Sep 2026· International Journal of Economics and Financial Issues· Vol 16, pp. 51-61· 0 citations
TL;DR
The risk of hallucination when using large language models (LLMs) for financial research is examined and the unreliability of off-the-shelf tools such as ChatGPT for tasks requiring precise financial reasoning is demonstrated.
Abstract
This article examines the risk of hallucination when using large language models (LLMs) for financial research and demonstrates the unreliability of off-the-shelf tools such as ChatGPT for tasks requiring precise financial reasoning. A simple, replicable experiment was conducted using the free online version of ChatGPT-4o. The same prompt concerning capital budgeting studies advocating dual discount rates for cash inflows and outflows was submitted on three consecutive days to assess the consistency and accuracy of the model’s responses. The results reveal extensive fabrication of academic references and significant logical inconsistency. Approximately half of the cited studies were fictitious and could not be verified. In addition, the model produced contradictory responses regarding the appropriate discount rates for different cash flows, with its recommendations reversing across days. These findings provide further evidence that off-the-shelf LLMs remain unreliable for precise financial applications and highlight the need for careful use and independent verification by researchers.
A narrative review distinguishes fabrication, the invention of an entirely non-existent source, from unfaithfulness, the citation of a genuine source in support of a claim it does not contain, and argues that the two failure modes require different detection strategies.
M. Rana, M. Alkhlewi, Turki Abdulaziz Alsohaibani et al.· European Journal of Prosthod...· 0 citations
The concern over hallucination, a situation where a model generates fluent but factually inaccurate or fabricated information-has grown with the steady development of Large Language Models like GPT and Gemini. Such results have the potential to erode safety, dependability, and trust in AI driven fields such as journali...
Shanya Kumari, P. Bagane, U. Waghmode et al.· International Conference on...· 2 citations
The study reveals that AI hallucinations have three kinds of manifestations, which include content distortion, evidence failure and improper argumentation, which comprises factual error, false or irrelevant citation, and superficially coherent arguments lacking evidence.
It is shown that epistemic uncertainty leaves a measurable trace within intermediate layers, where spikes in attention entropy are associated with reasoning breakdowns, and that attention dispersion is a promising complement to traditional hallucination detection methods.
This work proposes a five-step diagnostic protocol covering feature correlation, bootstrap stability, sparse versus dense ranking disagreement, intervention baselines, and cross-dataset evaluation as a minimum standard for sparse-neuron localization claims, and investigates prior work using the proposed approach.
Huseyin Cavus, Sebin Sabu, J. Spear et al.· arXiv.org· 0 citations
Bounded margins mitigate confident hallucinations during post-training, implemented through an entropy-dependent margin bound in direct preference optimization (DPO) and shown to mitigate confident hallucinations during post-training.