Page 698 - ISC PROCEEDINGS 21.4

changed their answers without reporting a corresponding increase in certainty -
suggesting that answer revision may have been driven less by genuine persuasion and
more by an implicit social pressure to align with the AI, even in the absence of felt
conviction. This finding adds nuance to Fernandes and Welsch's (2025) account of the
"illusion of competence," in which AI-assisted users exhibit inflated post-task confidence;
however, in our study, confidence remained stable, yet behavior still shifted in LLM’s
direction. This points to a form of behavioral compliance that operates independently of
metacognitive updating.
An approach is to measure dissonance arousal explicitly - for instance, through
ordinal scales – to quantify the existence of self-doubt or discomfort experienced upon
encountering a counterintuitive LLMs response, and to examine whether such arousal
mediates the rate of answer revision. Cognitive dissonance as a mediator of responses
can be modelled after Spencer et al. 's (1999) research about the effect of stereotype
threat on women, where several factors (self-efficacy, anxiety, evaluation apprehension)
are regarded as mediators of math results. The implication is that participants may
experience less dissonance while conducting tasks simultaneously with LLMs, which leads
to improved performances and higher self-confidence in one’s response because they are
aware of that mental process and therefore will avoid adopting it – that is the mechanism
behind the ‘stereotype threat’ towards women in Spencer et al. (1999), one which
women participants strive to disprove their inferiority relative to their male counterparts
in math.
A second direction concerns the moderating role of prior confidence: participants
who were highly confident in their initial answer may have experienced more acute
dissonance upon receiving a contradictory AI response, and may consequently have
revised their answers at different rates than less confident participants. Manipulating AI's
expressed confidence level while holding accuracy constant would allow researchers to
isolate the contribution of the AI's apparent certainty to dissonance induction. Finally, the
observed resistance to change in the hesitant condition, though insignificant, invites
replication with larger samples - it remains possible that verbalized uncertainty functions
as a dissonance buffer, attenuating the psychological conflict that arises from receiving a
contradictory but self-assured AI response, thereby reducing the pressure to revise one's
own answer.
4. Conclusion
This study evaluates the impact of behavioral nudges on students’ reliance on AI,
specifically examining whether such interventions can mitigate the risks of cognitive
offloading in academic research. Our findings reveal a critical tension between the
convenience of AI integration and the preservation of independent critical thinking.
Contrary to the intended goal of fostering responsible use, the experimental nudges -
specifically cognitive forcing and verbalized uncertainty - failed to reduce overreliance.
Instead, they exposed a deep-seated authority bias. Students interpreted AI hesitancy not
as a prompt for skepticism, but as a signal of intellectual humility, which paradoxically
lowered the perceived psychological cost of deferring to the system.
Furthermore, the significant shift from correct personal judgments to AI-endorsed
errors under cognitive forcing conditions (DAC and DAWGC) suggests that these nudges
may inadvertently intensify cognitive dissonance. Rather than utilizing the forced delay to
re- evaluate the AI’s logic, students appeared to prioritize dissonance reduction by
aligning their behavior with the authoritative AI output. This confirms that cognitive

697