Page 691 - ISC PROCEEDINGS 21.4

2. Theoretical framework, literature review and methodology
2.1. Theoretical background and literature review
2.1.1. Behavioral nudges
Nudges, as defined by Thaler & Sunstein (2008) in behavioral economics, are
psychological mediators that influence human behavior with a predictable pattern –
without having to resort to coercions or taxes. Nudges are prevalent in the economics of
food, gambling and various other sectors; for example, using a smaller spoon size in the
sugar bowl to regulate sugar level in hot beverages or placing healthy food option near
the cash register desk (Venema et al., 2020; Kroese et al., 2016). Benartzi et al. (2017)
provides justification for behavioral interventions as a more cost-effective and impactful
approach relative to tax imposition, educational programs or rewards. Given that
informational literacy are occasionally ignored and/or are overwhelming to process for
individuals with limited cognitive capacity – who are susceptible to information overload
when confronted with large volumes of complex data simultaneously (Sweller, 1988) –
behavioral nudges offer a more accessible and scalable alternative for guiding decision-
making without imposing additional cognitive burden.
The use of behavioral nudges in LLMs has been explored and evaluated in some
recent literature. One is verbalized uncertainty, which are LLMs responses characterized
by the expression of phrases with uncertain connotation such as ‘It could be...’,
‘somewhat’ or ‘kind of’. Xu et al. (2025) has found that medium verbalized uncertainty in
LLM expressions consistently leads to higher user trust, satisfaction, and task
performance compared to both high and low levels, and that users' perceptions of
uncertainty vary depending on the LLM's accuracy level. Kim et al. (2024) found that first-
person uncertainty expressions such as "I'm not sure, but..." reduced participants'
tendency to agree with the system's answers and increased overall accuracy, primarily by
diminishing overreliance on incorrect outputs - while expressions framed from a general
perspective produced weaker and statistically non-significant effects. Taken together,
these findings suggest that calibrating both the degree and framing of expressed
uncertainty functions as a subtle nudge, shaping how much users rely on AI-generated
suggestions.
Another one is cognitive forcing, which is based on the dual process theory – human
brains are prone to fast, automatic (Type 1) thinking rather than slower, more analytical
(Type 2) reasoning (Bellini-Leite, 2022) – and the role of cognitive forcing is to reverse this
pattern, interrupting users’ heuristic cognitive processes and forcing them to engage in
more deliberate analysis before arriving at the final decision. Bucinca et al. (2021) stated
that under cognitive forcing (on demand – AI suggestion, update – making decision before
AI, wait – waiting for AI to process the image before seeing the suggestion) performance
was no better than normal condition, where AI provided immediate answer. Meanwhile,
de Jong et al. (2025) found out that full and partial suggestions significantly improve
participants’ performance in both tasks, with accuracy in the full suggestions condition
significantly higher than that of partial ones in the second task. The aforementioned
findings suggest that the effect of cognitive forcing varies and more research needs to be
conducted to further explore this psychological tool.
Until now, there is only a few research assessing both conditions simultaneously,
except for Vejandla et al. (2025) where uncertainty is treated as a cognitive forcing trigger.
Also, in both Bucinca et al. (2021) and de Jong et al. (2025), the suggestions provided by
AI direct participants to the right answer, which fails to assess students' trust in LLMs

690