aceModerate evidence
More
Decision outcomes peaked at moderate chatbot humanlikeness in experiments with 832 people
Study title: AI Chatbot Anthropomorphism and Consumer Decision-Making: A Dual-Pathway Calibration Model
- Evidence type
- Controlled or validated comparison
- Sample
- n = 832
- Sources
- 1 sourcePeer-reviewed
- Published
- 3 min read
- Truvace record
- TRV-2026-2674cite or check it
Study title: AI Chatbot Anthropomorphism and Consumer Decision-Making: A Dual-Pathway Calibration Model
Researchers ran three online experiments with 832 adults in mainland China to test how humanlike design in AI chatbot advisers shapes consumer decisions. Participants chatted with a simulated adviser about an investment of 30,000 CNY, a health product or a smartphone, and their choices were scored against expert benchmarks.
Read the study abstract
Anthropomorphism has shown inconsistent effects on human judgment, yet the underlying cognitive mechanisms remain underspecified. This research develops and tests a Dual-Pathway Calibration Model (DPCM) integrating dual-process theory, construal level theory, metacognition theory, and regulatory focus theory to explain how anthropomorphic cues influence cognitive processing in human–artificial intelligence (AI) interaction. Three experiments (N = 832) manipulated anthropomorphism and cue inconsistency, examining…
The problem
In three online experiments with 832 adults in China, decision outcomes with an AI chatbot adviser followed an inverted U (quadratic b = −0.05, p = 0.001), so maximum humanlike design did not give the best results.
In brief
Across the studies, decision outcomes rose and then fell as the chatbot became more humanlike (quadratic b = −0.05, p = 0.001). Humanlike cues built closeness and affective trust (indirect effect = 0.18), but when cues clashed with the chatbot's actual responses, participants reported more uncertainty and scrutinised the advice more.
Whether relying on the chatbot helped depended on how well participants could tell good advice from bad: metacognitive calibration moderated the link between AI reliance and decision quality (β = 0.17, p = 0.007). Regulatory focus also shaped which route dominated. The authors describe the inverted U as an exploratory prediction, so the finding is early evidence from simulated advisory tasks.
Main points
- The three experiments had 286, 234 and 312 participants after exclusions, all recruited on the Credamo platform from mainland China, aged 18 to 55 with prior AI chatbot experience; data were collected between March and June 2025.
- Study 1 used a 3 (low, medium, high humanlikeness) × 2 (cue inconsistency absent or present) design in which participants allocated 30,000 CNY with an AI financial adviser chatbot.
- In Study 1, the serial path from high humanlikeness through reduced psychological distance and then affective trust to heuristic reliance was significant (indirect effect = 0.19, 95% CI [0.11, 0.29]).
- Study 2 used a health-product recommendation task and induced promotion or prevention focus by essay writing; promotion focus strengthened the social closeness pathway (ηp2 = 0.15) and prevention focus strengthened the cognitive evaluation pathway (ηp2 = 0.05).
- Study 3 asked participants to choose a smartphone within a 3,000–5,000 CNY budget and varied how much autonomy the chatbot's framing gave them, alongside three levels of humanlikeness.
- Decision quality was scored 0 to 100 by proximity to an option chosen by independent expert panels (financial planners, dietitians and health professionals, consumer-electronics specialists), with inter-rater agreement of ICC = 0.84–0.91.
- Metacognitive calibration, measured with a Brier score on eight factual questions, moderated how much AI reliance improved decision quality (β = 0.17, p = 0.007).
- When high humanlikeness was paired with cue inconsistency, participants reported more perceived uncertainty and shifted to analytical processing (ηp2 = 0.06).
The full story
The chatbot conditions varied by appearance, name and tone. The low condition showed an abstract AI icon with formal language; the medium used a cartoon avatar with a friendly nickname; the high used a realistic human photograph, a personal name and warm, empathetic language with self-disclosure. A pretest of 45 people confirmed the three levels differed.
Cue inconsistency was built from three subtle flaws: a recommendation partly mismatching stated risk preferences, a repeated question and a contextually mismatched reply. In Study 2, the test of this pathway used the 117 participants who received the inconsistent version.
Participants were randomly assigned to conditions by computer, and those who failed attention checks, finished too quickly or gave straight-line answers were excluded before the final samples. Study 3's calibration task was completed by a subset of 156 participants and was pre-registered as exploratory.
The authors read the results as support for moderate humanlike design and for interfaces that help consumers judge whether a recommendation is reliable. They also pooled all 832 participants, z-standardising measures within each study, to test the model across the three advisory domains.
Sources
1 on file- Behavioral Sciencesdoi:10.3390/bs16081269
Talk it over
Who carries this cost, and what would have to change to prevent it?
Decision outcomes peaked at moderate chatbot humanlikeness in experiments with 832 peoplehttps://truvace.com/item/ai-chatbot-anthropomorphism-and-consumer-decision-making-a-dual-pathway-calibrat
Who carries this cost, and what would have to change to prevent it?
The problem In three online experiments with 832 adults in China, decision outcomes with an AI chatbot adviser followed an inverted U (quadratic b = −0.05, p = 0.001), so maximum humanlike design did not give the best results.
Answer in the threadHow 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 832, from Behavioral Sciences.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat All three experiments used simulated advisory scenarios with participants recruited online from mainland China, and the inverted-U relationship was framed as an exploratory hypothesis.
Answer in the threadHave 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