The AI-Fraud Diamond: A Novel Lens for Auditing Algorithmic Deception

Benjamin Zweers, Diptish Dey, Debarati Bhaumik · 2025

Read The AI-Fraud Diamond: A Novel Lens for Auditing Algorithmic Deception by Benjamin Zweers, Diptish Dey, Debarati Bhaumik (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-13984.txt.

Introduction

Benjamin Zweers, Diptish Dey, and Debarati Bhaumik extend the classic fraud triangle to cover deception by AI systems, in a 2025 paper. Fraud analysis traditionally weighs pressure, opportunity, and rationalization. AI adds a fourth corner, technical opacity, since harm can emerge from hidden model behavior, bad training data, or careless deployment without clear human intent. The paper sorts AI fraud into five types, from data tampering to synthetic misinformation and ethics-washing.

Interviews with auditors at two major firms show the strain: too little technical skill, siloed teams, and blocked access to system internals. Those gaps blunt detection and accountability. The authors urge auditors to shift from checking outcomes to diagnosing system weaknesses. Their diamond framework aims to ground future audit methods for automated decision systems.

Catalog record

CREATORBenjamin Zweers, Diptish Dey, Debarati Bhaumik
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-13984

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The AI-Fraud Diamond: A Novel Lens for Auditing Algorithmic Deception Benjamin Zweers, Diptish Dey, Debarati Bhaumik Abstract As artificial intelligence (AI) systems become increasingly integral to organizational processes, they introduce new forms of fraud that are often subtle, systemic, and concealed within technical complexity. This paper introduces the AI-Fraud Diamond, an extension of the traditional Fraud Triangle that adds technical opacity as a fourth condition alongside pressure, opportunity, and rationalization. Unlike traditional fraud, AI-enabled deception may not involve clear human intent but can arise from system-level features such as opaque model behavior, flawed training data, or unregulated deployment practices. The paper develops a taxonomy of AI-fraud across five categories: input data manipulation, model exploitation, algorithmic decision manipulation, synthetic misinformation, and ethics-based fraud. To assess the relevance and applicability of the AI-Fraud Diamond, the study draws on expert interviews with auditors from two of the Big Four consulting firms. The findings underscore the challenges auditors face when addressing fraud in opaque and automated environments, including limited technical expertise, insufficient cross-disciplinary collaboration, and constrained access to internal system processes. These conditions hinder fraud detection and reduce accountability. The paper argues for a shift in audit methodology—from outcome-based checks to a more diagnostic approach focused on identifying systemic vulnerabilities. Ultimately, the work lays a foundation for future empirical research and audit innovation in a rapidly evolving AI governance landscape. Introduction and Motivation In many cases, it works exactly as designed; yet, it produces outcomes that reinforce inequality or facilitate fraud ( Benjamin 2019 ; Birhane and Prabhu 2021 ) . As artificial intelligence (AI) becomes more deeply embedded in business, government, and daily life ( Eubanks 2018 ) , it creates new vulnerabilities, ranging from data manipulation ( Yerlikaya and Şerif Bahtiyar 2022 ) and biased sampling ( Fukuchi, Hara, and Maehara 2020 ) to the deliberate exploitation of technical opacity ( Asatiani et al. 2020 ) . These forms of potential fraud are difficult to detect, often hidden within systems so complex that even developers cannot fully explain them ( von Eschenbach 2021 ; Lipton 2018 ) . Recent real-world cases illustrate how algorithmic decision-making can translate this opacity into harmful and fraudulent outcomes. For example, Amazon’s recruitment algorithm systematically disadvantaged women ( Reuters 2018 ) , credit scoring tools assigned lower limits to female applicants with identical financial histories as men ( Vigdor 2019 ) , and mortgage algorithms charged higher interest rates to Black and Hispanic applicants than to their White counterparts ( Fuster et al. 2022 ) . These are not isolated incidents; they reveal how for example biased data, flawed optimization goals, or lack of oversight can embed and even amplify existing inequalities, often with discriminatory effects. Amid growing concerns over AI accountability, regulators have pushed for ethical auditing frameworks, such as the EU’s Ethics Guidelines for Trustworthy AI ( Smuha 2019 ) and the U.S. Algorithmic Accountability Act ( 116th Congress 2019 ) , emphasizing transparency, fairness, and responsibility. However, most of these guidelines remain voluntary and non-binding, lacking the specificity needed to detect or address fraud in AI systems ( Organisation for Economic Co-operation and Development 2025 ; National Institute of Standards and Technology 2024 ) . While scholars like ( Floridi and Cowls 2019 ) have proposed principles for trustworthy AI, they often fall short of providing practical tools for fraud-focused audits. The academic focus has primarily been on using AI to detect fraud ( Dash et al. 2023 ) , rather than preventing fraud within AI systems themselves, thereby, leaving a critical gap in oversight and risk mitigation ( Morley et al. 2021 ; Raji et al. 2020 ) . Potential AI-related fraud is especially dangerous because it silently reproduces harm at scale. As AI systems influence access to resources such as housing, employment, credit, and public services, undetected fraud risks reinforce inequality and marginalization ( Benjamin 2019 ; Dastin 2018 ) . These harms often go unnoticed, hidden by system opacity and unclear accountability ( Fuster et al. 2022 ; Buolamwini and Gebru 2018 ) . They are difficult to detect and even harder to challenge. Left unchecked, they erode individual rights, social justice and public trust ( Chesney and Citron 2019a ) . Addressing this is not just a technical task; it is essential for preserving legitimacy in an automated society. The ethical horizon articulated by Rawls ( Rawls 1971 ) will fail to materialize due to lack of appropriate mechanisms, as envisioned by Bourdieu ( Claridge 2018 ) , which ultimately will lead to erosion of collective trust ( Luhmann 1997 ) . This paper examines the potential of audit mechanisms to detect and mitigate fraud-related risks arising during the development and deployment of AI systems. It argues that while existing ethical frameworks provide a foundational basis, they are insufficient for addressing deliberate manipulations obscured within the technical complexity of AI. The paper introduces a taxonomy of AI-fraud, critically analyzes the limitations of current audit practices in this context, and highlights the challenges posed by the technical opacity of AI systems and the inadequacies of the traditional Fraud Triangle. It proposes the AI-Fraud Diamond as a more suitable analytical tool, and presents preliminary validation through primary research involving practitioners from two of the Big Four consulting firms. Understanding AI-Fraud Fraud in AI systems is not always loud or obvious. It can be quiet, systemic, and deeply embedded, shaped by for example opaque algorithms, misused data, and unchecked organizational practices. Unlike traditional forms of fraud, AI-enabled deception often hides in complexity, evading detection while producing far-reaching consequences. This section maps the diverse ways in which AI-fraud can manifest, using a structured taxonomy to classify emerging patterns. From manipulated data and stolen models to synthetic deception and ethics washing, these categories help illuminate where and how fraud takes root in algorithmic environments. The section also introduces the classic Fraud Triangle as a conceptual lens to reflect on how traditional fraud theories might adapt to AI-driven risks. Fraud in AI systems involves the intentional misuse of data, models, or outcomes—whether to deceive, exclude, or gain unfair advantage. It is a distinct and systemic risk that requires its own approach to detection and oversight. To support a clearer understanding of how such fraud manifests, this section introduces a taxonomy grounded in the classification method of ( Nickerson, Varshney, and Muntermann 2013 ) . As presented in Table 1, the taxonomy consists of five categories: Input Data Manipulation, Model Exploitation & Evasion, Algorithmic Decision Manipulation, Synthetic Misinformation & Deception, and Unregulated AI & Ethics Fraud. Together, these categories are designed to capture the full range of known AI-related fraud mechanisms, from data manipulation at the input stage to systemic misuse beyond regulatory oversight. Each category represents a different mode of fraudulent behavior: for example, poisoning training data, bypassing model safeguards, or deploying misleading outputs without regulatory oversight. Fraud Categories Underlying Mechanism(s) Manifestations Input Data Manipulation Fraud, in which data is manipulated before or during the training of AI models to influence the output. ( Biggio and Roli 2018 ; Xu, Evans, and Qi 2017 ; Jagielski et al. 2018 ) Data Poisoning and Synthetic Data Generation Model Exploitation and Evasion Fraud in which attackers directly manipulate,

Read the complete text