The Fallacy of Explainable Generative AI: evidence from argumentative prompting in two domains

Musi, E ORCID: 0000-0003-2431-455X and Palmieri, R ORCID: 0000-0002-5122-3058 (2024) The Fallacy of Explainable Generative AI: evidence from argumentative prompting in two domains In: CMNA workshop at COMMA.

Text Fallacy_explainable.pdf - Open Access published version Download (562kB) | Preview

Abstract

This contribution presents a methodology to investigate the soundness of GPT-4 explanations through a combination of fallacy theory and linguistic refinement. It seeks to address the following research questions: Can we leverage Argumentation Theory to i) elicit differences between LLMs’ and human reasoning? ii) build prompting strategies to reduce hallucinations in explanations? To achieve this, we test four prompting strategies using GPT-4 across two domains (HR, loan), prompting the system to generate 30 explanations using analogical, causal, and counterfactual reasoning. We manually annotate the results to assess whether the justifications and the associated reasoning (argument scheme) are sound, fallacious and or influenced by contextual factors. Furthermore, we develop guidelines for prompt engineering to improve the argumentative quality of explanations.

Item Type:	Conference Item (Unspecified)
Divisions:	Faculty of Humanities & Social Sciences Faculty of Humanities & Social Sciences > School of the Arts
Depositing User:	Symplectic Admin
Date Deposited:	14 Oct 2024 07:34
Last Modified:	14 Jun 2025 16:07
URI:	https://livrepository.liverpool.ac.uk/id/eprint/3185048
Disclaimer:	The University of Liverpool is not responsible for content contained on other websites from links within repository metadata. Please contact us if you notice anything that appears incorrect or inappropriate.