City Research Online

Abstract

This thesis presents three empirical projects examining judgment and reasoning across evaluative, legal, and narrative domains, using both human participants and large language models.

The first project investigates the less-is-more effect in reward evaluation, that is, the paradoxical finding that adding a positively valued reward to another can reduce the overall perceived value of the combination. Drawing on ideas from quantum probability theory, we propose that this effect arises from the difficulty of jointly representing the components of a reward bundle, a property formalised as representational incompatibility. Across five experiments, we show that less-is-more effects co-occur with order effects at the level of individual participant–pair observations, where order effects refer to systematic sensitivities to the sequence in which rewards are presented. This co-occurrence is consistent with the prediction that less-is-more effects are related to the representational difficulty indicated by order sensitivity, and the association was observed across different presentation formats and experimental conditions.

The second project examines whether large language models exhibit cognitive biases analogous to those observed in human legal decision-making. Using a controlled judgment paradigm adapted from prior human research, we show that GPT-4o and GPT-5.2 exhibit systematic order effects, whereby judgments are influenced by the sequence in which incriminating and exonerating evidence is presented, as well as evaluation biases arising from the requirement to make intermediate judgments. These biases persist across prompt engineering interventions, suggesting that simple changes to task instructions do not provide a robust safeguard against procedural sensitivity in evaluative judgment tasks.

The third project investigates whether formal tools developed for studying causal reasoning can be used to study aspects of narrative reasoning. The approach treats narratives as systems of causally related events and combines causal elicitation with counterfactual judgment tasks, allowing predictions derived from causal models to be compared with participants’ responses. Across a series of pilot studies and experiments, the results provide an initial evaluation of causal graph modelling as a framework for studying how people understand and reason about narratives, while also identifying important methodological challenges.

Together, these projects demonstrate the value of formal approaches for investigating how people and artificial systems represent, evaluate, and reason about information.

Publication Type:	Thesis (Doctoral)
Subjects:	B Philosophy. Psychology. Religion > BF Psychology Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Health & Medical Sciences > Department of Psychology & Neuroscience School of Health & Medical Sciences > School of Health & Medical Sciences Doctoral Theses Doctoral Theses

Export

Downloads