Is trust the same as reliability?

No. Reliability is a system characteristic; trust is a person's attitude toward relying on it.

Can distrust be reasonable?

Yes, particularly where relevant evidence is weak or consequences exceed available safeguards.

Should every user trust a system equally?

Not necessarily. Their tasks, experiences, responsibilities, and exposure to consequences may differ.

AI Glossary / Governance, ethics and safety / Advanced

Trust

Definition

Trust is a person's confidence that an AI system can be relied upon in a particular situation. It concerns an attitude toward the system and its role, especially where outcomes are uncertain and the person accepts some vulnerability. Appropriate trust is supported by evidence and remains sensitive to limitations.

What it means

Trust is not simply liking an interface or believing that AI is generally useful. Someone may trust a tool to suggest meeting titles while withholding trust in its interpretation of an unfamiliar contract. The relevant task, consequences, and available safeguards matter.
Trust also differs from trustworthiness. Trust describes an attitude; trustworthiness concerns properties and practices that justify reliance. A persuasive system can attract confidence it has not earned. A reliable system can receive little confidence because users have encountered poor communication or lack useful evidence.
A third concept is reliance: what people actually do with the output. Accepting a suggestion does not reveal someone's attitude by itself. They might follow it because they are rushed, required to use it, or unable to find an alternative.

How it works

People form expectations from experience, reported performance, explanations, institutional reputation, and cues in the interface. These influences can help or mislead. Fluent wording, for example, does not establish factual accuracy.
Organizations can support appropriate trust by defining the intended use, evaluating performance under relevant conditions, communicating limitations, and providing effective review and recovery processes. NIST describes several trustworthiness characteristics, including validity, reliability, safety, security, transparency, and fairness. Evidence about one characteristic does not establish all the others.
Assess attitudes and behavior separately. Surveys can examine reported confidence, while observations can show when people accept, reject, or check advice. Compare those behaviors with the quality of the advice and the final outcome. The aim is well-founded reliance, rather than maximizing acceptance or questionnaire scores.

Three practical examples

A support assistant drafts answers from an approved knowledge base. Staff gain confidence after checking its references, but continue reviewing questions outside the documented scope. Trust is conditional on the task and evidence.
A stock-planning tool performs well for established products. A buyer checks its suggestions more carefully for a new product with little history. Lower reliance in that situation can be appropriate rather than a failure of adoption.
A writing assistant presents an incorrect statement in polished language. A user accepts it without checking. The action shows reliance, but neither establishes the user's underlying attitude nor demonstrates that the system deserved confidence.

Think of it this way

Think of trusting a colleague with a particular responsibility. You might rely on their expertise in one subject and seek another opinion in a different subject. Familiarity helps form expectations, but confidence should remain connected to demonstrated capability and the consequences of an error.

Worked example

Consider a constructed evaluation containing 100 AI suggestions: 80 correct and 20 incorrect. Two users review the same suggestions independently.
User A accepts 72 correct suggestions and 16 incorrect suggestions. Their overall acceptance rate is 88/100 = 88%. They accept 72/80 = 90% of the correct advice and 16/20 = 80% of the incorrect advice.
User B accepts 64 correct suggestions and four incorrect suggestions. Their overall acceptance rate is 68%. They accept 80% of correct advice and 20% of incorrect advice.
A higher overall acceptance rate therefore does not establish better judgment. User B accepts fewer incorrect suggestions, but also rejects more correct suggestions. Evaluating the complete workflow requires the cost of checking, the decisions made after rejection, and the consequences of different errors.
These figures measure observed acceptance in this test. They are not trust scores, and they do not prove why either user behaved that way. A separate assessment of attitudes could help explain the pattern without substituting for the behavioral evidence.

Illustrative case study

Fictional company Willow Stock introduces an assistant for replenishment suggestions. Early demonstrations involve familiar products with stable demand. Buyers become comfortable with the tool and start accepting suggestions for newly introduced products as well.
The team discovers that confidence has extended beyond the conditions covered by its evaluation. It separates performance reports for established and new products, displays relevant limitations, and gives buyers a practical route for additional review.
The team then examines both reported confidence and purchasing decisions. It checks whether buyers recognize uncertain situations and whether final decisions improve. This fictional case illustrates the importance of aligning reliance with evidence. It does not assume that displaying an uncertainty label automatically changes behavior or improves outcomes.

Common misconceptions and mistakes

“More trust is always better.” Excessive confidence can encourage reliance beyond demonstrated capabilities.
“Acceptance proves trust.” Workload, incentives, rules, and lack of alternatives can also influence behavior.
“A transparent system is necessarily trustworthy.” Disclosure does not by itself establish accuracy, safety, or fairness.
“One good experience justifies every use.” Performance and consequences vary across tasks and conditions.
“An explanation earns trust automatically.” A plausible explanation may be incomplete, misleading, or irrelevant to the decision.

Business use

Treat trust as a relationship between people, a system, and a defined use. Identify which decisions depend on its outputs and what evidence supports that dependence. Give users realistic opportunities to question results and recover from errors.
Track reported attitudes alongside actual behavior and outcome quality. Investigate both unwarranted acceptance and rejection of useful advice. When the system or workflow changes, reassess the evidence rather than assuming earlier confidence remains appropriate. Training should develop judgment about when and how to rely on the tool.

Check your understanding

Why is User A's 88% acceptance rate insufficient?

Answer: It combines acceptance of correct and incorrect advice.

What does a confidence survey measure?

Answer: Reported attitudes, rather than the full pattern of reliance or final decision quality.

Why define the task before evaluating trust?

Answer: Evidence supporting one use may not justify another.

Related terms

Learn next

Choose an AI-assisted decision. Specify evidence that would justify reliance, circumstances requiring additional checking, and separate measures for user attitudes, observed behavior, and final outcomes.

Sources

Russell and Norvig, AIMA glossary, trust: https://github.com/aimacode/aima-glossary/blob/master/glossary.md NIST, AI Risk Management Framework, characteristics of trustworthy AI: https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/ Scharowski and colleagues, Trust and Reliance in XAI—Distinguishing Between Attitudinal and Behavioral Measures, 2022 position paper: https://arxiv.org/abs/2203.12318

Previous: Truncated singular value decomposition Next: Trust region