From Answers to Integrity: Designing Financial AI That Refuses to Mislead

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteur principal: Minchev, Teodor
Format: Recurso digital
Langue:anglais
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901544890269696
author Minchev, Teodor
author_facet Minchev, Teodor
contents <p>This work introduces a governance-first framework for AI-assisted financial reasoning, addressing a critical failure mode in contemporary AI systems: epistemic overreach.</p> <p>As large language models increasingly participate in financial analysis, advisory-adjacent workflows, and decision-support environments, the dominant risk is no longer factual inaccuracy, but the generation of coherent, confident outputs beyond what available evidence, uncertainty, or regime visibility can justify.</p> <p>The paper formalizes an Epistemic Governance Layer built on three core principles:</p> <ul> <li> <p>Cognitive Risk Friction</p> </li> <li> <p>Minimum Analytical Evidence</p> </li> <li> <p>Honesty as a First-Class Output (including refusal and controlled silence)</p> </li> </ul> <p>Together, these mechanisms operate before answer generation, enforcing epistemic discipline at the point of reasoning rather than relying on post-hoc disclaimers or output moderation.</p> <p>Through conceptual explanation, comparative analysis, and practical examples, the work demonstrates how governance-first systems differ structurally from conventional financial AI assistants that optimize for responsiveness, fluency, and perceived usefulness.</p> <p>The framework is designed to be applicable in regulated and advisory-adjacent environments, including brokers, banks, asset managers, and research platforms, where controlling false authority, suitability leakage, and action-enabling language is critical.</p> <p>Rather than helping users decide faster, the approach prioritizes decision integrity — ensuring that AI systems know when not to answer, when to downgrade analysis, and when silence is the most responsible output.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18229279
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle From Answers to Integrity: Designing Financial AI That Refuses to Mislead
Minchev, Teodor
AI-assisted financial analysis Model uncertainty Risk-aware AI systems Advisory-adjacent AI AI explainability AI compliance design Financial regulation and AI Human-AI decision interfaces AI epistemology
<p>This work introduces a governance-first framework for AI-assisted financial reasoning, addressing a critical failure mode in contemporary AI systems: epistemic overreach.</p> <p>As large language models increasingly participate in financial analysis, advisory-adjacent workflows, and decision-support environments, the dominant risk is no longer factual inaccuracy, but the generation of coherent, confident outputs beyond what available evidence, uncertainty, or regime visibility can justify.</p> <p>The paper formalizes an Epistemic Governance Layer built on three core principles:</p> <ul> <li> <p>Cognitive Risk Friction</p> </li> <li> <p>Minimum Analytical Evidence</p> </li> <li> <p>Honesty as a First-Class Output (including refusal and controlled silence)</p> </li> </ul> <p>Together, these mechanisms operate before answer generation, enforcing epistemic discipline at the point of reasoning rather than relying on post-hoc disclaimers or output moderation.</p> <p>Through conceptual explanation, comparative analysis, and practical examples, the work demonstrates how governance-first systems differ structurally from conventional financial AI assistants that optimize for responsiveness, fluency, and perceived usefulness.</p> <p>The framework is designed to be applicable in regulated and advisory-adjacent environments, including brokers, banks, asset managers, and research platforms, where controlling false authority, suitability leakage, and action-enabling language is critical.</p> <p>Rather than helping users decide faster, the approach prioritizes decision integrity — ensuring that AI systems know when not to answer, when to downgrade analysis, and when silence is the most responsible output.</p>
title From Answers to Integrity: Designing Financial AI That Refuses to Mislead
topic AI-assisted financial analysis Model uncertainty Risk-aware AI systems Advisory-adjacent AI AI explainability AI compliance design Financial regulation and AI Human-AI decision interfaces AI epistemology
url https://doi.org/10.5281/zenodo.18229279