits.

The 100-person survey should be read alongside larger public opinion research. Pew Research Center reported in 2025 that 59% of the US public had not too much or no confidence in US companies to develop and use AI responsibly. Pew also found that Americans hold mixed views about AI’s effects on people and society, including concerns about whether people can identify AI-generated content.

The Stanford AI Index reported a similar tension at the global level. The share of respondents saying AI products and services offer more benefits than drawbacks increased from 55% in 2024 to 59% in 2025. At the same time, 52% said AI made them nervous.

Those findings help place the AGI survey in context. Public opinion is capable of recognizing benefits while still demanding safeguards. Support for useful AI does not automatically translate into support for the companies developing the most powerful systems.

The National Institute of Standards and Technology’s AI Risk Management Framework offers one established way to organize responsible development around trustworthiness, evaluation, and risk management. The framework is voluntary, so its value depends on whether organizations apply it consistently and whether stronger requirements emerge for higher-risk systems.

The public preference for independent oversight suggests that voluntary frameworks may need additional accountability when AI systems become more capable or more deeply connected to essential services.

The survey provides a useful snapshot of attitudes, but its sample size is small. It included 100 Americans between ages 18 and 63, and the available description does not establish that the sample was a probability-based sample representative of the entire US population.

That means the findings should be treated as directional. They can identify concerns worth investigating and show how respondents reacted to specific choices. They cannot establish a precise national opinion or support a conventional margin of error without a suitable sampling design.

The wording of survey questions also affects results. For example, respondents may interpret terms such as AGI, safety, controllability, and independent oversight in different ways. A larger study could test whether views change after participants receive a common definition or concrete examples of possible AGI risks.

The sample limitation does not erase the central signal. Even a small survey can show that safety and oversight are central to how many people evaluate advanced AI. Larger, representative studies can determine how broadly that view applies and how stable it remains.

The findings support a concrete set of expectations for companies and policymakers.

For AI companies, responsible development would involve more than publishing principles. It would require safety teams with the authority to delay releases, evaluations conducted by people who are not rewarded solely for launch speed, and clear disclosure when serious failures occur.

For policymakers, the public preference for outside accountability supports rules that focus on high-risk systems rather than every ordinary AI application. Requirements could scale with the system’s capabilities, access to tools, reach, and potential impact.

For independent researchers, meaningful oversight requires access. Evaluators need enough information, testing time, and technical support to find problems before a system reaches millions of users. A review process that receives only selected demonstrations cannot provide strong assurance.

A practical governance package could include:

These measures would add time and cost to development. That is a real tradeoff. The alternative transfers those costs to users, workers, institutions, and the public after a failure occurs. A capable system that lacks credible controls may also face greater delays later through lawsuits, regulation, or loss of market access.

The survey’s headline finding is straightforward: about 45% of respondents want capability and safety treated equally, and another 42% put safety first. That is a strong public standard for companies pursuing AGI.

The result does not settle the debate over whether AGI development should continue. It does establish what many people expect before that development earns wider acceptance. Companies need to show that safety work receives comparable attention, resources, and authority. Regulators and independent experts need a meaningful role in release decisions. Serious failures need to be disclosed with enough detail for outside observers to judge the response.

A company that reaches AGI first may gain a technical milestone. Public trust will depend on what follows: who tested the system, who could stop its release, what evidence supports the safety claims, and what happens when the system behaves unexpectedly.

The public appears willing to accept continued AI progress under those conditions. Shared standards, independent oversight, and transparent reporting can preserve competition while reducing the pressure to cut corners. Without enforceable safeguards, a technical lead will carry less weight than the industry expects.