Inside the AI Prompts DOGE Used to “Munch” Contracts Related to Veterans’ Health

Join ProPublica, a nonprofit newsroom dedicated to uncovering abuses of power, to receive our top stories upon publication.

An AI script developed by a Department of Government Efficiency employee flagged an internet service contract as cancelable due to unclear and conflicting instructions, rather than actual waste. The script, written by Sahil Lavingia, was tasked with canceling anything not directly supporting patient care, but lacked the knowledge to make such determinations. Mistakes were acknowledged by Lavingia, who worked at DOGE for two months, as the department rushed to review all VA contracts following President Trump’s executive order.

ProPublica obtained the code and prompts used in the contract review process, revealing significant flaws. The code utilized outdated models unsuitable for the task, resulting in errors like hallucinating contract amounts and not analyzing full contract texts. Experts criticized the use of off-the-shelf AI models without proper context of VA operations.

Lavingia admitted to flaws in his work, attributing them to time constraints and inadequate tools. The prompts provided to the AI model lacked clarity and essential information, leading to erroneous contract classifications. Despite efforts to correct errors, many contracts were inaccurately flagged for cancellation, including vital services like VA office internet provisions.

The analysis highlighted issues with the AI’s inability to discern critical contract details and determine “munchability” accurately. The lack of clear instructions and context contributed to flawed decision-making by the model. VA’s reliance on AI for contract review was questioned, with concerns raised about the potential impact on essential services and patient care.

The VA defended its AI usage for contract evaluations, citing it as a logical step forward. Plans to expand AI applications within the VA were hinted at in internal communications. If you have information on AI misuse in government agencies, contact ProPublica via Signal or SecureDrop. Investigative journalist Brandon Roberts, experienced in AI analysis, is available for consultation.