How much do people trust AI — and is that trust calibrated?
Trust in AI is not one attitude and it does not translate into one behaviour. Across five Human Clarity Institute samples, 32% to 39% reported high trust that AI would act in their best interests; a separate decision-dependence sample recorded 53%. Yet trust was not linked to less checking in the main calibration sample. The clearest differences appeared in acceptance, delegation and monitoring—making the relationship between trust and behaviour more informative than a trust score alone.
AI trust: the benchmark at a glance
How much do people trust AI?
AI trust is neither universal nor automatically uncritical. Across five HCI samples, 32% to 39% reported high trust that AI would act in their best interests; one separate decision-dependence sample reported 53%. High-trust respondents usually still checked outputs. The sharper calibration question is whether trust changes acceptance, delegation, monitoring and override.
What does this AI trust benchmark measure?
AI trust is a person's willingness to expect that an AI system will be reliable, useful or aligned with their interests under uncertainty. It is broader than believing one answer is accurate and different from simply using AI often.
HCI examines trust through several related measures: trust that AI will act in the user's best interests, trust in AI-generated information, confidence relying on outputs and the behaviours that follow—checking, accepting, delegating, monitoring, overriding and retaining responsibility.
Trust calibration usually means how closely a person's trust corresponds with a system's actual capability in a specific context. HCI's survey data do not observe the performance of each system respondents used. This benchmark therefore measures behavioural alignment between stated trust and checking, not objective calibration against system accuracy.
What proportion of people report high trust in AI?
Five samples cluster between 32% and 39%
Does frequent AI use relate to higher trust?
Yes, descriptively. High trust was more common among frequent users in each HCI dataset where the same frequency comparison was available. The size of the difference varied by instrument.
Frequent users reported high trust more often
Do people who trust AI still verify its outputs?
Checking remained high on both sides of the trust divide
Self-reported checking does not establish how thoroughly, consistently or effectively an output was verified. The separate AI Verification benchmark examines this behaviour in more detail.
Where does higher trust change AI behaviour?
Higher trust was most clearly associated with acceptance and delegation. In the Decision-Making and Digital Systems 2026 sample, respondents at the high and low ends of trust differed sharply even though high levels of verification and retained responsibility remained common.
Acceptance and high-stakes delegation showed large differences
Does higher trust reduce monitoring?
Confidence without monitoring was far more common with high trust
Trust appears to change how much oversight feels necessary without automatically removing attention to results or the sense of responsibility for them.
Can people trust AI and still retain control?
Yes. In HCI's Autonomy, Control & Perceived Independence 2026 sample (n=352), 93% of high-trust respondents reported feeling in control when using AI, compared with 54% of low-trust respondents. The ability to override remained high in both groups: 89% and 86% respectively.
The same coexistence appeared in delegated action. High- and low-trust respondents were almost equally likely to say they paid attention to outcomes and accepted personal responsibility for negative results, even while their monitoring confidence differed sharply.
What is the HCI Trust Calibration Grid?
The HCI Trust Calibration Grid is a descriptive framework that places stated trust alongside active checking. In the Trust Calibration in Information Environments 2026 sample, 392 respondents had valid responses on both measures. The four groups are patterns, not fixed personality types or ratings of good and bad AI use.
Trust with active checking
110 respondents reported both high trust that AI acts in their best interests and high active checking.
Cautious checking
229 respondents did not report high trust but did report high active checking.
Trust with limited checking
17 respondents reported high trust without reporting high active checking.
Limited trust with limited checking
36 respondents reported neither high trust nor high active checking.
What changes when trust combines with verification, reliance and agency?
Trust × Verification
High trust did not correspond with less checking in the central calibration sample. This makes verification behaviour an independent signal rather than the opposite end of a single trust scale.
Trust × Reliance
Acceptance, difficult-decision reliance, high-stakes delegation and reduced monitoring all differed sharply by trust group. The AI Reliance benchmark provides the broader context for when support becomes routine handover.
Trust × Agency
High trust frequently coexisted with reported override capability, perceived control and responsibility for outcomes. Trust alone therefore cannot establish whether a person has retained or surrendered decision ownership.
What does wider research show about trust in AI?
Global willingness to trust AI remains below a majority
The University of Melbourne and KPMG surveyed more than 48,000 adults across 47 countries between November 2024 and mid-January 2025. Sixty-six percent reported regular AI use, while 46% said they were willing to trust AI systems. Adoption and stated trust therefore did not move together as one measure.
Trust depends on what or whom people are being asked to trust
The nationally representative 2026 Bentley–Gallup survey found that 27% of U.S. adults trusted businesses a lot or some to use AI responsibly. In a separate Pew Research Center survey of 5,023 U.S. adults fielded in June 2025, 60% wanted more control over how AI is used in their lives and 17% felt comfortable with their current level of control. These questions concern institutions and personal control, not trust in the accuracy of one AI output.
Public opinion varies across countries and institutions
Stanford HAI's 2026 AI Index synthesised multiple public-opinion sources and reported that 31% of U.S. respondents trusted their government to regulate AI responsibly, compared with a 54% global average. This is regulatory trust rather than behavioural trust in AI, but it demonstrates why headline trust figures should not be treated as interchangeable.
Appropriate reliance is the established goal—not maximum trust
Lee and See's foundational 2004 review defines calibration in relation to how closely trust corresponds with an automated system's actual capabilities. Parasuraman and Riley's 1997 review distinguishes use from misuse and disuse, showing why both too much and too little reliance can be behaviourally important.
People can favour algorithmic advice under some conditions
Across six experiments, Logg, Minson and Moore found that laypeople incorporated more advice when they believed it came from an algorithm than when they believed the same advice came from another person. The work describes algorithm appreciation, while also showing that expertise and the identity of the human comparison can change the pattern.
Interface friction can change overreliance
In an experiment with 199 participants, Buçinca, Malaya and Gajos found that cognitive-forcing designs reduced acceptance of incorrect AI advice compared with simpler explainable-AI designs. The designs that reduced overreliance most were also rated least favourably, showing that behavioural calibration and user preference need not align.
What can be concluded from this AI trust benchmark?
High stated trust was a minority response in five HCI samples, ranging from 32% to 39%, although a separate decision-dependence sample recorded 53%. The difference between instruments is part of the finding: trust changes with sample and behavioural context.
Frequent users reported high trust more often, but the cross-sectional data cannot determine whether repeated use increased trust, trusted systems invited more frequent use, or another factor affected both.
Trust did not remove checking in the main calibration sample. It was more clearly associated with accepting outputs unchanged, delegating high-stakes decisions and reducing monitoring. At the same time, high trust often coexisted with override capability, attention to outcomes and retained responsibility.
AI trust is therefore best interpreted as one part of a wider behavioural pattern. Verification, reliance, oversight, override and agency must be measured separately. None of the HCI results tests whether an individual's trust matches the objective performance of a particular AI system.
How can I tell whether my trust in AI is well calibrated?
A trust score alone cannot answer whether your behaviour is well calibrated. The useful comparison is whether your trust sits alongside checking, source comparison, decision ownership, monitoring, override and the ability to act without AI.
The Human Clarity Institute AI Identity & Behaviour Assessment asks 39 questions across nine dimensions of AI behaviour. It takes about 3–4 minutes and provides free personalised results comparing your responses with all HCI participants, people in your age group and people who use AI as often as you do.
The assessment provides a behavioural reference point for reflection and comparison. It does not test the accuracy of the AI systems you use, diagnose dependence or determine whether trusting a specific output is correct.
Frequently asked questions about AI trust
How much do people trust AI?
Trust varies by question, sample and context. Across five HCI samples collected in 2025–2026, 32% to 39% reported high trust that AI would act in their best interests. A separate AI Decision Dependence & Cognitive Caution sample recorded 53%. These are sample descriptions, not national population estimates.
Is trusting AI the same as relying on AI?
No. Trust is an expectation about an AI system, while reliance concerns what a person does with it. Someone may report high trust while still checking and overriding outputs, or report limited trust while using AI frequently because a task or workplace requires it.
Does frequent AI use increase trust?
Frequent users reported high trust more often in three HCI datasets: 44% versus 18%, 49% versus 14%, and 68% versus 35%, depending on the instrument. The studies are cross-sectional, so they cannot show whether frequent use increased trust or trust encouraged more frequent use.
Do people who trust AI still verify its outputs?
Yes, in HCI's Trust Calibration in Information Environments 2026 sample. Active checking was reported by 87% of high-trust respondents and 86% of respondents without high trust. This is self-reported checking and does not measure its thoroughness or effectiveness.
What does calibrated trust in AI mean?
In the research literature, calibrated trust means trust that corresponds with a system's actual capability in a specific context. HCI's survey benchmark compares stated trust with checking behaviour, but it does not test the objective accuracy of each AI system. It therefore describes behavioural alignment rather than proving true calibration.
Does higher trust lead to more AI delegation?
Higher trust was associated with more delegation in HCI's Decision-Making and Digital Systems 2026 sample. Sixty percent of high-trust respondents were comfortable delegating high-stakes decisions, compared with 6% of low-trust respondents. The association does not establish that trust caused the difference.
Can someone trust AI and still retain control?
Yes. In HCI's Autonomy, Control & Perceived Independence 2026 sample, reported override capability remained high among both high- and low-trust respondents. Trust, perceived control, verification, monitoring and responsibility are related but separate behaviours and should be measured independently.
What is the HCI Trust Calibration Grid?
The HCI Trust Calibration Grid places stated trust alongside active checking. In one HCI sample, 28.1% showed trust with active checking, 58.4% cautious checking, 4.3% trust with limited checking and 9.2% limited trust with limited checking. The groups are descriptive patterns, not fixed types or evaluations.
How can I compare my AI trust with other people?
The Human Clarity Institute AI Identity & Behaviour Assessment asks 39 questions across nine dimensions and provides free personalised comparisons with all HCI participants, people in your age group and people who use AI as often as you do. It provides a behavioural benchmark, not a test of AI accuracy.
Related HCI benchmarks
Suggested citation: Human Clarity Institute. (2026). AI Trust Benchmark: Trust, Verification and Calibration. https://humanclarityinstitute.com/benchmarks/ai-trust/
Methodology: Learn how HCI builds benchmarks. Dataset sources are linked throughout the page and in the accompanying structured data.
How does your trust in AI compare?
Compare your responses with HCI participant benchmarks across nine dimensions of AI behaviour—including trust, verification, reliance, decision ownership and human agency.
See Where I Sit →- 01All participants in the benchmark
- 02People in your age group
- 03People who use AI as often as you do