This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
The Department of Government Efficiency (DOGE) deployed an AI tool using outdated LLMs to evaluate and flag Veterans Affairs contracts for cancellation. The tool suffered from significant hallucinations, including inflating contract values and misclassifying essential healthcare and research services as expendable. This resulted in the termination of at least two dozen critical contracts, negatively impacting veteran care and research projects.
A purported AI tool using LLMs was deployed to classify Veterans Affairs contracts as expendable based on limited text and simplified criteria. The system allegedly produced hallucinated values and flagged critical healthcare and research services for cancellation. Reportedly, at least two dozen flagged contracts were later terminated.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
DOGE and VA officials over-relied on a hastily built, unvalidated AI tool to make critical budgetary decisions affecting veteran healthcare.
The AI system failed to perform reliably, producing glaring errors and hallucinations regarding contract values and essential services due to technical limitations.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk was caused by the AI system's technical failures and hallucinations after it was deployed to analyze VA contracts.
Added categories peak at 2: Minor (Epistemic Harm, Direct)
The deployment of an unvalidated, flawed AI tool by DOGE to evaluate VA contracts led to severe hallucinations and the erroneous cancellation of critical veteran healthcare and research services. While causing direct harm to veteran care and procurement processes, the incident remains a domestic administrative failure with minor national security implications.
The AI tool acted as an Assistant, designed to parse contracts and flag non-essential ones to help humans make budgetary decisions. The developer stated, 'I just delivered the [list] to the VA employees' and confirmed that humans made the final termination decisions, meaning the AI augmented human work rather than acting autonomously.
High Risk: The AI system was used in healthcare administration and public services (Veterans Affairs), directly impacting access to essential healthcare services and medical research, which has significant implications for safety and fundamental rights.
The primary goal of the AI tool was to evaluate VA contracts and identify which ones were non-essential ('munchable') to assist in budget cuts and resource allocation. It also performed value estimation by attempting to extract and calculate the total value of each contract.
The reports from ProPublica and Seifeur provide highly detailed, consistent accounts of the AI tool's development, its specific technical flaws (hallucinations, 2500-word limit), the developer's background and admissions, and the specific real-world impacts (the cancellation of 24 contracts including cancer research). The role of the AI is explicitly documented.
None. The tool's failure is clearly a mix of technical limitations (7.3) and human overreliance/rushed deployment (5.1).
Technology
Outdated LLM Models
Used inexpensive, older models incapable of solving complex tasks.
Model Hallucinations
AI hallucinated contract sizes, inflating value of small contracts.
Flawed Prompt Engineering
Prompts lacked essential context about VA operations and legal mandates.
Data Inputs
Truncated Input Text
Analyzed only the first 2500 words, missing crucial contract details.
Inaccurate Value Extraction
AI selected wrong dollar values when multiple figures were present.
Ignored Public Databases
Failed to verify contract data against readily available public sources.
Human Factors
Lack of AI Expertise
Developer had no formal training in artificial intelligence.
No Healthcare Background
Developer had no experience in healthcare or government procurement.
Cognitive Overload on Staff
VA employees faced extreme pressure to justify contracts in hours.
Process and Methods
Rushed Software Development
The tool was built in a single day using AI-assisted coding.
Restricted Appeal Process
Staff restricted to 255-character responses to defend contracts.
Opaque Review System
VA staff could not see who made final decisions to cancel contracts.
Management
Speed Prioritized Over Safety
Management pushed for rapid cuts without evaluating operational risks.
Inadequate Risk Assessment
Deployed unvalidated AI tools for critical budgetary decisions.
Opaque Decision-Making
Leadership failed to provide transparency on final cancellation choices.
Regulatory Environment
Aggressive Federal Timeline
Executive order required a complete contract review within 30 days.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
