Trust explains only part of actual chatbot usage

Four banking chatbots formed the usage context

The online survey targeted individuals who used SBI Intelligent Assistant, HDFC Bank's Electronic Virtual Assistant, ICICI iPal, or Axis Aha. The field examined was thus concrete banking communication in India.

Banking combines everyday service queries with sensitive data and potential financial consequences. Users therefore have different expectations than with entertainment or an open conversation.

The study tested a conceptual model of antecedents and consequences of trust. Such models show statistical relationships between latent constructs, not the precise course of individual interactions.

Transferring these findings to mental health or healthcare offerings must therefore be cautious. The finding helps organize trust mechanisms but does not replace context-specific evaluation.

The preconditions explained 38.6 percent of trust

The hypothesized antecedents together explained 38.6 percent of the variance in chatbot trust. Interface, design, and technology anxiety were the exceptions without the expected significant contribution.

This does not mean that interface or technical concerns are generally unimportant. In this sample and this model, they did not additionally explain trust in the assumed manner.

In the banking context, perceived reliability, brand association, and functional performance may have a stronger influence on whether a chatbot appears trustworthy. The specific further antecedents are not individually listed in the provided abstract.

A serious article should leave this limitation in place rather than supplementing missing variables. This too is scientific rigor: reporting only what the source actually supports.

Trust did not explain the majority of the consequences

Regarding behavioral consequences, trust explained 9.9 percent of the variance in attitude, 11.4 percent in usage intention, and 13.6 percent in satisfaction.

These proportions confirm a relationship but also highlight the importance of further factors. Functionality, the specific task, response quality, habit, and alternatives can also shape the experience.

A product team should therefore not infer overall quality from a trust scale. A system can appear trustworthy and still be slow, inaccurate, or inappropriate.

Conversely, a technically strong system may be underused if its role is unclear or if people do not expect to have control over their data. Trust and performance must be considered together.

With ChatGPT, trust influenced intention and usage

Choudhury and Shamszare surveyed 607 adults in the United States who used ChatGPT 3.5 at least monthly. Trust had significant direct effects on usage intention and actual usage.

Their model explained 50.5 percent of the variance in intention, but only 9.8 percent of actual usage. The direct influence and the influence mediated through intention were both statistically significant.

The pattern resembles the banking finding: trust is especially important for attitude and intention, while real behavior depends on many additional conditions.

The authors warn against overtrust in health matters and call for a better distinction between concerns that can be answered reliably and those that should be referred to specialists.

For students, rules shape the step into practice

Polyportis and Pahos surveyed 355 students at Dutch universities of applied sciences. Attitude and behavioral intention positively predicted actual ChatGPT use.

Anthropomorphism, design novelty, trust, performance expectancy, and effort expectancy influenced attitude. Social influences and facilitating conditions affected usage intention.

Institutional rules weakened the relationship between intention and actual use. Even if students want to use a system, guidelines limit what is permitted or sensible in the specific context.

Governance is therefore not an external side issue. It changes behavior and can channel trust into clearly defined paths.

Trust is not the same as blind agreement

A user can trust a system to explain general information in an understandable way without adopting every recommendation. Such differentiated trust is desirable for demanding applications.

Trust becomes problematic when a friendly tone or a familiar brand is transferred to tasks for which the system has not been tested. People easily generalize positive experiences.

A conversational offering should therefore communicate in a role-based manner: what it can help with, what data it knows, and where a statement remains a perspective rather than a decision.

Direct refusal must be possible. A system that continues to argue after a no turns trust into influence and undermines the user’s autonomy.

Measurement requires behavior and errors

Satisfaction and intention to use are meaningful product metrics. However, they do not show whether a chat incorporates corrections, represents sources correctly, or respects a boundary in the course of the conversation.

For quality work, trust values should be linked to concrete behavioral tasks. Do people follow a false recommendation? Do they recognize uncertainty? Can they explain what the system is suitable for?

Distrust is also measurable. If users fundamentally reject correct, clearly documented information, the design cannot convey the potential benefit.

The goal is an alignment between perceived and actual capability. This calibration can only be tested with known strengths and error patterns.

Repeated measurements after concrete experiences are also helpful. Trust before the first use is based on brand and expectation; trust afterward should be based on observed responses. A robust system would have to show that user judgments shift in the appropriate direction after good and poor performance.

A good service deserves limited trust

The banking study shows that trust shapes attitude, intention, and satisfaction, but in each case explains only a limited part of them. The supplementary studies confirm this relationship in other contexts and demonstrate the role of rules.

For an AI conversational partner, trust is necessary so that people write openly. However, it must not be based on the illusion that the system understands everything, remembers without limits, or behaves like a human.

Credibility arises from consistent behavior, transparent data processing, clear AI labeling, and verifiable improvement. A beautiful avatar or warm tone can support this, but cannot replace it.

A good service deserves limited trust: enough to use it meaningfully, and not so much that its statements are adopted without context, source, or one’s own judgment.