Consumer AI survey · US and UK · April 2026
AI chatbots: navigating distrust and usefulness in everyday life
83.9% of active users are concerned about 2 or more AI chatbot issues while continuing to use AI chatbots weekly or more.
80.9% of AI chatbot users feel uncertain at some point when using AI chatbots, and 42.4% of all users say it's specifically because they can't tell if the response is accurate.
35.5% of users rarely or never check AI chatbot responses that sound confident.
89.5% feel the responsibility and workload of checking AI chatbots falls on them, not the AI chatbot.
51.4% of non-users say they still would not use AI chatbots even with a strong security guarantee.
1,500 users
Surveyed across the US and UK
Users of the tools:
ChatGPT
Gemini
Copilot
Claude
and others
United States
United Kingdom
United States (n=750)
Female
51.3%
Male
48.7%
United Kingdom (n=750)
Female
50.5%
Male
49.5%
Reflections from survey participants
“Nothing makes me feel better about AI. I do not understand why people are trying to replace all human interaction with AI, it’s just downright wrong.”
Male, 55–64, USA
“I’d feel a lot better if AI tools were more ‘open book’ about where they get their information and how they use my data. Right now, it feels like a black box. I want to know exactly what is being saved, who can see it, and have a simple, one-click way to delete my entire history whenever I want.”
Male, 25–34, UK
“I am currently pregnant and honestly Chat GPT has helped with a lot of my basic questions when my OB office wouldn’t answer.”
Female, 18–24, USA
“Nothing really, I believe you take just as much of a risk interacting with AI as you do when interacting with humans, probably even less because humans can be jaded and dishonest or deceitful on purpose. AI doesn’t tend to seek revenge or get jealous...”
Female, 25–34, USA
“Nothing in the tech industry is being regulated anymore and all the big tech bros are quite vocal about the harm they wish to society and humanity in general in pursuit of even more money for themselves.”
Female, 55–64, USA
“What worries me most about AI chatbots is knowing too much personal data about you. And acting on the person’s behalf, without authorization. There needs to be a moratorium on all AI.”
Female, 25–34, USA
“If there was a regulation to ensure that people double check the chatbot info or a popup that says please verify with one more page to ensure the accuracy of the chatbot.”
Male, 35–44, USA
“It generally gives quick and accurate responses, but it’s my responsibility to fact check everything and look for biases.”
Male, 35–44, USA
“It’s like having a super fast research assistant who’s amazing at brainstorming but occasionally gets a bit too confident about facts they’ve made up.”
Female, 25–34, UK
“It feels like having instant access to a knowledgeable, patient assistant who can help you think, learn, and create on demand — but one that can sometimes be confidently wrong, lack real-world judgement, and or miss context in subtle ways.”
Male, 35–44, UK
“I do not trust AI chat bots at all. They can easily give wrong information and can only do so much until you get stuck in a loop, such as when calling customer service and dealing with a chat bot.”
Female, 35–44, USA
“When it gets something wrong and I correct it and it goes ‘of course, you are right ...’. That makes me doubt it’s overall reliability.”
Female, 45–54, UK
Introduction
We all use products and services we don’t trust
AI chatbots occupy a strange place in everyday life. They are increasingly familiar, frequently useful, and widely questioned. People turn to them for quick answers, advice, and decisions, yet remain uneasy about whether their responses can be trusted. Using AI chatbots does not mean that users trust them, or the companies behind them. Especially in an age where personal information is harvested everywhere. Every app tracks something, every site asks for consent it knows nobody will read, every purchase and scroll gets logged by someone, somewhere. AI chatbots do not sit outside that system. They sit at the center of it, and they ask for more than almost anything else does. Not just what you click, but what you actually think, worry about, and want to know.
The danger lies in the fact that users are sharing extremely personal and private parts of their lives to AI chatbots. We found that at least 33.9% of users are asking AI for personal advice on health, money, and relationships; the exact categories people are most guarded about everywhere else.
That is why simply asking users, "do you trust AI?" is not useful. Trust is so complex. It is situational and contextual, meaning that trust changes with the task, the stakes, and the moment. Someone may distrust AI in principle but accept its answer when it sounds convincing, the question feels simple, or checking would take too much time. What people say about trust and what they do with an answer can be very different.
We surveyed 1,500 consumers in the United States and United Kingdom to examine this gap. Instead of treating trust as a single attitude, we looked at the behaviors that surround it: when users verify an answer, what prevents them from checking, what makes them accept a response and whether those habits change as AI becomes part of more areas of their lives. We also asked non-users what keeps them away and whether stronger security guarantees would change their minds.
AI chatbot: any software program that simulates human conversation. Early chatbots used rigid, pre-written rules to answer basic questions (like customer service bots). Modern AI chatbots leverage generative large language models (LLMs) — such as ChatGPT, Claude, and advanced corporate support agents — to understand complex human intent, context, and nuance. Despite these significant technological leaps, the term "chatbot" still fundamentally describes the conversational nature of the user interface rather than the specific backend intelligence powering it.
Theme 1 · Feeling responsible vs taking action
Almost everyone says checking is their responsibility. Yet 70.1% check only sometimes, rarely, or never.
70.1% report checking only sometimes, rarely, or never, despite 89.5% assigning themselves full or partial responsibility for verification. This gap is consistent with automation bias, a documented effect in human-computer interaction (HCI) research where users reduce vigilance toward a system’s output as that output appears fluent and confident, independent of its actual accuracy. Verification is effortful. It requires deliberate cognitive processing, whereas accepting an answer at face value does not. Under time or attention pressure, the lower-effort path wins even when a person holds the stated belief that they should verify. The responsibility judgment and the checking behavior are governed by separate cognitive processes, one reflective, one automatic. Only the reflective one has shifted.
Roughly 9 in 10 (89.5%) say it is fully or mostly their responsibility. Only about 1 in 10 lean towards it being the AI’s job.
How often users check AI chatbot answers
Figure 1 · Self-reported responsibility to verify AI chatbot answers and frequency of verification
The stakes attached to this gap are not evenly distributed. Reported use remains concentrated in low-stakes tasks such as search, quick answers, and learning, but 33.9% of users report using AI chatbots for personal advice on health, money, or relationships, and 40.1% report work use. Verification effort does not scale with the sensitivity of the task. The same cognitive shortcut that reduces checking on a casual query operates on a financial or medical one, because the decision to check is driven by how confident a response sounds, not by what the response concerns.
What stops users of AI chatbots from checking the answers?
The two largest barriers, no reliable source to check against (37.5%) and the answer already looking good enough (34.4%), are structural rather than motivational. Verification depends on two things: an external standard to measure the answer against, and enough domain knowledge to apply that standard. Time (19.2%) and not knowing how to check (15.4%) rank well below both, which suggests the constraint is not effort or skill. It is the absence of a reliable fixed reference point to check against.
When people do check, most look outside the tool: 76.8% search online and 53.5% rely on their own knowledge. One in five instead ask the same chatbot again. This is not equivalent to independent verification. A second query to the same model draws on the same training data and the same failure modes as the first, so it can raise a user’s confidence without changing the accuracy of the answer underneath it. In research on advice-taking and forecast aggregation, this distinction is central: corroboration only reduces error when the sources being consulted are independent of one another. Asking the same system twice does not meet that condition.
Figure 2 · Reasons for not verifying AI chatbot answers
It is not how often you use AI. It is how many different tasks you use it for.
41.4% of broad users rarely or never check an answer, against 25.8% of scoped users. The reversal is symmetric: 36.8% of scoped users check regularly, against 24.7% of broad users. The category in between, checking sometimes, barely moves, 33.8% versus 37.4%. Whatever separates these two groups operates at the extremes of the scale, not across it.
Broad users are also the majority here, 64% of the sample. This is not a fringe pattern sitting off to one side. It describes how most people who use AI chatbots are currently using them.
Broad users: describe themselves as using AI chatbots for many tasks, or freely for many kinds of tasks.
Scoped users: describe themselves as using AI chatbots only for certain kinds of tasks, or only for low-stakes or simple tasks.
Figure 3 · Checking answers frequency among broad vs. scoped AI users
One plausible reading: once someone uses AI across many different kinds of tasks, a single, generalized sense that "this tool is fine" can take over from a task-by-task judgment. Someone who only ever asks it about recipes builds a narrow, specific expectation, tested against one kind of answer repeatedly. Someone who asks it about recipes, code, medical symptoms, and legal questions in the same week has no comparable basis for judging any one of those answers, and appears to default to trusting the tool generally rather than assessing each domain on its own terms. Breadth of use may be substituting for depth of judgment.
A second, equally plausible reading runs the other way. People who already checked less might be the ones who felt comfortable letting AI into more parts of their life in the first place, precisely because they were not scrutinizing each answer closely enough to notice when it failed. In that version, low verification is not a consequence of broad use; it is a pre-existing habit that broad use simply exposes more of.
The chart cannot distinguish between these two stories, and its own note says so: association, not causation. What it does establish is that the pattern is real, sizable, and sitting inside the group that now makes up most AI chatbot users, not a minority carve-out.
Theme 2 · Cognitive overload and checking
Once an answer sounds right, almost half don’t question it further.
The expectation was that heavier cognitive load would increase vigilance: the harder a task felt, the more a person would check. It does not. Most people who stop checking do not cite fatigue, time pressure, or distraction. They cite the answer itself. Once it sounds sufficient, checking stops, regardless of how much capacity the person had left.
Figure 1 · Ranked reasons for not questioning an AI chatbot’s answer
47.3% say that if a response sounds right, they do not question it further, more than double the next highest reason. The next three, relying on quick answers over careful ones (22.6%), not having the time or energy to check (21.8%), and going with the first response without much thought (18.8%), cluster tightly together and describe the same underlying constraint: limited time or effort. "Sounds right, stop there" is a different kind of reason. Of the people who cite it, 61.2% give no other reason alongside it. This is closer to satisficing, stopping once an answer clears a "good enough" bar, than to fatigue. The signal that ends the search comes from the answer, not from how much energy the user had left. Only 9.5% recognize none of these five behaviors in themselves, this is close to universal, not a trait that separates careless users from careful ones.
The more users lean on AI, the lower the bar for accepting an answer.
This shortcut becomes more common as AI occupies more of a person’s activity. The increase appears with usage frequency, but it is sharper when comparing people who use AI for a narrow set of tasks with those who use it freely across many parts of life and work.
Acceptance of answers by how often they use AI
Acceptance of answers by how widely they use AI
Figure 2 · Acceptance rate of AI chatbot answers, by frequency and breadth of AI use
Acceptance rates climb from 44.1% among weekly users to 58.1% among several-times-a-day users, but almost all of that increase happens in one step. Moving from weekly to daily use adds 12.4 points. Moving from daily to several times a day adds only 1.6. Frequency’s effect saturates early. Breadth does not: it runs from 28.6% among people who keep AI to low-stakes tasks to 59.7% among people who use it for many kinds of tasks, a 31.1-point range, more than double the spread produced by frequency alone. This is the same "good enough" threshold from Figure 1, but shaped by exposure across contexts rather than by the individual moment of use. The survey cannot establish causality, and greater exposure does not automatically produce a higher verification standard.
Most AI users have engaged in more than one risky behavior.
35.3% have used AI to fact-check AI, more than the 30.9% who relied on an answer without checking it at all. That is not really a safe behavior. Asking a second instance of the same kind of system to confirm the first does not produce independent evidence, so the single most common response to this complacency issue is not actually verification. 33.2% have trusted an answer because it sounded confident rather than because they checked it, and 25.2% have followed AI advice on a major real-life decision. These behaviors are not spread thinly across different people. 76.6% of AI users report doing at least one of them, and of those, 73.0% report doing two or more, averaging 2.6 each. The risk compounds within a subset of users rather than scattering as isolated incidents across the population.
Figure 3 · Prevalence of risk behaviors among AI chatbot users
Theme 3 · People who don’t use AI chatbots and security guarantees
Would security promises encourage people who don’t use AI chatbots to start using them?
The original hypothesis was that non-users stay away because they are cautious. So fix the fear, gain the user. The data shows that the two biggest barriers are almost equally split between distrust and irrelevance, and those impacted by these respond to a security pitch very differently.
Two equally important barriers: distrust and simply not needing AI
The two largest barriers among non-users are effectively tied: some do not trust AI chatbots, others simply do not see a need for them. The two groups overlap more than the headline numbers let on. 47 non-users, one in six, gave both reasons. Strip that overlap out and you get 73 people who name distrust alone and 72 who name irrelevance alone, nearly identical in size. Someone who distrusts AI has usually thought it through and decided against it. Someone who calls it irrelevant may never have thought about it hard enough to form an objection in the first place. Both end up not using the product, for reasons that don’t share a cause, and a 0.4 point gap between the two means neither can claim to be the bigger barrier.
Figure 1 · Prevalence of stated barriers to AI chatbot use among non-users
Security motivates those who don’t feel a need for AI today, more than those who do not trust it.
A promised security layer doesn’t land the same way for everyone. It does more for people who were never fully convinced AI is unsafe, and less for people who’d already made up their mind. The 47 non-users who cite both reasons are excluded from both lanes.
Still would not use. Distrust barrier: 70%. Irrelevance barrier: 51%. Distrust is 18.5pp harder to move, significant at p<0.05
Is the gap real?
The 18.5 pp difference in "still would not use" has a 95% confidence interval of ±15.6pp. It clears zero, so the distrust group is genuinely harder to move. But the subgroups are small; the true gap could be anywhere from about 3 to 34 points.
Figure 2 · Likelihood of reconsidering AI use after a security guarantee, by main barrier
Among non-users who cite distrust alone, 25% say a security layer would make them more likely to use AI. Among those who cite irrelevance alone, it’s 33%, about a third higher. Flip the question and the same pattern holds: 70% of the distrust-only group say they’d still avoid AI even with the guarantee, against 51% of the irrelevance-only group. The 47 people who gave both reasons track closer to the distrust group; only 21% say they’d be more likely to use it. Whatever drives someone to distrust AI in the first place doesn’t seem to loosen its grip just because a feature promises to catch bad content.
What non-users would want: scam detection and data protection first
When non-users describe the security features that would matter most, they do not lead with accuracy scores or technical explanations. They prioritize concrete protections — detecting scams, preventing theft or misuse of data, blocking harmful manipulation — the kind of capabilities a traditional security product offers, not an AI‑accuracy feature.
Figure 3 · Prevalence of desired security features among non-users
Four features clear 37%: catching AI‑generated scams (46.1%), keeping data from being stored or shared (45.0%), blocking harmful content (42.9%), and flagging answers that might be wrong (37.5%). Everything else, warning before sharing sensitive info, explaining how data is used, citing sources, scoring confidence, tops out at 28.9%. People aren’t asking to be let in on how the system works. They’re asking for it to stop doing the things that could hurt them. There’s a real tension with Figure 1 here: only 16.1% of non-users say scams are why they’re staying away today, yet scam protection is the single most requested feature if they reconsidered. The worry seems to sit under the surface. It just doesn’t surface until you ask a different question.
Methodology
How the survey was run and how to read the statistics
Sample
1,500 consumers across the United States and United Kingdom, surveyed April 2026.
Bases
AI users n=1,045–1,165 depending on question; non-users n=280. Exact bases under each figure.
Confidence intervals
Intervals in tooltips are 95% Wald intervals for a single proportion. Group differences use two-proportion z‑tests, noted where shown. Unit charts round to whole units, with exact values labeled.
Reading the splits
Subgroup comparisons are associations, not causation. Multi-select questions sum above 100%. US and UK combined; regional gaps within a few points.
A note from the author
The challenge has less to do with AI itself and more to do with a mismatch: models are becoming increasingly fluent and embedded in virtually everything we use. But this is happening faster than our ability to verify this has kept up. The giveaway cues we used to rely on — such as the hesitation, inconsistency, or phrasing — are becoming difficult to detect. What’s left sounds correct and convincing regardless of whether it is actually right or wrong.
I don’t think the fix is asking people to be more careful. Users already believe checking the accuracy of the content is somewhat their responsibility. That responsibility sits with the people building these tools, not the people using them. A system that can sound confident regardless of whether it’s right should be the one required to show its uncertainty, not the user.
Amel Bourdoucen, D.Sc. (Technology)
Senior User Researcher, Human-Centric Cybersecurity and Privacy