AI decision making appears to have hit a boundary that everyday AI use has not. A Rutgers survey of 5,154 US adults found that use of major AI tools has risen sharply, while fewer than one in ten respondents were comfortable letting AI make final, unsupervised decisions in areas such as hiring, loans, college admission, parole or who receives a medical procedure first.
The result captures an important split in public attitudes. People can use AI to write, search, plan or summarise and still want a human to remain accountable when the outcome changes somebody’s job, money, education or liberty.
That is a more useful picture than asking whether the public is simply pro-AI or anti-AI. Adoption and caution are increasing together.
AI decision making feels different from AI assistance
The Rutgers researchers separated assistance from final authority. Respondents were more willing to let AI help a person make a consequential decision than to let the system decide on its own. For hiring, for example, about a third accepted an assisting role, while acceptance of fully autonomous final decisions remained much lower.
That distinction mirrors how many organisations currently deploy AI. A model might rank applications, summarise evidence or flag unusual cases while a named employee formally makes the decision. The practical question is whether that human review is meaningful or merely a final click on a machine-generated recommendation.
LiveAIWire has previously covered evidence that people can struggle to remain independent once an AI recommendation is already on the screen. Human oversight only works if the human has enough information, authority and time to disagree.
Use has climbed quickly
The survey found that 72 per cent of respondents had used at least one major AI tool, up from 50 per cent in a comparable Rutgers survey in November 2024. More than a quarter said they used AI every day.
That is a rapid shift from novelty to routine behaviour. It also explains why arguments about AI governance can no longer assume that sceptical people are simply non-users. The survey found substantial concern even among frequent users.
The research was carried out online between 5 July and 1 August 2026. It used a non-probability sample with quotas and statistical weighting designed to better reflect the US adult population. As with any online survey, the figures describe the sampled and weighted responses rather than a perfect census of national opinion.
People want to know when AI is judging them
Disclosure produced some of the clearest agreement in the survey. Roughly three quarters of respondents said employers should disclose when AI is used to evaluate workers or job applicants. Similar majorities wanted companies to tell customers and schools to tell students when AI is being used in consequential ways.
That demand is understandable because automated systems can be invisible. A person may never know that an application was screened, a conversation analysed or a score generated before a human saw the case. Disclosure does not by itself make the system fair, but it gives the affected person a chance to ask what happened.
The issue is particularly relevant in recruitment, where LiveAIWire has examined how small changes in a CV can alter AI-assisted hiring outcomes. A process can feel administrative to an employer while being life-changing to the person being scored.
Frequent users are not automatically relaxed about the risks
One of the more striking findings is that familiarity did not erase concern. The survey reports that about half of respondents were worried about AI, and worry remained common among daily users. That suggests experience can produce both appreciation and caution rather than a simple march towards acceptance.
It also undercuts the assumption that resistance is mainly caused by ignorance. People can find AI genuinely useful while also believing there are decisions it should not make without human responsibility.
That tension is likely to intensify as systems move from answering questions towards acting on behalf of users and organisations. The more capable the assistant becomes, the more important it is to define where assistance ends and authority begins.
Confidence in understanding AI exceeds tested knowledge
The Rutgers survey also asked people about their understanding of AI. Many daily users said they understood the basics, yet their average performance on factual knowledge questions was much lower than that confidence might imply.
That is not unusual for a fast-moving technology. People learn interfaces before they learn architecture. Someone can be highly competent at using a chatbot without knowing how training data, probability, retrieval or model errors work.
The gap matters when users are asked to judge whether an AI-supported decision is trustworthy. Effective oversight requires more than knowing which button to press. It requires understanding that the output can be persuasive, consistent and still wrong.
The public seems to be asking for accountability, not a ban
The survey should not be read as a rejection of AI. Tool use is rising. People are accepting assistance in many settings. The stronger resistance appears when the system becomes the final authority over another person.
That is a design signal for companies as much as a policy signal. A bank, employer, school or public body that introduces AI can preserve more trust by making the human role visible, explaining when automation is used and creating a route for challenge when the system produces an important outcome.
LiveAIWire previously reported that disclosing that a persuasive message came from AI can change how people respond to it. Transparency does not solve every problem, but hidden automation creates an additional trust problem before the underlying decision is even examined.
The useful dividing line is consequence
AI is already ordinary enough that asking whether people trust it in general is becoming too crude. Trust depends on the task. Letting a system rewrite an email is not equivalent to letting it reject a loan. Asking for restaurant suggestions is not equivalent to deciding who receives parole.
The Rutgers findings suggest that the public intuitively recognises this hierarchy. The more consequential and less reversible the outcome, the stronger the demand for a responsible person to remain involved.
That may turn out to be one of the more durable rules for mainstream AI adoption: people are often happy to let machines help, but much less willing to let them become the person in charge.
The survey also exposes a knowledge problem
The researchers did not only ask whether people liked AI. They tested basic knowledge as well. Daily users were especially likely to say they understood how AI works, but their average score on factual questions was much lower than that confidence suggested. Familiarity with an interface is not the same as understanding the system behind it.
That matters because consequential AI is often introduced with language such as score, recommendation or risk flag that sounds objective. A user who overestimates their understanding may give an output more authority than it deserves, while a sceptical user may reject a useful tool for the wrong reason. Better public literacy should therefore include both capabilities and limitations.
A separate report on the Rutgers findings emphasised the same contrast between rising use and resistance to fully automated decisions. The survey is not a referendum on one product or company. It is evidence that people distinguish between convenience and control.
That distinction could become more important as AI agents move from suggesting actions to carrying them out. The public may tolerate a system recommending a flight, for example, while expecting a much higher standard before it can reject a job candidate or alter access to credit.
Why the human role needs to be more than ceremonial
A company can say that a person remains “in the loop” while giving that person almost no realistic opportunity to challenge the machine. If an employee sees only a final score, is expected to process hundreds of cases quickly or is penalised for disagreeing with the recommendation, human review exists mainly on paper.
The survey’s preference for human involvement is therefore only the starting point. Meaningful oversight requires access to the relevant evidence, a route to question the model and responsibility for the final outcome. Otherwise the organisation can preserve the appearance of human judgement while the automated system still determines what happens.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.