AI Policy & Safety

The AI Trust Paradox: Usage Surged While Skepticism Grew

By The Agent5 founder·July 1, 2026·How we fact-check

More people are using AI than ever before, yet global surveys consistently show trust in the technology is declining. Understanding why these two trends move in opposite directions is one of the most important questions in AI policy today.

ShareX LinkedIn WhatsApp

Key takeaways

Global AI usage is at record highs, but multiple large surveys show trust in AI systems declining simultaneously, producing a persistent paradox that researchers call the AI trust gap.
Hallucinations and consequential errors, particularly in legal, medical, and financial contexts, are a primary driver of eroding trust, and current research indicates they cannot be fully eliminated from large language model architectures.
Shadow AI (ungoverned, unsanctioned use) is a direct consequence of the trust paradox and creates accountability gaps that are arguably riskier than open, monitored deployment.
Geographic, generational, and income divides mean the trust deficit is not evenly distributed and will not self-correct uniformly across populations as AI usage grows.
The EU AI Act and similar frameworks represent a regulatory forcing function that may gradually close the gap between usage and genuine accountability, with the most consequential enforcement dates arriving in 2026 and beyond.

More people are using AI tools than at any point in history. Yet the more they use them, the less they seem to trust them. That contradiction sits at the heart of the current AI moment, and it matters enormously for anyone trying to reason clearly about where AI goes from here.

This is not a minor data quirk. It shows up across multiple large, methodologically rigorous studies conducted across dozens of countries. It shows up among developers, office workers, students, and consumers. And it has concrete consequences for policy, for business strategy, and for the way AI technology actually gets embedded into everyday life. If you want to think in probabilities about AI's future, the trust question is one you cannot skip.

The Numbers Are Stark and Consistent

The most comprehensive recent data point comes from a global study led by researchers at the University of Melbourne in collaboration with KPMG. The project surveyed more than 48,000 people across 47 countries between November 2024 and January 2025. Its headline finding captures the paradox cleanly: 66 percent of respondents already use AI with some regularity, yet less than half (46 percent) say they are willing to trust AI systems. The same research team had run a comparable study just before ChatGPT launched in late 2022. When they compared the two data sets, the direction of travel was unambiguous: people have become less trusting and more worried about AI as adoption has increased.

A December 2025 YouGov survey of U.S. adults produced similarly striking numbers. Roughly 70 percent of American adults report using AI tools at least occasionally, with 35 percent using them weekly. Yet fewer than one in five (18 percent) say they would trust an AI system to make a decision or take an action on their behalf. More telling still: when asked how their trust has changed over the prior year, 25 percent said it had decreased, while only 21 percent said it had increased. Net trust, in other words, is moving in the wrong direction even as usage climbs.

The developer community, often assumed to be AI's most enthusiastic constituency, shows the same inversion. According to Stack Overflow's 2025 developer survey, more than 84 percent of developers reported using or planning to use AI tools, up significantly from 2023. Yet the share who said they trust those tools fell to 29 percent in 2025, down 11 percentage points from the year before. Usage up, trust down. The pattern is consistent enough that analysts have started calling it the AI trust paradox.

Why the Two Lines Are Moving Apart

The natural instinct is to assume that more experience with a technology leads to more confidence in it. For most technologies, that is exactly what happens. You learn the quirks, develop mental models, build routines, and familiarity breeds comfort. AI is behaving differently, and there are several reasons why.

The first is exposure to failures. When someone uses a navigation app for years and it occasionally gives a suboptimal route, the consequence is minor and forgiving. When someone uses an AI system for legal research, a medical question, or a financial decision and receives a confidently stated wrong answer, the stakes are different. High-profile hallucination incidents have moved from being niche technical complaints to mainstream news. A Canadian tribunal forced Air Canada to honor a discount after its AI customer-service chatbot confidently cited a nonexistent bereavement fare policy. A database of hallucinated legal citations maintained by researchers at HEC Paris and Sciences Po had logged over 1,300 instances of fabricated or incorrect citations appearing in legal proceedings by early 2025. A 2024 study at the University of Mississippi found that nearly half of AI-generated citations submitted by students had incorrect titles, dates, or authors. These are not abstract risks. They are accumulating in public awareness.

The second reason is a mismatch between what AI promises and what it reliably delivers. The global KPMG-Melbourne study found that 66 percent of users rely on AI output without properly evaluating its accuracy, and 56 percent report making mistakes in their work due to AI. These users are not being careless. Many of them are simply encountering the gap between a system that sounds authoritative and a system that actually is. When you encounter that gap enough times, trust erodes.

The third factor is context sensitivity. The Statista and KPMG data both show that people make sharp distinctions based on what AI is being asked to do. Using AI to summarize a document or draft a first email is one thing. Using it to make healthcare decisions, issue legal advice, or take autonomous action on someone's behalf is quite another. As AI expands its scope from informational to consequential, it runs into much sharper skepticism.

The Geography of Distrust

The trust deficit is not evenly distributed. The Edelman Trust Barometer Flash Poll released in November 2025, which surveyed respondents in Brazil, China, Germany, the UK, and the United States, found dramatic geographic variation. In China, 72 percent of respondents express trust in AI. In the United States, that number drops to 32 percent. In Germany and the UK, trust levels sit near or below 40 percent.

Brazil and China stand out as high-trust environments where optimism is fueling rapid adoption. The developed Western economies, by contrast, show skepticism as the dominant sentiment. Researchers point to several explanations: differences in how societies perceive institutional risk and control, prior exposure to tech-sector scandals and privacy failures, job-market anxieties that make automation feel threatening rather than helpful, and lingering erosion of trust in experts and institutions following recent global disruptions.

Generational and income divides layer on top of the geographic ones. In the UK, there is a 41-point gap in AI trust between the 18-to-34 age group and those 55 and older. Older adults, lower-income households, and women report lower trust levels across multiple surveys. These are not peripheral demographics. They represent the majority of the population in most countries, which means the trust deficit is not a problem that time alone will automatically solve.

The Shadow AI Problem

One counterintuitive consequence of the trust paradox is that low organizational trust in AI does not stop people from using it. It just pushes them to use it in ways that bypass oversight. The KPMG-Melbourne study found that half of the U.S. workforce reported using AI tools at work without knowing whether it was allowed. More than four in ten (44 percent) admitted to knowingly using AI improperly at work. More than half (53 percent) said they had presented AI-generated content as their own.

This is what researchers and practitioners have started calling shadow AI: unofficial, unsanctioned, ungoverned use that flies under organizational radar. It creates a compliance and accountability gap that is arguably more dangerous than controlled, monitored AI deployment. When AI is used openly with governance structures in place, errors can be caught, patterns can be learned, and policies can be adjusted. When it is used in the shadows, the feedback loop breaks.

The same dynamic that produces shadow AI also produces a trust paradox at the organizational level: companies publicly express caution about AI risk while their employees quietly automate large portions of their work without oversight. According to the Diligent Institute's Q4 2025 Business Risk Index, 60 percent of legal, compliance, and audit leaders cite technology as their top risk concern. Yet only 29 percent of organizations have comprehensive AI governance plans in place. Concern is high. Action is not keeping up.

What Hallucinations Are Actually Teaching People

The term "hallucination" entered mainstream vocabulary alongside ChatGPT, describing instances where AI models produce plausible-sounding but factually incorrect outputs. What is important to understand about hallucinations is not just that they happen, but that they are not random bugs that better engineering will entirely eliminate. Research published in 2024 and 2025 has formalized this mathematically: because large language models generate text by predicting probable sequences from learned statistical patterns, they will by design sometimes produce outputs not grounded in fact. The question is not whether hallucinations will be eliminated, but how frequently they occur and in which contexts.

Stanford's Human-Centered AI Institute, in its 2026 AI Index Report, found 362 documented AI incidents in 2025, up from 233 in 2024, a 55 percent year-over-year increase and the highest annual count in the AI Incident Database's history. Columbia Journalism Review tested eight generative AI search tools and found that more than 60 percent of tested news-citation queries produced incorrect answers. Purpose-built legal AI tools still hallucinated on between 17 and 34 percent of challenging legal research queries, according to Stanford HAI research.

These numbers are informing the public's updating of its expectations. Users who encounter an AI confident hallucination in a high-stakes context do not just distrust that output. They update their mental model of AI systems in general. The trust damage is cumulative and, in many cases, proportionate to the stakes of the domain where the failure occurred.

Regulation Is Racing to Catch Up

The trust deficit has not gone unnoticed by policymakers. The KPMG-Melbourne global study found that 70 percent of the global public believes AI regulation is needed, representing a clear public mandate. Governments have responded, most prominently in the European Union.

The EU AI Act, the first comprehensive legal framework on AI worldwide, entered into force in August 2024 and has been rolling out in phases since then. Prohibited AI practices, including social scoring systems and certain forms of biometric surveillance, became enforceable in February 2025. Obligations for general-purpose AI model providers took effect in August 2025. The largest tranche of requirements, covering high-risk AI systems, activates in August 2026, when national and EU-level enforcement formally begins. Penalties for violations of prohibited AI practices can reach 35 million euros or 7 percent of global annual turnover, exceeding even the GDPR fines that reshaped corporate data practices globally.

The implementation picture is imperfect. By the August 2025 deadline, many EU member states had not yet designated the required national oversight authorities. Governance infrastructure is lagging behind the regulatory text. But the direction of travel is clear: AI deployment is moving from a largely self-governed activity to one with meaningful external accountability. The question is how fast that accountability structure can mature relative to the pace of deployment.

Familiarity Is the Underappreciated Variable

Here is where the picture becomes more nuanced, and where the simple narrative of "rising usage, falling trust" risks missing something important. The Edelman Flash Poll data shows that while aggregate trust numbers are low in developed markets, hands-on experience with AI is one of the most reliable predictors of increased trust at the individual level. When AI helps users understand complex ideas, trust in the technology rises sharply in most markets. People who have used generative AI and found it useful report far higher trust than those who have not used it at all.

This creates an interesting dynamic. Distrust in AI among non-users is often rooted in anticipatory concern rather than lived negative experience. Among respondents who resist AI, only around 18 percent report having had a genuinely bad experience. The distrust is largely imagined in advance. And when those same people encounter AI in a context where it performs well, and where their employer provides training and honest framing about what AI can and cannot do, their trust rises.

This is the experience gap: an asymmetry between people who have had mostly positive, low-stakes AI interactions (and who tend to trust it more) and those who have encountered it primarily in high-stakes, underprepared contexts (and who trust it less). Managing which side of that gap people land on turns out to matter enormously.

What Probably Happens Next

This is where reasoning in probabilities becomes genuinely useful. The current pattern, widespread use combined with declining or stagnant trust, is not a stable long-term equilibrium. Something has to give. Here is how to think about the most likely directions.

The highest-probability scenario over the next three to five years is a gradual bifurcation rather than a single clean resolution. In sectors with strong governance structures, clear accountability, and well-calibrated AI deployment (certain financial services, some healthcare applications, enterprise productivity tools with human-in-the-loop oversight), trust is likely to stabilize and gradually recover as track records accumulate and failures are corrected in visible ways. In sectors where AI is deployed quickly without adequate oversight, where hallucinations carry serious consequences, and where users have no mechanism to distinguish reliable outputs from unreliable ones, trust is likely to continue eroding.

Regulation will act as a forcing function in the EU and, increasingly, in other jurisdictions that adopt similar frameworks. As the EU AI Act's high-risk provisions activate in August 2026 and enforcement begins at the national level, companies that have genuinely built governance infrastructure will have an advantage, not just a compliance checkbox but a credibility signal to users who have learned to be skeptical.

The Edelman finding that familiarity with AI correlates with higher trust, combined with the finding that most AI resistance is anticipatory rather than experience-based, suggests a plausible trust recovery mechanism: structured, well-supported first experiences with AI that demonstrate genuine value before they expose users to high-stakes failure modes. Employers who run serious AI training programs, provide honest framing about limitations, and deploy AI in low-risk contexts first are essentially manufacturing the conditions under which the trust loop can close.

For people who want to think clearly about AI rather than just react to it, the trust paradox is a useful frame. It reveals that adoption rates are not the same as acceptance rates, that widespread use does not guarantee widespread confidence, and that the gap between the two is exactly where both the risks and the opportunities concentrate. Getting smart about AI means understanding not just what the technology can do, but which conditions allow it to do those things in ways that people can actually rely on.

The probability that trust and usage realign quickly and smoothly is low. The probability that governance, transparency, and accumulated experience slowly close the gap is considerably higher, but only if the institutions building and deploying AI treat trust as an engineering problem rather than a communication one.

Sources

Frequently asked questions

Several forces push usage upward regardless of trust: competitive and social pressure to adopt AI tools, employer mandates, the convenience of AI-powered products that people use without consciously choosing them, and the fact that many everyday AI interactions (autocomplete, recommendations) are low-stakes enough that distrust does not translate into avoidance. Trust, in this context, measures something different from habit or necessity. People can use a tool routinely while still doubting whether they can rely on it for important decisions.

Shadow AI refers to the use of AI tools at work without organizational knowledge, approval, or governance oversight. Research from the University of Melbourne and KPMG found that half of U.S. workers used AI tools at work without knowing if it was permitted. This matters for trust because unmonitored AI use means errors go uncorrected and unlearned from, accountability is unclear when something goes wrong, and the organization cannot build the institutional knowledge needed to deploy AI responsibly at scale.

Current mathematical research suggests complete elimination of hallucinations is not achievable given how large language models fundamentally work. Because these systems generate text by predicting statistically probable sequences from learned data, they will sometimes produce outputs that are not grounded in fact. The practical goal is reducing hallucination rates in high-stakes domains through techniques like retrieval-augmented generation, multi-model verification, and human oversight, and being transparent with users about where uncertainty exists.

The relationship is complicated. At the aggregate level, usage has increased while average trust has declined. But at the individual level, people who use AI and find it genuinely helpful in low-stakes contexts tend to report higher trust. The Edelman Flash Poll found that helping users understand complex ideas through AI produced large trust gains. The implication is that how someone's first experiences with AI are structured matters enormously. Positive, well-supported first encounters can build trust; high-stakes failures without context can destroy it.

The EU AI Act takes a risk-based approach: the higher the potential harm from an AI application, the more stringent the requirements. It bans certain AI practices outright (social scoring, certain biometric surveillance), requires transparency and human oversight for high-risk systems, and imposes governance obligations on providers of general-purpose AI models. The theory is that mandatory accountability structures, enforced by financial penalties that exceed even GDPR levels, will create external pressure for AI systems to be built and deployed in ways that are actually trustworthy rather than merely claimed to be.

Watch the rate of consequential public failures versus the rate of demonstrably beneficial outcomes in high-visibility domains. Trust in AI is primarily shaped by stories, and the stories people encounter about AI in healthcare, law, finance, and education are disproportionately influential. If the dominant narrative shifts from hallucination-caused harm to AI-assisted diagnosis or AI-enabled research breakthroughs, the trust trajectory will likely improve. If high-profile failures continue to outpace compelling success stories, skepticism will deepen even as usage continues to grow.