Amount awarded: $10,450

Abstract: Many users now turn to AI for support in navigating their personal lives (Anthropic, 2025). Yet when we outsource cognitive work to technology, our brain-bound capacities for it adapt, change, or diminish (Sparrow et al., 2011; Dahmani & Bohbot, 2020). This project combines behavioural and neural measures to investigate what change may occur when we outsource moral deliberation to an LLM.

This project investigates how outsourcing moral deliberation (MD) to large language models (LLMs) might impact our capacity for MD. The primary aim is to ask whether outsourcing the processing of MD to an LLM changes that processing, measured through behaviour and neural activity. The project aims specifically at moral deliberation as opposed to moral decision making. The motivation for this is that moral deliberation is an integral part of moral cognition, and yet, manifests as a theoretically and neurally distinct phenomenon: deliberation elicits activation of distinct brain regions, for example (Schaich Borg et al., 2011), and Arendt (1971) identifies deliberation as a separate but complementary element in moral competence. However, deliberation has proven difficult to operationalise for experimental purposes (and difficult to conceptualise robustly (e.g., Quilty-Dunn & Krakauer, 2026)). Conceptually, we take moral deliberation to be the conscious, effortful process of identifying and weighing relevant moral considerations over time. To be clear, this need not result in any decision, judgment, or action. This may involve an array of cognitive capacities, articulated through Arendt (1971) who describes moral deliberation as a form of metacognitive, self-reflexive dialogue. However, there's no agreed mechanistic account of deliberation (Quilty-Dunn and Krakauer, 2026; Senghor and Racine, 2022). We therefore target plausible proxies of deliberation. While we utilize a multi-methodological approach, we put behaviour first, and use neuroimaging to complement and constrain our interpretation of the behavioural data (Krakauer et al., 2017). Behaviourally, we test subjects’ ability to reproduce and reformulate their deliberations, and probe them regarding their felt sense of ownership and accountability. We also use NLP to measure the convergence toward homogenization of their written deliberations (taking homogenization to be a further proxy for reduced deliberation, via increased reliance on LLM outputs). The study will involve an EEG element though this will be used not to draw any strong inferences about internal processing per se, but to support us in interpreting behavioural data. Strategically, the project is intended to serve as a pilot, a foundational and rigorous point of departure for further work undertaken by a multi-site international consortium, utilizing diverse methods including expanded behaviour measures, and functional magnetic resonance imaging (fMRI). In service of this, our primary goal with the pilot is to establish stable conceptual, empirical, and infrastructural foundations by leveraging the pilot’s deliverables (e.g., a linguistically standardized stimulus set, a containerised analysis pipeline, and effect-size estimates), in order to seed a public open call via a dedicated website at year-end and secure further public funding.

Description: Many users now turn to AI for support in navigating their personal lives (Anthropic, 2025). Yet when we outsource cognitive work to technology, our brain-bound capacities for it adapt, change, or diminish (Sparrow et al., 2011; Dahmani & Bohbot, 2020). This project combines behavioural and neural measures to investigate what change may occur when we outsource moral deliberation to an LLM.