Page 696 - ISC PROCEEDINGS 21.4

Table 4. Proportion and test statistics of correct answer (A) and incorrect answer (C)
between DAC before - DAC after and DAWGC before – DAWGC after
DAC DAWGC
before after before after
N 42 42 30 30
Proportion of A 0.3571 0.2143 0.4667 0.2
z-statistics – -1.4486 – -2.1911
p-value (right-tailed) – 0.9263 – 0.9858
Proportion of C 0.4048 0.6429 0.2333 0.5333
z-statistics – 2.1847 – 2.3898
p-value (right-tailed) – 0.0145* – 0.0084**
Source: Calculation results by the authors' team
3.1.2. Mean confidence analysis (Wilcoxon)
To evaluate confidence level, we use Wilcoxon signed-rank test for our variables of
comparison to satisfy all assumptions – dependent variable in ordinal scale, two
categorical ‘related groups’, independence and continuous dependent variable
measurements.
Statistical analysis yields insignificant change in confidence in answer in both DAC
and DAWGC (p-value = 0.3708 and 0.9557 respectively). We conclude that self-confidence
level in our own answer does not change significantly when participants have access to
LLM response in both conditions. This contradicts H2.1 and H2.2.
Table 5. Changes in confidence rate from DAC before to DAC after, and DAWGC before
to DAWGC after
DAC DAWGC
obs s.ranks expected obs s.ranks expected
Positive 10 338 276 7 170 172.5
Negative 6 214 276 8 175 172.5
Zero 26 351 351 15 120 120
All 42 903 903 30 465 465
z 0.895 -0.056
p-value 0.3708 0.9557
Source: Calculation results by the authors' team
3.2. Discussion
On verbalized uncertainty, our findings diverge from those of both Xu et al. (2025)
and Kim et al. (2024). While Kim et al. (2024) demonstrated that first-person uncertainty
expressions reduced participants' tendency to defer to AI responses, participants under
HLC showed no statistically significant improvement in accuracy relative to the CC (p =
0.906), and in fact trended toward worse performance. One plausible explanation lies in
our experimental design: unlike the binary agree/disagree paradigms employed by Kim et
al. (2024), our participants faced a four-option multiple choice question in which the
hesitant LLM response steers participants toward the wrong answer. It is possible that
hedging language functions less as a cue for critical re-evaluation and more as a social
signal of intellectual humility - one that paradoxically renders AI more persuasive by
appearing more trustworthy and self-aware. This interpretation aligns with Cheng et al.
(2025)'s finding that users tend to trust more accommodating AI responses, and with Xu
et al. (2025)'s observation that medium-level uncertainty consistently outperformed both

695