Americans Demand Stronger AI Safety as New Poll Shows Deep Concern Over Catastrophic Risks.

A national survey published on September 22 points to a substantial confidence problem for artificial intelligence companies: 73 percent of Americans expressed concern that the industry is not doing enough to prevent serious harm to society. The online poll included 1,277 US adults and had a reported margin of error of about three percentage points. It also found 55 percent support for slowing AI development.

Those findings measure public attitudes, not the statistical likelihood of a catastrophic event. That distinction is essential. People can favour stronger safeguards without agreeing on precisely which future risks are most probable, how quickly they might emerge or which restrictions would be justified.

Public concern is not a technical forecast

An opinion poll answers questions about what respondents think when they are surveyed. It cannot establish whether a particular model will escape its controls, whether a proposed safety measure will work or when a hypothetical disaster might occur. Treating a concern figure as a prediction of catastrophe would give the results a meaning they do not have.

At the same time, attitudes matter in their own right. AI services are being offered to people who may have limited visibility into their design or evaluation. Trust becomes difficult when a company asks users to accept broad assurances while the consequences of an error are borne by an employee, customer or member of the public.

The poll is therefore most useful as evidence of an accountability gap. A large share of respondents want more confidence that serious risks are being addressed. The industry's challenge is to explain what protections exist, how they are tested and what happens when they fail.

Different harms require different safeguards

The phrase AI safety covers problems that should not be treated as interchangeable. A chatbot inventing information, an automated service exposing private data and an autonomous agent exceeding its permissions raise different technical and organisational questions. A measure designed for one problem may do little to solve another.

For example, checking the factual accuracy of an answer does not establish that an agent should be allowed to alter a database. Restricting tool permissions may contain an unauthorised action but does not ensure that advice is correct. Effective oversight requires specific descriptions of the system, its access and the consequences of a mistake.

This is also why dramatic language can obscure practical work. A discussion focused entirely on distant catastrophic scenarios can overlook immediate failures that people already understand. Conversely, attention only to everyday errors can leave more consequential capabilities insufficiently examined. Both need evidence rather than competing slogans.

What a meaningful safety claim should contain

A useful safety statement should say what was evaluated, under which conditions and with what limitations. It should distinguish internal testing from independent assessment and explain whether protections were active during the tests. A summary that reports only a favourable score leaves important questions unanswered.

The same principle applies to incidents. Organisations need a process for recording failures, determining their causes and assessing whether affected users require notification or remediation. Publishing an improved model does not automatically resolve the weaknesses of the application in which an earlier version was deployed.

Oversight must have a route to action

Independent review is valuable only if reviewers can examine relevant evidence and their findings can influence decisions. A review that arrives after an irreversible deployment, or cannot reach the details needed to reproduce a problem, may provide less assurance than its label suggests.

For an organisation deploying AI, internal responsibility must also be clear. Someone needs authority to restrict access, stop an unsafe workflow and decide when testing is sufficient to resume it. Assigning those duties in advance is more useful than assuming a technical team and a management team will resolve responsibility during a crisis.

Slowing development is not one single policy

Support for a slower pace does not by itself identify what respondents want slowed. Research, release of a powerful model and permission for a system to act in a sensitive environment are different decisions. A proposal can be strict about deployment while allowing research into reliability and beneficial uses to continue.

Policy discussions should specify the capability or application being governed and the evidence that would trigger a restriction. Otherwise, a broad call for caution can become either an impractical prohibition or a vague promise that changes little. Clarity is also important for smaller developers who need to understand their obligations without navigating rules designed only around the largest companies.

Rebuilding confidence through observable behaviour

AI companies cannot settle a public-confidence problem solely by emphasising potential benefits. Users need a way to distinguish a service that has earned trust from one that merely communicates confidently. Useful signals include documented testing, honest limitations, accessible complaint routes and evidence that known failures lead to changes.

The survey does not prescribe a complete regulatory framework. It does show why assurances of responsible development need to be backed by practices that people outside a company can examine. Progress and caution are not necessarily opposites; reliable systems can make beneficial adoption easier.

The constructive response to these findings is neither to dismiss public concern nor to present fear as scientific proof. It is to make responsibility concrete: identify the risks, test the controls, explain the remaining uncertainty and provide a way to intervene. Confidence in AI will depend on that record of behaviour more than on another promise that safety is a priority.

PUBLISHED

BY

SUYASH PACHAURI,

FOUNDER & OWNER,

GLOBAL BOLLYWOOD | THE HOLLYWOOD SCOPE