Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Human-like finance chatbot made advice feel less risky to novices, study of 1,256 finds

Study title: Anthropomorphism and risk perception: The role of trust and domain knowledge in decision-support AI

Evidence type
Controlled or validated comparison
Sample
n = 1,256
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2503cite or check it

Study title: Anthropomorphism and risk perception: The role of trust and domain knowledge in decision-support AI

Researchers at the Chinese University of Hong Kong, Shenzhen and the Hong Kong University of Science and Technology (Guangzhou) ran an online experiment with 1,256 analysed participants who sought investment advice from a simulated AI financial-advisor chatbot. The bot varied its avatar, name and communication style across eight conditions. People who perceived the bot as more human-like reported more cognitive trust and more affective trust, and cognitive trust was linked to lower perceived risk of following the advice.

Read the study abstract

Anthropomorphic design is increasingly used to make conversational AI systems more approachable and socially engaging, yet its effects on users’ risk judgments remain unclear. Drawing on theories of anthropomorphism and cognitive–affective trust, we develop a moderated mediation model explaining how perceived anthropomorphism relates to risk perception in AI decision support. We conducted a large-scale between-subjects online experiment (N = 1,256) using a simulated financial advisor chatbot with a 2 2 2 anthrop…

The problem

In an online experiment with 1,256 people, the more human-like a simulated financial chatbot seemed, the more participants trusted it and, through cognitive trust, the less risky they judged its investment advice, an effect strongest for those with low financial knowledge.

In brief

The pattern matters because the advice in question involves real money. The authors argue that users may feel safer with an AI that can still produce severe factual errors, and that this mismatch between perceived safety and actual risk echoes earlier findings on overtrust in AI-generated content. They conclude that human-like cues should be calibrated rather than simply maximised in high-stakes decision support.

Domain knowledge changed the picture. Participants with lower self-reported financial knowledge showed a stronger risk-reducing path through cognitive trust, while knowledgeable participants were less swayed by human-like cues. The measures were questionnaire ratings of perceived risk, not actual investment outcomes, and the authors say the results suggest but do not prove that trust reduces risk perception.

Main points

  1. Participants were recruited through Prolific; 1,430 completed the study and 1,256 passed attention checks (50.4% female, mean age 38.1, more than 90% from the US).
  2. Each person was randomly assigned to one of eight bot designs: human photo or computer image, the name 'David' or 'bot#156', and human-style or machine-style communication.
  3. Communication style was the main driver of perceived human-likeness; the most human-like design scored higher than a mixed design (Cohen's d = 0.405) and the fully machine-like one (d = 0.350), while a human name alone had little effect.
  4. Perceived human-likeness predicted both trust types, and cognitive trust reliably predicted lower perceived risk (b = -0.230); affective trust showed a small positive link to risk (b = 0.092).
  5. The design cues did not shift trust directly, and the direct effect of perceived human-likeness on risk became non-significant (b = 0.025) once trust measures were included.
  6. The risk-reducing route through cognitive trust was stronger at low domain knowledge and weakened, but stayed significant, at high knowledge.
  7. The authors suggest pairing recommendations with calibration, provenance and plain-language rationale, and tailoring tone to user expertise.

The full story

Participants were told to ask the advisor how to invest their savings, with a suggestion window offering step-by-step prompts, and the chatbot then produced a concrete investment recommendation. Immediately afterwards they rated the bot's human-likeness with the Godspeed questionnaire, rated cognitive and affective trust, reported their knowledge of financial markets, and rated on a 7-point scale how risky it would be to follow the advice. The session took 43 minutes on average.

The manipulation check showed the eight designs differed in perceived human-likeness (F(7,1248) = 2.8912, p = 0.005). Design conditions on their own produced no significant differences in either trust type, so the analysis used participants' perceived human-likeness as the predictor, tested with regression and structural equation models.

Perceived human-likeness strongly predicted affective trust (standardized b = 0.596) and cognitive trust (b = 0.364). Domain knowledge raised both trusts but dampened the effect of human-likeness on each. At low domain knowledge the direct slope from human-likeness to risk was 0.042 and not significant; at high knowledge it was 0.161.

The authors say anthropomorphism is not just an engagement feature and that systems in high-risk tasks should use it responsibly so users do not accept hallucinated or incorrect output without scrutiny. They propose competence-first framing, warmth only alongside evidence, and a short onboarding self-assessment so knowledgeable users see more formal, evidence-heavy messages.

Sources

1 on file
  1. International Journal of Human-Computer Studiesdoi:10.1016/j.ijhcs.2026.103956
    Peer-reviewed2026-10-28

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

Human-like finance chatbot made advice feel less risky to novices, study of 1,256 findshttps://truvace.com/item/anthropomorphism-and-risk-perception-the-role-of-trust-and-domain-knowledge-in-d

  1. Who carries this cost, and what would have to change to prevent it?

    The problem In an online experiment with 1,256 people, the more human-like a simulated financial chatbot seemed, the more participants trusted it and, through cognitive trust, the less risky they judged its investment advice, an effect strongest for those with low financial knowledge.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a controlled comparison with a sample of 1,256, from International Journal of Human-Computer Studies.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The authors note they inferred mental processes from psychometric questionnaires, and say the results suggest but do not prove that trust reduces risk perception; domain knowledge was self-reported and the advisor was a simulated chatbot.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Business