[Paper Review] Insights for an AI Whistleblower Office from 30 Case Studies

Ethan Beri, Mauricio Baker|arXiv (Cornell University)|Mar 1, 2026

Ethics in Business and Education0 citations

TL;DR

A dataset of 30 whistleblower case studies is analyzed to design an AI whistleblower office, yielding policy recommendations on rewards, protections, anonymity, processing capacity, and messaging.

ABSTRACT

Whistleblower programmes are a promising tool for uncovering noncompliance with AI regulations. This paper aims to help policymakers design an AI whistleblower programme by giving them an understanding of whistleblowers' motivations, and of the overall whistleblowing process. We take an empirical approach, assembling a dataset of 30 case studies of whistleblowers. This dataset includes dozens of features of each case, which range from 1978 to 2020 and span 15 industries. Our findings suggest that whistleblower programmes will be more effective if they financially reward whistleblowers, provide protections for whistleblowers, enable whistleblowers to report anonymously, are adequately staffed and funded, and provide advice to potential whistleblowers. We provide ten concrete policy recommendations for an AI whistleblower programme at the end of this paper.

Motivation & Objective

• Understand whistleblowers’ motivations and the whistleblowing process in AI-relevant contexts.
• Characterize demographics, organizations, and types of wrongdoing in case studies.
• Identify factors affecting reporting and retaliation to inform policy design for AI whistleblower programs.

Proposed method

• Assemble a dataset of 30 whistleblower case studies drawn from public lists (1978–2020, 15 industries).
• Extract 58 fields per case across seven thematic groups (whistleblower, organisation, wrongdoing, whistleblowing, relationship dynamics, motivation).
• Code motivations with cautious, evidence-backed judgments and include an evidence column for each motivation field.
• Analyze demographics, organization sizes, sectors, retaliation, anonymity, and process progression using descriptive statistics and ranges.
• Acknowledge limitations such as sampling bias toward high-profile cases and the small n (n=30).
• Provide policy-relevant recommendations grounded in observed patterns.

Figure 1: Prevalence of motivating (left) and demotivating (right) factors. Note we took a conservative, careful approach to ascribing motivation: each case used a range of sources, and we preferred a sparse field to an incorrect one. More on this in Methodology and dataset.

Experimental results

Research questions

• RQ1 What motivates whistleblowers to report wrongdoing in AI-related contexts?
• RQ2 How do whistleblowing processes unfold from initial tip to external reporting?
• RQ3 What protections, anonymity, and incentives affect whistleblower participation and safety?
• RQ4 What design features would make an AI whistleblower office effective and trustworthy?

Key findings

• Whistleblowers are predominantly insiders (≥90%) and often middle-aged, well-educated; many are employees of large organizations (>10,000 employees in 57–73% of cases).
• Most wrongdoing is ongoing (80%), with typical duration around 3 years for cases with known end dates, and insider relationships are common (≥90%).
• Retaliation is common (57–67%), including harassment, unjust termination, and death threats (13–23%); anonymity is rare (13%).
• Motivations are largely moral (≥87%), but often include multiple motives (including financial, social, or protection motives); demotivating factors include retaliation and distrust in action.
• Effective AI whistleblower Office features include financial rewards (10–30% of sanction amounts in SEC/CFTC-style programs as a reference), strong protections, anonymous reporting options, adequate staffing, and clear messaging/advice for potential whistleblowers.
• Ten concrete policy recommendations are provided, including how rewards should relate to sanctions, protection mechanisms, anonymity options, tip processing capacity, and outreach strategies.

Figure 2: Whistleblower ages at the time of making their first report, bracketed (to include estimated ages, in cases where exact ages were unavailable).