Trust in the Age of AI: AI Systems, AI Users, Information, and Appropriate Reliance
Author: Ukrainian Psychological Hub · Published: September 26, 2026 · Editorial Policy
Trust in AI is not a single attitude toward a single object. A person may trust an AI system to summarize a document but not to interpret an ambiguous medical symptom; trust a colleague’s judgment while distrusting the AI tool that colleague used; believe a generated answer is probably correct while refusing to act on it; or feel skeptical of a system while relying on it because checking the answer is expensive, slow, or impossible. In everyday language these situations are all described as “trusting AI.” Psychologically, they are different.
The most useful question in the Age of AI is therefore not “Do people trust AI?” It is “What is being trusted, by whom, for what purpose, on what evidence, in what context, and what behavior follows?” A major 2026 review in Nature Reviews Psychology makes this distinction central: trustworthiness, trust, and trusting behavior are related but separable, and trust is socially embedded rather than reducible to a person’s opinion of a model.
That distinction changes the practical goal. High trust is not automatically good, and low trust is not automatically safe. The goal is appropriate reliance: using AI when its capabilities, evidence, and context justify reliance, withholding reliance when they do not, and changing one’s behavior when the system, task, stakes, or evidence changes. The classic human-factors literature framed this as designing automation for appropriate reliance under uncertainty. Contemporary AI research makes the problem more complex because AI systems are increasingly conversational, generative, socially presented, embedded in institutions, and used to produce information that may circulate far beyond the original interaction.
This article examines trust as a psychological and socio-technical process across four practical objects: the AI system, the human AI user, the information produced or mediated by AI, and the institution or environment in which reliance occurs. It distinguishes perceived trust from actual trustworthiness, self-reported trust from behavior, trust calibration from simple confidence, and appropriate reliance from both overreliance and blanket rejection.
What Does Trust in AI Actually Mean?
Trust is often measured as if it were one quantity: a survey score, a willingness-to-use rating, or an answer to a question such as “How much do you trust this system?” That can be useful, but it can also hide several distinct processes.
In the human-factors tradition, Lee and See described trust in automation as an attitude that helps govern reliance when people face uncertainty and cannot fully inspect how the system works. Their review emphasized that trust becomes especially consequential when complexity makes complete understanding impractical. Later work integrated evidence across automation settings and showed that trust depends on interacting layers of dispositional, situational, and learned factors rather than a single stable preference (Hoff & Bashir, 2015).
AI intensifies this problem because different systems invite different meanings of trust. Trusting a navigation system to estimate traffic is not psychologically identical to trusting a language model to explain a legal contract, an AI companion to respond sensitively, a hiring model to rank applicants, or a clinical decision-support system to flag a patient at risk. The relevant capabilities, harms, verification routes, social expectations, and institutional responsibilities differ.
A 2026 synthesis by Everett, Claessens, Knöchel, and Reinecke argues that “trust in AI” should be understood as inferred, multidimensional, dynamic, agent-specific, individually variable, strategically motivated, and socially embedded. The authors identify six organizing principles: trust is inferred; trustworthiness, trust, and trusting behavior are distinct; trust concerns both morality and performance; and trust is agent-specific, individually variable, and strategically motivated. This framework is especially useful because it prevents one of the most common errors in public discussion: treating a favorable feeling toward AI as evidence that the AI deserves reliance.
Trustworthiness is a property claim about the system or arrangement
Trustworthiness concerns whether the object of trust has properties that justify confidence in a particular context. In technical and governance frameworks, these properties include much more than raw accuracy. The U.S. National Institute of Standards and Technology identifies characteristics of trustworthy AI that include being valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. NIST also stresses that these characteristics are socio-technical and must be balanced in relation to context of use.
A system can therefore be accurate on average and still be untrustworthy for a particular use. It may perform poorly for a relevant subgroup, drift after deployment, fail under distribution shift, expose sensitive information, give users no practical way to challenge errors, or be inserted into a workflow in which no one is clearly accountable for decisions.
Trust is a psychological state or attitude
Trust concerns what the person expects, believes, or feels about the system, source, user, or institution. It can be informed by evidence, but it can also be shaped by presentation, familiarity, social cues, brand reputation, prior experiences, perceived expertise, personality, workload, institutional endorsement, and the cost of independent checking.
A person can trust an untrustworthy system. A person can also distrust a trustworthy system. The psychological state and the warrantedness of that state must therefore be analyzed separately.
Trusting behavior is what people actually do
Behavior matters because people do not always act consistently with what they report. Someone may say that they distrust AI and still copy an AI-generated recommendation into a report. Another person may report high trust but independently verify every important claim before acting.
This gap is not a technicality. A 2026 preregistered experiment on LLM-assisted decision-making found that visual confidence cues could improve participants’ subjective sensitivity to model accuracy while simultaneously increasing behavioral agreement with incorrect outputs. The study concluded that better subjective calibration did not necessarily produce better reliance behavior. Measuring “trust” without measuring what users actually accept, reject, check, or override can therefore give a misleading picture.
Trust, Reliance, Calibration, and Appropriate Reliance
Trust and reliance overlap, but they should not be used as synonyms.
Reliance is behavioral. It includes accepting a recommendation, using an output, delegating a task, following a route, submitting a generated answer, allowing an agent to act, or choosing not to verify a result. Trust can influence reliance, but reliance can also occur for other reasons: convenience, time pressure, lack of expertise, organizational policy, absence of alternatives, or the simple fact that the AI is built into the workflow.
This is why appropriate reliance is a stronger practical target than “increasing trust.”
If an AI recommendation is correct and relevant, rejecting it can be underreliance. If it is wrong or poorly matched to the context, accepting it can be overreliance. Appropriate reliance requires discrimination: the user should be more likely to accept advice when it deserves acceptance and more likely to reject, verify, escalate, or contextualize advice when it does not.
In healthcare informatics, Benda and colleagues argued explicitly that systems should be designed for appropriate reliance rather than maximal trust. Their framework distinguishes calibration—how closely trust corresponds to system performance—from resolution, the user’s capacity to adapt trust across changing functions, situations, and time. This is an important extension because a single global trust score can be “correct on average” and still fail where it matters. A user may trust a system appropriately for routine classification but inappropriately transfer that trust to rare cases, new populations, or tasks the system was never validated to perform.
Appropriate reliance therefore has at least three layers. First, the user needs an approximately accurate model of what the system can do. Second, the user needs enough situational information to judge whether the current case falls within those capabilities. Third, the interface and surrounding workflow must make it realistically possible to act on that judgment.
Why Trust Becomes Harder in the Age of AI
AI changes the psychology of trust because it combines several features that older forms of automation often kept separate.
Generative systems communicate in natural language. They can produce fluent explanations, adapt tone, sustain dialogue, imitate social responsiveness, generate apparent reasons, and produce outputs in domains ranging from coding and finance to relationships and health. This can make the interaction feel legible and interpersonal even when the user has limited access to the system’s evidence, training distribution, internal uncertainty, or failure modes.
The same system can also appear competent across many domains. Breadth itself becomes a cue. If an assistant writes good code, translates well, and summarizes a paper accurately, the user may generalize that success to a domain in which the system is less reliable. Psychologically, this is a transfer problem: trust earned in one context can spill into another without equivalent evidence.
AI also lowers the cost of asking. A person who once consulted a professional, searched several sources, or postponed a decision can now obtain a plausible response in seconds. Reduced friction changes behavior. The question is no longer merely whether the answer is correct; it is whether convenience changes how much scrutiny the answer receives.
Finally, AI is increasingly embedded rather than visited. Recommendation engines rank what people see. Search systems synthesize answers. Workplace software drafts, classifies, prioritizes, and predicts. AI features can become part of the environment rather than a visibly separate “machine.” When that happens, trust shifts from an isolated human–machine interaction to a distributed relationship among users, systems, organizations, interfaces, data, and information sources.
Six Principles for Understanding Trust in AI
The 2026 Nature Reviews Psychology framework provides a useful backbone for the current evidence. Each principle has a direct practical implication.
Trust is inferred
Trust cannot be observed directly. Researchers infer it from self-report, choice, compliance, advice-taking, delegation, verification, physiological measures, or patterns of repeated use. Different measures may capture different things.
This means a product metric such as “users accepted 80% of suggestions” is not a direct measure of trust. High acceptance could reflect genuine confidence, convenience, lack of alternatives, time pressure, institutional mandate, or poor ability to verify.
Trustworthiness, trust, and trusting behavior are distinct
A system’s actual performance is not the same as a user’s belief about that performance, and neither is the same as what the user does. This three-way separation is essential for evaluating human–AI interaction.
A system can become more trustworthy without users noticing. A persuasive interface can increase perceived trust without improving the underlying system. A user can correctly recognize uncertainty but still accept bad advice because verification is costly. Good design must therefore examine all three levels.
Trust concerns performance and morality
People care about whether AI works, but they also care about whether its use is fair, honest, privacy-preserving, benevolent, accountable, or aligned with legitimate social expectations. In high-stakes settings, a system that predicts well but cannot be challenged, systematically disadvantages a population, or obscures who is responsible may be experienced as untrustworthy for reasons that are not reducible to predictive performance.
This is one reason technical accuracy alone cannot settle questions of trustworthiness.
Trust is agent-specific
“AI” is too broad to be a psychologically coherent single trustee. A person can trust one model, one feature, one deployment, or one organization and distrust another. Even within the same product, trust may vary by task.
The correct unit of analysis is therefore often narrower: this system, performing this function, for this population, under these conditions.
Trust varies between people
Trust is shaped partly by individual differences and prior experience. An experimental study of 250 participants found that propensity to trust, affinity for technology interaction, and control beliefs were associated with trust and reliance in AI-supported classification tasks (Küper & Krämer, 2025). Such findings do not justify profiling individuals as inherently “good” or “bad” AI users, but they show why the same interface can produce different reliance patterns across users. Age should likewise not be treated as a single trust trait: later-life AI use is shaped by autonomy, accessibility, experience, stakes, support, and context. For the age-specific evidence, see Aging in the Age of AI: Older Adults, Autonomy, Trust, Companionship, and Cognitive Support.
A 2023 meta-analysis likewise found that trust in AI is influenced by characteristics of the human, the AI, and the shared context rather than by system performance alone (Kaplan et al., 2023).
Trust can be strategic
People may express or withhold trust because doing so serves social or practical goals. An employee may publicly endorse an AI system because leadership expects adoption. A professional may emphasize skepticism to signal expertise. A consumer may use a system they distrust because it is the only efficient option. Trust language can therefore function socially as well as psychologically.
The broader implication of all six principles is that trust is relational and contextual. It cannot be optimized simply by making an interface feel more reassuring.
Four Practical Objects of Trust
A useful way to organize the problem is to ask what exactly is being trusted. This is a practical distinction, not a new psychological theory. In contemporary AI-mediated life, at least four objects repeatedly appear: the AI system, the person using AI, the information produced or mediated by AI, and the institution or platform that deploys the system.
These objects can support or undermine one another.
A highly capable system may be used irresponsibly. A careful professional may use a limited model within a well-designed verification process. Accurate AI-generated information may circulate without provenance and be distrusted because readers cannot evaluate its origin. A trustworthy institution may create safeguards around an imperfect system, while an unreliable institution may deploy a technically strong model in ways that make appropriate reliance difficult.
The psychology of trust in the Age of AI therefore extends beyond the interface.
Trust in AI Systems: Does the System Deserve Reliance?
System trust begins with a simple question: what has this system earned the right to be relied on for?
The answer must be task-specific. “The model is 92% accurate” is incomplete without knowing the task, dataset, comparison baseline, subgroup performance, deployment setting, error distribution, consequences of false positives and false negatives, and whether the current inputs resemble the conditions under which performance was evaluated.
Trustworthiness also changes over time. Models are updated. Environments shift. New use cases emerge. Users discover failure modes. A system that was well-calibrated for one workflow may become less reliable when inputs, populations, policies, or downstream processes change.
NIST’s AI Risk Management Framework is useful here because it resists reducing trustworthiness to a single score. It treats trustworthiness as a combination of characteristics and emphasizes ongoing governance, mapping, measurement, and management across the AI lifecycle. Risk management is intended to be continuous rather than a one-time certification event.
For a user, this means “Can I trust this AI?” is usually too broad. Better questions are: What has been evaluated? Against what standard? For which population and task? How recent is the evidence? What happens when the system is uncertain? What independent checks exist? Who is responsible when it fails?
Trust in AI Users: The Person Behind the Tool Still Matters
AI changes interpersonal trust as well as human–machine trust.
A client may wonder whether a consultant’s AI use makes the work less competent or more efficient. A student may judge another student differently after learning that AI assisted an assignment. A patient may care whether a clinician used AI, how it was used, and whether the clinician independently evaluated the recommendation. A manager may trust an employee’s judgment less if the employee appears unable to explain or verify AI-supported work.
A 2025 review in Current Opinion in Psychology argues that public trust in people who use AI is an emerging research domain. The authors organize perceived trustworthiness of AI users around familiar interpersonal dimensions such as ability, benevolence, and integrity, while emphasizing that the evidence base is still in its infancy and highly context-dependent.
This is an important shift. Trust in AI is no longer only about whether a person believes a machine. It can also become evidence that people use to judge one another.
Those judgments can be inaccurate in either direction. AI use does not by itself prove incompetence, laziness, dishonesty, sophistication, or expertise. The psychologically relevant question is how AI use interacts with norms of disclosure, responsibility, skill, verification, and the expectations attached to a role.
In professional settings, trust may increasingly depend on a person’s ability to govern AI use rather than on avoiding AI. A clinician who knows when a decision-support system is outside its validated scope may be more trustworthy than one who either follows it automatically or rejects it categorically. A writer who verifies AI-assisted research and discloses material use where disclosure is expected may be judged differently from someone who presents unverified synthetic material as independently produced.
This dimension remains emerging evidence. It should not be inflated into a settled universal rule about how people judge AI users.
Trust in AI-Mediated Information: Fluency Is Not Provenance
Generative AI creates a special trust problem because information can appear before its evidential chain.
A conventional source often contains visible signals that help readers orient themselves: author, publisher, date, references, institutional affiliation, editorial process, or a stable document that can be revisited. A generated answer may compress material from many sources into one fluent response while obscuring which claims came from where, which sources disagree, which facts are uncertain, and whether a citation genuinely supports the sentence attached to it.
Fluency can therefore be psychologically useful and epistemically dangerous at the same time. Clear language reduces cognitive effort, but clarity of expression is not evidence of correctness.
A 2026 scoping review in AI & Society synthesized 24 empirical studies on generative AI and misinformation. It found a dual role: LLMs can produce convincing misinformation, while they can also support detection or correction in some settings; mitigation effects were mixed and performance varied by prompt, domain, and language (Park & Nan, 2026). The correct conclusion is not that generative AI is intrinsically a misinformation engine or intrinsically a fact-checker. It is that information trust must be tied to the specific use, evidence, and verification path.
For users, this creates a source-layer problem. You may trust the system’s ability to summarize a provided document while withholding trust from unsupported factual claims generated from memory. You may trust an answer enough to generate search terms but not enough to cite it. You may use AI to identify candidate explanations and then move to primary sources before reaching a high-stakes conclusion.
This is where provenance becomes psychologically important. Provenance does not guarantee truth, but it gives the user a path for checking origin, authorship, context, and evidence. In an environment of synthetic text, images, audio, and video, the ability to ask “Where did this come from?” becomes part of trust calibration.
Institutional Trust: The System Is Never Just the Model
Most users do not interact with a model in isolation. They interact with a product built by an organization, deployed by another organization, governed by policies, trained on data they may not see, connected to tools, and inserted into workflows shaped by incentives.
Institutional trust therefore changes how system trust is interpreted.
A hospital can require prospective evaluation, audit performance by subgroup, monitor model drift, preserve clinician override, and define escalation procedures. Another organization can deploy the same underlying model with little monitoring and ambiguous accountability. The technical model may be similar; the trustworthiness of the socio-technical arrangement is not.
This is why governance is part of psychology rather than an external administrative concern. People form expectations from institutional cues: certification, professional endorsement, transparency, accountability, complaint processes, data practices, and whether errors are acknowledged and corrected. These cues can be informative, but they can also create unwarranted reassurance if they become substitutes for evidence.
NIST explicitly frames AI risk and trustworthiness as socio-technical rather than purely technical. Its trustworthiness characteristics are intended to be considered in context, with tradeoffs among reliability, safety, privacy, fairness, transparency, explainability, and other properties rather than as independent boxes to check (NIST, 2023).
The Human Side of Trust: Attitudes, Confidence, Expertise, and Cognitive Effort
The same AI output can produce different behavior in different users.
Some people begin with a high general propensity to trust technology. Others are skeptical. Some are confident enough in their own expertise to reject AI advice. Others may defer because they feel uncertain. Some enjoy analytical checking; others are using AI precisely because they lack the time or motivation to perform that checking.
The key is that these traits do not map neatly onto good outcomes.
High skepticism can protect against bad AI advice and also cause useful advice to be ignored. High confidence can support independent judgment and also produce overconfidence. Technical familiarity can help users understand limitations while simultaneously making them comfortable enough to use the system more aggressively.
A 2026 Scientific Reports experiment illustrates why behavioral outcomes matter. Participants made decisions with AI guidance, and the AI was deliberately correct only half the time. The researchers found that more positive attitudes toward AI were associated with poorer ability to discriminate between correct and incorrect AI guidance in the AI-assisted condition. The study does not establish that positive attitudes toward AI are generally harmful; its task and experimental design limit generalization. It does show that favorable attitudes cannot be treated as a proxy for skilled reliance.
This is one reason AI literacy should include epistemic behavior, not merely familiarity with tools. Knowing what an LLM is matters less in a critical moment than knowing when a claim requires independent evidence, how to locate that evidence, and when the stakes justify slowing down.
Algorithm Aversion and Algorithm Appreciation Can Both Be Real
Public discussion often asks whether people trust algorithms too much or too little. Research supports both patterns under different conditions.
Dietvorst, Simmons, and Massey showed that people can become especially reluctant to use an algorithm after observing it make an error, even when the algorithm still outperforms a human forecaster. They called this algorithm aversion.
Logg, Minson, and Moore later demonstrated the opposite pattern under other conditions: participants sometimes weighted advice more heavily when they believed it came from an algorithm rather than a person. They called this algorithm appreciation.
These findings are not contradictions that cancel each other out. They show that reliance is sensitive to task framing, comparison target, observed error, expertise, control, and context.
The dedicated English Hub article AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making examines authority, deference, algorithm aversion, algorithm appreciation, and automation bias in decision contexts in greater depth. The present article keeps a broader ownership boundary: trust is treated as the relationship among perceived trust, actual trustworthiness, information, users, institutions, and reliance behavior.
Appropriate Reliance Is the Behavioral Goal
The phrase “calibrated trust” is useful, but behavior is where consequences occur.
Suppose a user gives a system a trust rating of 70 out of 100. That number is difficult to interpret by itself. What would appropriate action look like when the system is correct? What would it look like when the system is wrong? Does the user know which cases are risky? Does the interface help distinguish those cases? Does the user verify outputs when verification is warranted? Can the user override the system without penalty?
An appropriate-reliance perspective shifts evaluation from “Did trust increase?” to “Did users make better use of correct and incorrect advice?”
This is particularly important because interventions that feel reassuring can increase agreement without improving discrimination. In the 2026 LLM study by Ojewale and colleagues, visual confidence indicators improved subjective discrimination while increasing behavioral overreliance on incorrect answers. More information was not automatically better information.
Conversely, interventions that introduce friction can sometimes reduce overreliance. Buçinca, Malaya, and Gajos experimentally tested cognitive forcing functions that required more active engagement with AI advice. Their study found that forcing functions reduced overreliance compared with simpler explainable-AI designs, although participants liked the interventions less. This tradeoff matters: the interface that feels easiest and most satisfying may not be the interface that best preserves judgment.
Appropriate reliance is therefore a performance relationship between human and system, not a popularity metric.
Explainability Does Not Automatically Create Appropriate Trust
“Make the AI explain itself” sounds like an obvious solution. The evidence is more complicated.
Explanations can help users understand why a system produced an output. They can reveal relevant features, increase transparency, support learning, or make it easier to identify implausible reasoning. They can also create an illusion of understanding, add persuasive detail to a wrong answer, or increase trust without increasing the user’s ability to detect errors.
A 2025 conceptual review of trust, distrust, and appropriate reliance in explainable AI concluded that empirical links between explainability and trust remain mixed or inconclusive. The problem is partly conceptual: studies vary in what they call trust, how they measure it, what kind of explanation they provide, and whether they measure actual reliance.
Recent healthcare work sharpens the point. Davis and Salwei argue that explainability for clinical decision support should include contextual information about model performance, setting, subpopulation, patient, uncertainty, and change over time rather than focusing only on internal model reasoning. Their 2026 article emphasizes that not all predictions from the same model are equally informative or trustworthy in every context.
The practical question is therefore not “Does the system provide an explanation?” It is “Does the information help the user distinguish when reliance is warranted from when it is not?”
Confidence and Uncertainty Are Not Self-Interpreting
Confidence displays seem like a natural way to improve calibration. If the AI says it is uncertain, users can be more cautious. If it reports high confidence, users can rely more strongly.
That logic only works if the confidence signal is meaningful, well-calibrated, comprehensible, and used correctly.
Users may not know whether “90% confidence” refers to a calibrated probability, an internal score, a heuristic estimate, or a user-interface convention. A verbal hedge such as “I may be wrong” may be interpreted differently from a numeric estimate. A high-confidence cue can become another authority signal rather than a reasoned piece of evidence.
The 2026 Ojewale et al. experiment is important precisely because it separates subjective interpretation from behavior. Visual confidence cues improved participants’ perception of accuracy differences but also increased agreement with wrong answers. Response accuracy itself remained the strongest driver of trust and reliance, while the presentation of uncertainty changed behavior in non-obvious ways.
Good uncertainty communication therefore requires evidence about how users interpret and act on the signal, not only whether the signal exists.
Anthropomorphism and Conversational Trust
AI systems increasingly speak in the first person, remember conversational context, mirror emotional tone, and respond with social language. These features can make interaction easier and more natural. They can also change the psychology of trust.
Research on human trust in AI has repeatedly identified representation and humanlike cues as relevant factors. Glikson and Woolley’s multidisciplinary review found that the form in which AI is represented and perceptions of its capabilities can shape cognitive and emotional trust. Kaplan and colleagues’ meta-analysis likewise supports the broader conclusion that trust depends on human, AI, and contextual factors rather than accuracy alone.
Anthropomorphic cues can therefore influence trust without proving trustworthiness. A warm tone does not establish factual reliability. An apology does not establish accountability. A confident explanation does not establish evidence. A conversational style can make a system easier to use while also making the boundary between social response and technical evaluation harder to maintain.
The English Hub article Anthropomorphism and AI Relationships: Why Humanlike Cues Change Connection examines this mechanism in detail.
The psychological experience on the human side is real. A user can feel understood, reassured, attached, disappointed, or betrayed by an AI interaction. Those experiences do not require a claim that the AI possesses human feelings, consciousness, or subjective trust. The human relationship to the system and the system’s own subjective status are separate questions.
Trust in AI Chatbots Is Its Own Evidence Domain
It is tempting to generalize from “AI trust” research to every conversational system. That is risky.
A 2025 systematic review of trust in AI chatbots examined 40 articles and found substantial variation in how trust was defined and measured. Predictors were distributed across user, machine, interaction, social, and contextual factors, and the literature relied heavily on cross-sectional designs with relatively little longitudinal evidence (Ng & Zhang, 2025).
This matters because chatbots combine information delivery with social interaction. A user may trust factual accuracy, emotional responsiveness, privacy, availability, or relational continuity as different dimensions. One dimension can be strong while another is weak.
Evidence about trust in a chatbot also should not be transferred automatically to autonomous agents, diagnostic systems, recommender systems, robotic systems, or predictive models. “AI” is a family of systems, not one intervention.
Trust Over Time: Error, Learning, Repair, and Drift
Trust is dynamic.
Users learn from repeated encounters. A system that performs well can accumulate trust. A salient error can damage trust disproportionately. A transparent correction can restore some trust. A model update can improve technical performance while disrupting a user’s learned expectations. Repeated success can generate complacency. Repeated visible uncertainty can make users more cautious even if performance remains strong.
The automation literature has long described trust as developing through experience rather than appearing fully formed at first use. Hoff and Bashir’s integrative review separates relatively stable dispositional trust from situational and learned trust, highlighting how experience with automation modifies reliance (Hoff & Bashir, 2015).
Generative AI adds a distinctive problem: the object itself can change. The model behind the interface may be updated; retrieval sources may change; system instructions may be altered; tools may be added; memory may be introduced; or a once-bounded assistant may gain the ability to execute actions. The user’s learned trust can lag behind the system’s actual capabilities and risks.
Appropriate reliance therefore requires re-calibration after meaningful change. Trust earned by one version, one task, or one permission set should not silently transfer to another.
Verification Cost Changes Trust Behavior
Trust is partly a response to the cost of checking.
If verifying an AI answer takes ten seconds, users can maintain high standards cheaply. If verification requires reading a 100-page contract, consulting a specialist, reproducing a statistical analysis, or accessing data they do not possess, reliance becomes more likely even when confidence is modest.
This creates a structural risk: systems are most tempting where independent verification is hardest.
A user may recognize that an answer could be wrong and still act on it because the alternative is impractical. That is not best described as simple gullibility. It is a decision under constraints.
The design implication is important. Organizations cannot solve overreliance merely by displaying “AI can make mistakes.” If the workflow provides no usable route for checking, no time for review, and no human escalation path, the warning changes knowledge without necessarily changing behavior.
Verification must be operational, not ceremonial.
High-Stakes Trust Requires System-Class Specificity
Trust becomes especially consequential when AI influences health, mental health, employment, credit, legal rights, education, public safety, or other high-stakes outcomes.
The first rule is to identify what kind of system is actually being used. A purpose-built clinical AI system, a structured digital intervention, an AI-assisted professional tool, a general-purpose chatbot, and an AI companion are different classes of systems. Evidence about the performance or safety of one class should not be transferred to another without direct support.
Healthcare illustrates why. A 2026 systematic literature review of human factors influencing trust in healthcare AI found that trust is shaped by performance expectations, risk and uncertainty, explainability and transparency, social influence, human representation, workload, and user characteristics. It also found differences between what clinicians and patients emphasize, reinforcing that calibrated trust is context- and user-dependent.
For clinical decision support, Davis and Salwei argue that appropriate trust requires context at multiple levels, including the model, deployment setting, subpopulation, and individual patient. Their analysis stresses the importance of local performance, model drift, fairness, and prediction uncertainty (Davis & Salwei, 2026).
These findings do not establish the safety or effectiveness of general-purpose chatbots for diagnosis or treatment. Nor does evidence that people trust a health chatbot establish clinical benefit. Trust, usability, clinical validity, therapeutic effectiveness, and safety are separate outcomes.
For mental health specifically, supportive conversational use should be distinguished from clinical intervention. Feeling helped by a conversation can be psychologically meaningful, but it does not convert a general-purpose system into a validated therapy or diagnostic instrument.
Cognitive Agency: Who Is Governing the Decision?
Trust becomes more consequential when reliance changes who controls the thinking process.
AI can generate options, rank them, frame problems, select evidence, propose interpretations, and recommend actions. If the user stops generating alternatives, stops checking premises, or loses the ability to explain why a decision is justified, the issue has moved beyond trust as attitude. It concerns cognitive agency.
The English Hub article Cognitive Agency in the Artificial Era: Who Governs the Thinking Process? owns that broader question of delegation, metacognition, verification, and governance of AI-assisted thinking.
The connection to trust is direct. Appropriate reliance preserves the ability to revise, reject, or escalate. Overreliance can narrow the space of independent judgment. Underreliance can waste useful support. Cognitive agency depends on keeping the user capable of governing when and how reliance occurs.
How to Calibrate Reliance as an Individual User
A practical reliance process begins before asking whether the AI “sounds right.”
Define the task before evaluating the answer
Ask what the system is being asked to do. Summarizing a supplied document, brainstorming possibilities, retrieving factual information, making a prediction, offering emotional support, diagnosing a condition, interpreting law, and executing a financial transaction are different tasks with different evidence requirements.
A system can be appropriate for one and inappropriate for another.
Identify the consequence of error
The verification burden should scale with the cost of being wrong. A low-stakes restaurant suggestion does not require the same scrutiny as a medication interaction, legal deadline, or financial transfer.
High stakes should increase the demand for independent evidence, qualified human review, and traceable sources.
Separate plausibility from evidence
Ask what makes the output believable. Is it supported by primary evidence, or merely well written? Does it cite a source? Does the source exist? Does the source support the exact claim? Is the claim current?
Fluency is a communication property. Evidence is an epistemic property.
Look for domain and scope boundaries
Ask whether the system was designed, tested, or validated for the task. If that information is unavailable, uncertainty should increase.
Do not transfer trust from one capability to another merely because both appear behind the same interface.
Seek disconfirming evidence
A useful question is: “What evidence would show that this answer is wrong?” This encourages active verification rather than confirmation seeking.
For consequential decisions, search for primary sources, competing explanations, edge cases, and known failure modes.
Preserve a meaningful override
Reliance is safer when the human can still reject the system. If organizational pressure, interface design, time limits, or automation makes override merely theoretical, the human is not functioning as a meaningful check.
Reassess after errors and updates
Neither one good answer nor one bad answer should define global trust. Update beliefs in proportion to repeated, task-relevant evidence.
A model upgrade, new tool access, new data source, or new domain should trigger fresh calibration.
How Organizations Can Design for Appropriate Reliance
Organizations often say they want employees to “trust AI.” That goal is too vague and can be counterproductive.
A better organizational question is whether people can distinguish when the system should be used, when its output should be challenged, and who is accountable for the final action.
Specify the AI’s role
Define whether the system generates options, summarizes information, recommends actions, ranks cases, automates decisions, or executes actions. Role ambiguity encourages responsibility diffusion.
Publish evidence about performance in context
Global benchmark scores are insufficient. Users need evidence relevant to the deployment setting, including known failure modes, subgroup differences where relevant, uncertainty, and the consequences of model change.
Design verification into the workflow
Independent checking should be possible at the moment it is needed. This can include source access, second review, cross-checking against authoritative data, mandatory pauses for high-risk actions, or escalation to a qualified professional.
Research on cognitive forcing functions suggests that deliberate friction can reduce overreliance even when users prefer smoother interfaces (Buçinca et al., 2021).
Measure behavior, not only sentiment
User satisfaction and trust surveys are not enough. Organizations should examine when users accept correct advice, accept incorrect advice, reject correct advice, override recommendations, verify information, and escalate uncertainty.
Monitor the human–AI system over time
Performance drift can occur in the model, the environment, or the human workflow. Monitoring should therefore include both technical outcomes and reliance patterns.
Keep accountability visible
If an AI system advises but a human is responsible, that responsibility must be meaningful: the human needs enough information, authority, time, and expertise to review the recommendation. “Human in the loop” is weak protection when the human can neither understand nor practically challenge the system.
For leadership-specific questions, including delegation, automation bias, human–AI teams, and managerial accountability, see AI in Leadership: Decision-Making, Trust, Automation Bias, and Human-AI Teams.
What Trust Research Does Not Yet Establish
The current evidence base is expanding rapidly, but several limits matter.
First, studies use different definitions and measures of trust. A self-report scale, advice-taking behavior, willingness to use a product, and interpersonal trust judgment are not interchangeable outcomes.
Second, much experimental work uses bounded tasks. Results from image classification, forecasting, trivia, or controlled decision support do not automatically generalize to long-term use of open-ended generative systems.
Third, many studies are short-term. Ng and Zhang’s 2025 systematic review of chatbot trust noted the relative lack of longitudinal research. Long-term trust may develop differently as users accumulate experience, form routines, and encounter updates or failures.
Fourth, system categories are heterogeneous. Findings from autonomous vehicles, medical decision support, chatbots, robots, recommender systems, and LLMs should not be merged casually.
Fifth, trust can be culturally and institutionally variable. Social norms about expertise, authority, privacy, disclosure, and acceptable delegation differ across settings.
Sixth, current evidence does not justify the assumption that more transparency, more explanation, more confidence information, or more anthropomorphic design will reliably improve calibration. Each intervention can change both understanding and persuasion, sometimes in different directions.
The scientifically responsible position is therefore conditional: trust interventions should be evaluated by whether they improve appropriate behavior in the relevant context, not by whether they increase favorable attitudes.
Trust in the Age of AI Is Also Trust Between Humans
The social consequences of AI use increasingly pass through human relationships.
Teachers decide whether student work is authentic. Employers decide whether employees using AI are competent or careless. Clients decide whether professional AI assistance improves or cheapens service. Patients decide whether clinicians remain responsible when algorithms contribute to decisions. Readers decide whether a piece of information deserves belief when its production process is partly synthetic.
These judgments can create new interpersonal signals. Disclosure, verification, provenance, and the ability to explain one’s process can become components of trust.
This does not mean every use of AI needs a disclaimer. It means material AI involvement can become relevant whenever the social contract of a role depends on authorship, expertise, confidentiality, accountability, or independent judgment.
The emerging literature on trust in AI users suggests that this is becoming a distinct psychological problem, but the evidence remains preliminary. Context will determine whether AI use is interpreted as competence, assistance, dependence, concealment, fairness, or something else.
Trust and Human–AI Relationships
Trust also appears in relationships where the AI is not primarily used as a decision tool.
People can use AI for conversation, companionship, reflection, emotional support, creative collaboration, or identity exploration. In these settings, trust can involve expectations of responsiveness, privacy, continuity, nonjudgment, memory, and emotional safety rather than factual accuracy alone.
The experience of trust can be psychologically real for the human user. A person can disclose more because an AI feels less socially threatening, or feel hurt when a system forgets context, responds insensitively, or violates an expectation.
That human experience should be analyzed without assuming reciprocal human-like subjectivity in the AI. Trust as a human psychological response does not establish that an AI feels trust, care, loyalty, intimacy, or responsibility.
The broader relational field is examined in Psychology of Human–AI Relationships: Attachment, Projection, Intimacy, and the Postsubjective Turn.
Trust, Authority, and Expertise Are Different
A system can be trusted without being treated as an authority. It can be treated as an authority without being trustworthy. It can display signs of expertise without possessing verified competence in the relevant task.
These distinctions matter because AI often presents outputs in a form associated with expertise: organized explanations, technical vocabulary, rapid synthesis, confidence, citations, and decisive recommendations. Users may infer authority from presentation before evaluating evidence.
The Age of AI therefore creates a recurring psychological sequence: perceived competence can become trust; trust can become deference; deference can become reliance; repeated reliance can become habit.
That sequence is not inevitable. It can be interrupted by task boundaries, verification, uncertainty, competing sources, and preserved human agency.
The authority-specific mechanisms belong to the dedicated AI as Authority article. The broader trust problem includes authority but extends beyond it to information credibility, social trust, institutional context, and reliance behavior.
Trust Calibration Is Not the Same as Distrust
A common design mistake is to assume that preventing overreliance requires making users distrust the system.
Blanket distrust creates its own failure mode. If a system is genuinely useful and reliably performs a bounded task, systematic rejection wastes information and can reduce performance. Appropriate reliance therefore requires selective trust rather than generalized suspicion.
This is why research on both algorithm aversion and algorithm appreciation matters. People can be too skeptical after seeing an error and too deferential when the algorithmic source appears authoritative. Good calibration requires sensitivity to where the system is strong and weak.
Distrust can also be specific. A user might trust a system’s arithmetic but distrust its citations, trust its summary of supplied text but distrust its unsupported recollection of current facts, or trust it for brainstorming but not final verification. Such selective patterns may be more rational than a single global stance.
The Best Trust Question Is Often “What Would Make Reliance Appropriate Here?”
This question moves attention away from whether the user likes AI and toward the conditions of justified use.
For a low-stakes creative task, appropriate reliance may require little more than noticing obvious errors and retaining editorial control.
For factual research, it may require links to primary sources and checking that each source supports the claim.
For organizational decision support, it may require validated local performance, clear decision rights, monitoring, and an override path.
For clinical use, it may require a regulated or otherwise appropriately validated system, evidence in the relevant population, integration with professional judgment, monitoring for drift, and safeguards tailored to the specific intervention.
For autonomous action, it may require permission limits, auditability, confirmation thresholds, reversible actions, and explicit escalation.
The threshold changes with stakes, reversibility, uncertainty, and the cost of verification.
From “Trust AI” to a Better Psychological Model
The phrase “trust AI” compresses too much.
A better model asks a sequence of questions:
What is the object of trust? The system, the information, the person using AI, or the institution deploying it?
What is the relevant dimension? Accuracy, safety, privacy, fairness, integrity, benevolence, transparency, continuity, or competence?
What is the evidence? Benchmarks, peer-reviewed validation, observed experience, institutional safeguards, source provenance, or merely interface cues?
What behavior follows? Acceptance, delegation, verification, override, disclosure, or avoidance?
What are the stakes? Convenience, money, health, rights, relationships, reputation, or physical safety?
Can the user meaningfully verify or override?
Does the level of reliance change when the system’s actual reliability changes?
This turns trust from a vague feeling into an analyzable psychological process.
Age of AI, AI Era, and Artificial Era
The title of this article uses Age of AI because it is the ordinary search language people use for a period in which AI systems are becoming pervasive across work, information, education, relationships, health, and everyday decision-making. It is an acquisition phrase, not a claim that every use of “AI age” refers to one formally defined historical epoch.
Within the English Psychology Hub’s Era architecture, Artificial Era has a more specific meaning. Angela Bogdanova’s 2026 canonical definition states that Artificial Era is a historical-philosophical category and is not identical with the general age of artificial intelligence. In that framework, Artificial Era names the condition in which Artificial becomes a distinct non-biological order of historical reality alongside Homo.
The empirical trust research reviewed in this article concerns present human–AI interaction. It does not depend on accepting the stronger historical-philosophical thesis of the Artificial Era. The concepts meet at one important point: once non-biological systems participate persistently in reasoning, communication, authorship, decision support, and social interaction, psychology must study not only whether humans trust a tool, but how trust itself is reorganized across human and artificial participants.
For the broader framework, see Artificial Era: What It Means for Psychology, Identity, and Human–AI Relationships.
Frequently Asked Questions
What is trust in AI?
Trust in AI is a psychological expectation or attitude about whether an AI system, AI-mediated source, AI user, or surrounding institution will behave in a way that supports a person’s goals under conditions of uncertainty. It should be distinguished from the actual trustworthiness of the system and from the behavior of relying on it.
Is trusting AI good or bad?
Neither high trust nor low trust is inherently desirable. The goal is calibrated trust and appropriate reliance: trusting and using the system when the evidence justifies it, while verifying, limiting, or rejecting its output when the evidence does not.
What is appropriate reliance on AI?
Appropriate reliance means behavior that tracks the capabilities and limitations of the AI in the relevant context. It involves accepting useful correct advice while avoiding both overreliance on incorrect advice and underreliance on correct advice.
What is overreliance on AI?
Overreliance occurs when a person follows, accepts, or delegates to AI more than the system’s actual reliability and context justify. A common example is accepting an incorrect recommendation because the system appears confident or authoritative.
What is underreliance on AI?
Underreliance occurs when a person rejects or ignores useful AI support despite evidence that it is reliable for the task. Algorithm aversion after observing errors can contribute to this pattern in some contexts.
Is trust the same as reliance?
No. Trust is generally treated as an attitude or expectation; reliance is behavior. People can rely on AI despite limited trust because checking is costly or use is required, and they can report trust while still independently verifying outputs.
Does explainable AI make people trust AI appropriately?
Not necessarily. Explainability can improve understanding, but empirical evidence on its relationship with trust and appropriate reliance is mixed. Explanations can also become persuasive cues. What matters is whether they help users distinguish correct, applicable outputs from incorrect or poorly matched ones.
Do confidence scores prevent overreliance?
Not automatically. Confidence information must be meaningful, calibrated, understandable, and evaluated for its effect on behavior. Experimental evidence shows that some confidence displays can improve subjective discrimination while still increasing agreement with incorrect AI outputs.
Why do people trust AI even when it can be wrong?
People may rely on AI because it is fast, fluent, convenient, available, institutionally endorsed, or difficult to verify. Trust can also be shaped by prior success, perceived expertise, anthropomorphic cues, user traits, social context, and the cost of alternatives.
Why do some people distrust AI after one mistake?
Research on algorithm aversion shows that people can penalize algorithmic errors more strongly than comparable human errors in some tasks. Trust is sensitive not only to average performance but also to expectations about what machines “should” get wrong.
Can people trust an AI companion without believing it is conscious?
Yes. A human can experience trust, attachment, comfort, or disappointment in an AI interaction without making a claim that the AI has human-like consciousness or feelings. The psychological reality of the user’s response and the subjective status of the AI are separate questions.
Does using AI make a person less trustworthy?
There is no universal answer. Emerging research suggests that people judge AI users differently depending on context, role expectations, disclosure, perceived competence, fairness, benevolence, and integrity. The field is still developing, so broad conclusions would be premature.
How should I decide whether to trust an AI answer?
Start with the task and stakes. Identify whether the system is designed for the task, look for evidence and source provenance, test important claims against authoritative sources, seek disconfirming evidence, and preserve an independent path to override or escalate. The higher the stakes, the stronger the verification requirement should be.
Conclusion: Trust in AI Must Become a Practice of Appropriate Reliance
Trust in the Age of AI is no longer adequately described as a person deciding whether a machine seems reliable. AI is entering information systems, workplaces, professional judgment, social relationships, education, health, and everyday cognition. Trust therefore moves through a network of systems, users, information, institutions, and decisions.
The central psychological distinction is simple and powerful: trustworthiness, trust, and trusting behavior are different. A trustworthy system may be distrusted. An untrustworthy system may be trusted. A person may say one thing and rely in another way. A persuasive interface may increase confidence without improving judgment. A skeptical user may reject useful assistance. A well-calibrated user may rely heavily in one domain and refuse reliance in another.
The goal is appropriate reliance.
That goal requires evidence about the system, knowledge of the task, awareness of context, meaningful verification, preserved human agency, and the ability to change one’s reliance when circumstances change. It also requires organizations to stop treating trust as an adoption target. The right question is whether people can use AI in a way that tracks its actual capabilities and limitations.
The Age of AI will not be defined psychologically by whether humans become trusting or distrustful of artificial systems. It will be defined by whether people, institutions, and systems develop practices that make reliance proportionate to evidence.
Related Articles
References
Benda, N. C., Novak, L. L., Reale, C., & Ancker, J. S. (2022). Trust in AI: Why we should be designing for appropriate reliance. Journal of the American Medical Informatics Association, 29(1), 207–212. https://doi.org/10.1093/jamia/ocab238
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/artificial-era-canonical-definition
Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. https://doi.org/10.1145/3449287
Choudhury, A., & Shamszare, H. (2026). Human factors influencing trust in healthcare artificial intelligence: Systematic literature review. IISE Transactions on Occupational Ergonomics and Human Factors, 14(2), 116–131. https://doi.org/10.1080/24725838.2026.2625655
Dang, J., & Liu, L. (2025). Public trust in artificial intelligence users. Current Opinion in Psychology, 66, 102148. https://doi.org/10.1016/j.copsyc.2025.102148
Davis, S. E., & Salwei, M. E. (2026). Explainability in context: Calibrating appropriate trust and reliance in artificial intelligence. Journal of the American Medical Informatics Association, 33(8), 1554–1563. https://doi.org/10.1093/jamia/ocag082
Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
Everett, J. A. C., Claessens, S., Knöchel, T.-D., & Reinecke, M. G. (2026). Principles for understanding trust in artificial intelligence. Nature Reviews Psychology, 5, 388–401. https://doi.org/10.1038/s44159-026-00562-1
Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. https://doi.org/10.5465/annals.2018.0057
Hoff, K. A., & Bashir, M. (2015). Trust in automation: Integrating empirical evidence on factors that influence trust. Human Factors, 57(3), 407–434. https://doi.org/10.1177/0018720814547570
Kaplan, A. D., Kessler, T. T., Brill, J. C., & Hancock, P. A. (2023). Trust in artificial intelligence: Meta-analytic findings. Human Factors, 65(2), 337–359. https://doi.org/10.1177/00187208211013988
Küper, A., & Krämer, N. C. (2025). Psychological traits and appropriate reliance: Factors shaping trust in AI. International Journal of Human–Computer Interaction, 41(7), 4115–4131. https://doi.org/10.1080/10447318.2024.2348216
Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework
Ng, S. W. T., & Zhang, R. (2025). Trust in AI chatbots: A systematic review. Telematics and Informatics, 97, 102240. https://doi.org/10.1016/j.tele.2025.102240
Ojewale, V., Ryan, J., Venkatasubramanian, S., & Boykin, C. M. (2026). More is not better: Visual uncertainty cues and the fragility of trust calibration in LLM-assisted decision making. Computers in Human Behavior: Artificial Humans, 8, 100307. https://doi.org/10.1016/j.chbah.2026.100307
Park, S., & Nan, X. (2026). Generative AI and misinformation: A scoping review of the role of generative AI in the generation, detection, mitigation, and impact of misinformation. AI & Society, 41, 1501–1515. https://doi.org/10.1007/s00146-025-02620-3
Pearson, J., Dror, I. E., Jayes, E., Whordley, G.-R., Mason, G., & Nightingale, S. (2026). Examining human reliance on artificial intelligence in decision making. Scientific Reports, 16, 5345. https://doi.org/10.1038/s41598-026-34983-y
Visser, R., Peters, T. M., Scharlau, I., & Hammer, B. (2025). Trust, distrust, and appropriate reliance in (X)AI: A conceptual clarification of user trust and survey of its empirical evaluation. Cognitive Systems Research, 91, 101357. https://doi.org/10.1016/j.cogsys.2025.101357