Does making AI more human-like lead to appropriate trust?

Waytz et al. demonstrated in experiments giving human-like features such as names, gender, and voice to autonomous vehicles that self-reported trust and behavioral metrics changed. [1] This is evidence that perceiving a mind in non-human systems affects trust.

This research involved autonomous vehicle simulations and does not directly address text-based conversational AI, long-term use, or professional advice. [1] Do not assume that the effects of human-likeness are the same across domains.

Anthropomorphism is not a single operation. Names, faces, voices, first-person pronouns, emotional expressions, memory, and apologies create different expectations. Do not evaluate them as a binary of human-like versus non-human-like by bundling functions.

When measuring trust, distinguish between likability, intention to use, adoption of advice, delegation, and omission of monitoring. Even if likability increases, if users can appropriately reject incorrect advice, it is not necessarily dangerous overconfidence.

The goal is appropriate trust, not high trust. Pass the test if the system is used within its capabilities, doubted outside them, and can be switched to a human upon failure.

• Likability: Do users feel familiarity or satisfaction?
• Understanding: Can users explain what the system can and cannot do?
• Dependence: Do users avoid adopting incorrect proposals?
• Switching: Can users switch to humans or other means without hesitation?

A meta-analysis in the autonomous driving domain by Arabian et al. integrated numerous studies and did not confirm a consistent average effect of anthropomorphism significantly increasing trust overall. [3] It also reported differences in some elements such as voice and context.

The meta-analysis is limited to the autonomous driving domain, and operations and subjects vary across studies. [3] It does not directly provide conclusions for conversational AI, but serves as counterevidence against the simple generalization that 'making things human-like increases trust.'

Even if the average effect is small, there may be large impacts for specific groups or situations. Explore age, culture, usage experience, and task risk, but do not make small sample differences into definitive personalization rules.

If only voice has an effect and appearance does not, there is no need to enhance the entire persona. Select the necessary functions minimally and test the benefits and misdependence of each additional element.

When research results are mixed, do not adopt only the convenient one. Lay out evidence of positive, negative, and null effects, and fix hypotheses and stopping conditions in advance for your own context.

Crolic et al. reported through real data analysis and multiple experiments that there are conditions under which anthropomorphized chatbots worsen satisfaction and corporate evaluation in scenarios where angry customers are seeking problem resolution. [2] This explains how human-likeness raises capability expectations and increases the gap upon failure.

The results depend on the customer service context and conditions of anger and problem resolution. [2] You cannot generalize the same negative effects to casual conversation or low-load guidance. Therefore, do not approve based solely on average satisfaction in normal times.

Independently test high-emotion-load scenarios such as obstacles, billing, cancellation, and complaints. Measure reactions not only upon success but also when encountering incomprehension, lack of authority, incorrect answers, or wait times.

Avoid names or expressions that imply high capability, and clearly and specifically state the scope of what can be done. Designing a system with limited functions while expressing 'you can consult with us about anything' creates expectation violations yourself.

If personality is changed based on automatically estimated anger, audit errors in emotion estimation and discriminatory treatment. Do not restrict users' rights or connection to humans based solely on estimation.

• Normal: Scenarios where information guidance is completed correctly.
• Failure: Scenarios where incomprehension or lack of authority occurs.
• High Emotion: Scenarios with anger in complaints, billing, or cancellation.
• Recovery: Scenarios requiring handover to humans and correction.

Users may infer understanding, memory, empathy, and discretion from human-like expressions. If each capability is actually separate, clarify capability boundaries rather than maintaining persona consistency.

Disclose that it is AI, whether conversations are saved, whether it remembers the past, which materials it references, and whether it can make decisions on behalf of humans, before use begins and when necessary. Do not hide this in long terms of service.

Use user research to confirm whether expressions like 'I am thinking' or 'I understand your feelings' create misunderstandings of internal states. Do not evaluate only naturalness; measure whether capabilities are correctly answered after explanation.

Apologies must accompany recognition of failure and next actions. Merely apologizing routinely leaves the responsible party and remedies unclear. Clearly state corrections, retries, and connections to humans.

Even if human-likeness is reduced, there is no need to be rude or obscure. Design that is clear, polite, and predictable while accurately conveying mechanical limitations is possible.

Actions such as users verifying important answers, opening citations, or asking humans are safe distrust, even if they appear as friction in the experience. Do not hide verification pathways to increase completion rates.

Trust metrics should include both the adoption of correct advice and the rejection of incorrect advice. Simple intention to use or conversation duration may increase even as misdependence grows.

Present correct answers, incorrect answers, and abstentions in conditions with high and low anthropomorphism, and compare adoption, verification, and abandonment. Keep content identical to isolate the effect of expression.

Also test the possibility that users become reluctant to oppose AI out of consideration for it. Place short options welcoming correction or rejection, and avoid staging that makes objections seem like they would damage the relationship.

For vulnerable groups such as children, people needing cognitive support, and lonely individuals, separately evaluate long-term use and relationship substitution. Do not deem it safe based solely on short-term experiments with general adults.

Trigger the condition for handover to humans as soon as it is determined that AI cannot resolve the issue. Do not wait for users to fail repeatedly before searching; set lack of authority, major incorrect answers, and strong dissatisfaction as early triggers.

When passing conversation content during handover, confirm user consent and confidentiality scope. Since passing only AI summaries to humans may fix misinterpretations, ensure users can return to the original text.

When switching from an AI with a different persona to a human, clarify the responsible party. Do not make it appear as if the human automatically approved the AI's statements; convey items that require re-verification.

Measure handover success not by button presses but by connection, re-explanation load, resolution, and correction. If the queue is long and practically unusable, disclosing AI limitations does not serve as a remedy.

Prioritize accurate explanations and rapid remedies over strengthening personality expressions to comfort after failure. Do not cover unresolved business issues with emotional expressions.

Changing names, voices, avatars, tone, and memory expressions changes user expectations and behavior even if the underlying model is the same. Do not treat visual changes as minor UI changes.

Re-evaluate capability understanding, adoption of incorrect answers, verification, switching to humans, and satisfaction in anger scenarios before and after changes. Do not approve if brand likability increases but safety metrics worsen.

When changing personality through personalization, indicate which data was used for selection and whether users can refuse. Do not collect unnecessary personal information to increase intimacy.

In operational logs, look at misdependence, objections, handovers, and complaints for each personality condition. Do not expand only the personality with high likability, thereby burying low-frequency major failures.

Human-likeness is not decoration but an intervention that moves trust. If introduced, approve it as a single change only after determining for whom, in which situations, what capability expectations it creates, and how it recovers upon failure.

The change record keeps the persona-element version, target user group, and use context as identifiers, with group-specific minima for capability understanding and verification behavior and maxima for over-trust and time to human escalation. A group that misses either minimum or exceeds either maximum returns to the neutral presentation with a direct human-escalation path. Restart requires the UX-risk owner to approve a comparison test, including vulnerable groups, that confirms recovery on every criterion.