The Numbers Are Not Subtle

In 2023, roughly 65% of US adults held a positive view of artificial intelligence. By early 2026, that figure had fallen to around 42%. That is not a polling blip. That is a sustained, multi-survey shift in how ordinary people feel about a technology that is being woven into their daily lives without much of their input.

The drop is sharpest among adults aged 35 to 55. These are working people with mortgages and career expectations , precisely the group with the most at stake when automation enters the office.

When you ask people what they are unhappy about, the answers are consistent across surveys. Sixty-one percent cite job displacement as their primary concern. Fifty-eight percent say they distrust AI-generated content , a finding that reflects both the volume of synthetic content now online and a general sense that the information environment has become harder to work through. Fifty-two percent flag privacy. And underneath all of it is something harder to quantify: a sense that this technology is being pushed on them without their consent.

That last concern does not show up cleanly in any single survey question. It surfaces in the language respondents use in open-ended answers. Words like "imposed," "forced," and "no choice" appear with notable frequency when people describe their relationship with AI tools they now encounter at work, in customer service, and in content feeds. The sentiment is not just about specific harms. It is about agency , or the absence of it.

Why the Industry Can Afford to Shrug

Here is the structural reality that makes consumer sentiment largely irrelevant to AI's commercial trajectory. The money is in enterprise. Microsoft does not sell Copilot licences one anxious consumer at a time. It sells them in bulk to procurement officers at Fortune 500 companies, who are weighing cost reduction and productivity metrics, not public approval ratings.

The people who are unhappy are not, in most cases, the people signing the contracts. This is a familiar dynamic in the history of enterprise software.

Cloud computing is the closest parallel. For years, IT departments resisted it , security concerns, loss of control, vendor dependency. The surveys of IT professionals in 2010 and 2011 were scathing. But CFOs were looking at capital expenditure versus operating expenditure, and the CFOs won. By the time IT sentiment shifted, cloud was already the default. The people making the purchasing decisions were simply different people from the ones expressing the skepticism.

The same structure appears to be playing out with AI. CIOs and operations executives are buying. Workers and consumers are worried. The two groups are having different conversations, and only one of those conversations closes deals. This is not cynicism , it is an accurate description of how large organisations adopt technology. The approval of the people affected by a purchasing decision has rarely been a requirement for that decision to proceed.

Where Public Opinion Actually Has Teeth

There is one channel through which consumer unhappiness could bite the industry, and it runs through politics. Sustained public dissatisfaction, if it becomes politically salient, gives legislators a mandate to act. That is the lever that converts low approval ratings into something with actual consequences for corporate strategy.

This is not hypothetical. The EU AI Act was partly a product of public pressure. Several US states have introduced AI-specific labour protection bills in the past two years. None of them have passed in their original form, but the legislative pipeline is real. If approval ratings keep falling and the issue starts showing up in polling as a voter priority, the regulatory risk calculus for AI companies changes considerably.

The question is whether the political cycle moves faster than the embedding cycle. AI is already inside enough enterprise workflows that pulling it out would be operationally disruptive. If regulation arrives after the technology is structurally embedded, its practical scope narrows. Companies can argue that removing or restricting AI would harm their competitive position, their customers, and their employees. That argument gets stronger the deeper the technology is embedded.

History suggests that regulations imposed on deeply embedded technologies tend to address practices at the margins rather than the technology itself. The internet was regulated around specific harms , child safety, fraud, privacy , not around its existence. AI regulation will likely follow a similar pattern. The window for systemic constraint may be shorter than critics assume.

The Labour Problem Is More Immediate

There is a more near-term problem that gets less attention than the regulatory risk. AI unpopularity is strongest among workers who feel surveilled or replaced. And those workers are still showing up to work every day, often alongside the AI systems they distrust.

Companies rolling out AI tools over employee objection face a specific set of problems: resistance to adoption, errors that get attributed to the tool rather than corrected, and quiet disengagement from the workflows the AI is supposed to improve. This is not a PR problem. It is an operational one.

Industrial relations researchers have a term for the low-grade friction that develops when workers feel that new systems are being imposed on them. It does not look dramatic from the outside. But it consistently shows up as slower-than-expected productivity gains, higher error rates, and elevated turnover in roles adjacent to the AI deployment. The CFO who approved the licence does not always see that cost clearly in the spreadsheet.

Some of the most publicised AI rollout failures in the past two years have had this quality. The technology worked as intended. The workforce did not adopt it as intended. When investigators looked at why, the consistent finding was that workers had not been involved in the decision, did not trust the tool, and had found ways to route around it that preserved their sense of control. The AI sat in the workflow. The humans worked beside it rather than with it. The productivity gains were minimal.

The Embedding Question

History does suggest that resistance to convenient technologies tends to soften once those technologies become part of daily life. People who said they would never use GPS now cannot work through without it. People who predicted that social media would never replace face-to-face socialising now manage their relationships partly through those platforms. The general pattern of initial resistance followed by normalisation is well-documented.

But that pattern has preconditions. The technology has to deliver enough visible value to enough people that the cost of resistance starts to feel higher than the cost of adoption. And the adoption has to happen in a way that feels at least partly chosen rather than purely imposed. GPS felt like a tool you picked up because it was useful. Social media spread through networks of individual choices. The experience of adoption shaped the attitude toward the technology.

AI's current trajectory is mixed on both counts. The productivity gains are real in some contexts and overstated in others. Developers who use AI coding tools report genuine time savings. Knowledge workers who have tried to use AI for complex analysis often report frustration with outputs that look right but require significant verification. The experience is uneven in ways that the aggregate approval numbers partly obscure.

The deployment pattern , top-down, driven by enterprise purchasing decisions rather than individual choice , is exactly the kind that tends to generate sustained resistance. When people feel a technology was chosen for them by someone with different interests, their relationship with it starts adversarial. That starting position is harder to overcome than the starting position of voluntary adoption, even when the tool turns out to be useful.

What to Watch

The metric that matters most is not the approval rating itself. It is whether the disapproval translates into political priority. AI showing up as a top-five economic concern in voter surveys heading into any major election cycle would change the industry's ability to ignore public sentiment. As long as it remains a background concern that people feel but do not organise around, the enterprise adoption engine keeps running regardless of how the polls look.

The second metric worth watching is productivity realisation. If enterprise AI deployments start producing documented, measurable productivity gains that workers can see and feel, the emotional temperature changes. People are willing to accept tools they initially distrusted if those tools make their working day better. The current problem is that too many AI deployments are not doing that , they are adding workflow steps and monitoring layers that workers experience as friction, not as help.

The industry is betting that embedding happens fast enough, and delivers enough value clearly enough, that public sentiment becomes irrelevant before it hardens into regulatory constraint. That bet is probably right. Probably.

But the workers who are being quietly surveilled and slowly replaced will still be there, whatever the surveys say.

Their opinions did not close the deal. But they will decide how well it works in practice.