Survey Reveals Who Respondents Want Judging Whether AGI Is Safe to Release

A survey of 100 Americans found that most respondents want someone outside the company building an advanced AI system to help decide whether it is safe to release. Their preferred judges were government regulators and independent scientific experts, with the company itself receiving much less support.

A decision about releasing artificial general intelligence would affect workers, institutions, national security, and everyday users. The survey results suggest that Americans are uncomfortable leaving that decision solely to the companies developing the systems. About 70.5% of respondents assigned primary responsibility to someone outside the developer, while 29.5% said the company building the system should have the final say.

The result fits a broader pattern in the survey. Respondents remain open to continued AI development, but they want release decisions tied to independent review, shared safety requirements, and public accountability. The message is straightforward: reaching AGI first will carry little public legitimacy if a company cannot show that its safety decisions are transparent and enforceable.

The survey gave respondents several choices about who should decide when an AGI system is safe to release. Three options landed close together:

The narrow gap among the top three choices matters. Respondents did not settle on one institution as the obvious authority. Instead, they divided their confidence among public regulators, outside experts, and the developers themselves.

That split points toward a layered review system rather than a single gatekeeper. Government agencies could set legal requirements and enforce them. Independent experts could test systems, examine evidence, and challenge company claims. Developers would still contribute technical knowledge and remain responsible for their systems, but they would have less control over the final release decision.

The least popular option was an international regulatory organization. Only 12.7% selected it. That result could reflect discomfort with transferring authority to a global body, uncertainty about how such an organization would operate, or a preference for institutions that are easier for Americans to identify and hold accountable.

The survey does not establish why respondents chose each option. It does show that self-certification has limited appeal when the technology is described as capable of matching or exceeding human performance across many fields.

A survey of 100 Americans suggests that the public is open to continued AGI development but does not want companies to decide alone when powerful systems are safe to release. Government regulators and independent scientific experts received support nearly equal to that of developers, while most respondents preferred slower progress or stronger external oversight over an unrestricted race. The findings point toward a layered model in which companies provide evidence, independent experts test it, and regulators hold authority to delay or approve deployment. Because the sample may not be representative, the results should be treated as an indication of public concerns rather than a precise national measurement.

A company developing AGI would have strong incentives to declare its system ready. It may face pressure from investors, competitors, customers, employees, and national policymakers. Those pressures can influence how safety evidence is interpreted and how much uncertainty the company is willing to accept.

That creates a basic governance problem. The organization with the most detailed knowledge of a system may also be the organization with the strongest reason to release it. Technical expertise and decision-making independence are related, but they are not the same thing.

The survey captures public concern about that overlap. Only 29.5% of respondents believed the developer should have the greatest responsibility for deciding whether an AGI system is safe to release. A much larger share favored an outside party, even though respondents disagreed about which outside party should take the lead.

A credible process would need to answer several practical questions:

Without clear answers, the phrase “independent oversight” can become a label rather than a working safeguard. Oversight has to include access, authority, technical capacity, and a process for resolving disputes.

Government regulators and independent scientific experts received similar support, but they would serve different functions.

Regulators can create binding rules. They can require testing before deployment, set reporting obligations, impose penalties, and investigate failures. They also carry democratic authority that private companies and informal expert groups do not have. Their weakness is technical capacity. Agencies may struggle to keep pace with rapidly changing systems or to recruit enough specialists to evaluate them.

Independent scientists can examine technical claims with more flexibility. They may be able to reproduce tests, challenge assumptions, and identify weaknesses that internal teams missed. Their independence can also make their conclusions more credible to the public. Their weakness is enforcement. An expert panel can issue a serious warning, but another institution would need the power to act on it.

A workable release process could combine both roles. Scientific reviewers could evaluate capability, reliability, misuse risks, and control measures. Regulators could turn those findings into a decision with legal force. A company could submit evidence and participate in the process without having the final authority to certify its own system.

That arrangement would also reduce the pressure placed on individual experts. If a safety review has legal standards, published procedures, and documented evidence, no single scientist has to carry the full burden of saying yes or no to a system with wide social consequences.

The survey’s most important finding is not which institution ranked first, but that self-certification lacks legitimacy when the stakes are high. Developers possess essential technical knowledge, yet their commercial and competitive incentives make them a poor sole judge of readiness. A credible system should combine company responsibility with independent technical review, binding regulatory authority, transparent reporting, and ongoing monitoring. The sample is too small to settle public opinion, but its consistent preference for safety, cooperation, and outside accountability reflects a governance problem that companies and policymakers cannot solve with public-relations assurances alone.

The question about release authority sits within a larger rejection of an unchecked AGI competition. Only 12.4% of respondents wanted companies to continue developing AGI as quickly as possible in the scenario presented.

The remaining respondents selected some form of constraint:

Those numbers show that the public preference is conditional development. Many respondents are willing to see research continue, but the conditions matter. They want safety work to keep pace with capability gains, and they want outside institutions involved before the most powerful systems reach broad deployment.

The survey also shows that support for continuing development depends on the terms of competition. When respondents were asked which competitive structure would inspire the most confidence, 55.2% chose companies competing while cooperating on shared safety standards. Another 11.2% favored governments coordinating development, and 4.4% preferred an international organization managing AGI development.

Together, those choices gave coordinated or cooperative structures 70.8% support. Purely competitive arrangements received much less enthusiasm.

This is a practical distinction. Shared safety standards would allow companies to keep competing on products, research, and efficiency while establishing common minimum requirements for testing, reporting, model security, and controlled deployment. The model preserves competition while limiting the incentive to cut safety work in order to move faster than rivals.

The survey found a clear gap between perceived ability to develop AGI and perceived ability to develop it responsibly.

OpenAI was named 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 predicted that OpenAI would win the race, only 36% also trusted the company most on responsible development.

Google had a closer relationship between capability and trust. Half of respondents thought Google was most likely to develop AGI first, and 54.6% trusted Google most to develop it responsibly. Google was the only company to receive majority support for responsible development.

Those results do not prove that one company is safer than another. They measure public perceptions shaped by corporate history, public communication, product behavior, leadership decisions, and media coverage. Perception still matters because public acceptance affects regulation, adoption, and the political room companies have to operate.

A company that reaches a major technical milestone without public confidence could face serious resistance. Regulators may impose tighter requirements. Customers may hesitate to rely on its systems. Partners may demand additional testing. Employees and researchers may question whether internal safety concerns receive enough weight.

Technical achievement can open the door to deployment. Trust determines how widely that door stays open.

Respondents also made clear that capability alone is a poor measure of progress. Only 7.9% chose building the most capable AI system as the top priority when asked to weigh capability against safety and controllability.

Another 44.6% said both were equally important, while 42% prioritized safety alone. In total, 86.6% placed safety on par with or above capability.

That preference has direct implications for release decisions. A system that performs impressively but cannot be reliably monitored, constrained, or shut down will not satisfy most people. Capability increases the potential value of a system, but it can also increase the scale of mistakes, misuse, and failures.

For the public, the basic standard appears to be control before expansion. Companies would need to demonstrate that they understand the system’s limits, can detect harmful behavior, can respond to unexpected failures, and can prevent unauthorized access or use.

Those demonstrations should involve more than internal benchmarks. Useful evidence could include adversarial testing, independent replication, evaluations of deception and autonomy, cybersecurity assessments, misuse testing, and controlled pilots. The exact tests would vary by system, but the principle remains stable: release should depend on evidence about behavior and control, not a company’s confidence in its own engineering process.

The findings offer a useful snapshot of attitudes, but they need to be read with care. The survey included 100 Americans between the ages of 18 and 63. The available description does not establish that the sample was randomly selected or nationally representative.

That matters for how the percentages should be used. A small or non-probability sample can reveal patterns and generate questions, but it should not be treated as a precise estimate of the views of every American. A margin of error would be inappropriate without a probability-based sample and a known sampling design.

The survey also measures reactions to a specific set of questions. Wording, answer choices, the order of questions, and how AGI was explained can all influence responses. Some respondents may imagine an AGI system as a near-term commercial product, while others may picture a highly autonomous system with major effects on government, employment, or national security.

These limitations do not erase the central result. The response pattern is consistent across several questions: respondents favor safety, outside accountability, transparency, and cooperation on minimum standards. Larger, representative surveys would help measure how widespread those views are and how they change when people receive more information about specific oversight models.

Pew Research Center has reported similar skepticism about the ability of US companies to develop and use AI responsibly. Its 2025 research found that 59% of the public had not too much or no confidence in US companies on that question. The broader trend suggests that the survey’s concern about corporate credibility reflects a wider public debate rather than an isolated reaction.

External oversight does not remove responsibility from AI companies. Developers control much of the evidence, documentation, and early testing that an outside reviewer will need. They also decide whether internal safety concerns reach senior leadership and whether a known problem is disclosed promptly.

Companies can improve their credibility by publishing clearer safety cases, documenting limitations, reporting incidents, and allowing qualified independent researchers to test their systems. They can also separate safety approval from product and revenue teams so that a group responsible for finding problems has room to delay a launch.

Shared standards would help companies avoid a race in which every developer fears that stronger safeguards will put it at a commercial disadvantage. If major firms agree to common minimum practices, a company can slow a release for safety reasons without handing an automatic advantage to a less cautious rival.

The survey’s preferred structure captures that balance: companies continue to compete, but they cooperate on the rules that limit catastrophic mistakes. Competition can remain intense in areas such as performance, price, usability, and specialized applications. Safety requirements would set the floor beneath that competition.

Public confidence will depend less on the names of oversight institutions than on their ability to reject a release when the evidence falls short. An independent reviewer who can only recommend caution has limited power. A regulator that lacks technical staff may struggle to identify a problem. A company that publishes reports while ignoring outside findings will deepen suspicion.

The key test is operational:

Can an outside party inspect the evidence, identify a serious risk, require changes, and delay deployment when the company has strong reasons to move ahead?

If the answer is yes, public oversight has substance. If the answer is no, the process remains close to self-regulation regardless of how many panels, audits, or reports exist.

The survey shows that respondents are willing to consider continued AGI development under those kinds of conditions. They are also prepared to slow development when safeguards are missing. That gives companies and policymakers a clear choice about the direction of the race.

Technical leadership can produce headlines and investor confidence. Independent, transparent, and enforceable safety decisions will determine whether the public accepts what comes next.