DOGE was found to have used an error-ridden AI tool to determine whether medical service contracts could be cancelled
The Department of Government Efficiency (DOGE) is working to reduce waste in the US government and is investigating various government contracts to determine whether they can be canceled.
DOGE Developed Error-Prone AI to Help Kill Veterans Affairs Contracts — ProPublica
https://www.propublica.org/article/trump-doge-veterans-affairs-ai-contracts-health-care
Inside the AI Tool Used by DOGE to Review Veterans Affairs Contracts — ProPublica
https://www.propublica.org/article/inside-ai-tool-doge-veterans-affairs-contracts-sahil-lavingia
ProPublica Explains How DOGE's AI Cut Support for Veterans Care - emptywheel
https://www.emptywheel.net/2025/06/06/propublica-explains-how-doges-ai-cut-support-for-veterans-care/
Lavingia is the founder of Gumroad , an internet-based payment service, and joined DOGE as a software engineer involved in contract review for the Department of Veterans Affairs on March 17, 2025. Lavingia feels that government agencies have a greater impact than private companies, and has long wanted to write code for the federal government.
Originally, to streamline the tedious task of 'finding PDFs of contracts, contacting the person in charge and determining whether they are necessary,' Lavingia created an AI-based tool. The AI tool, named 'MUNCHABLE,' had the functions of 'iteratively processing PDFs of contracts,' 'extracting text,' 'using large-scale language models to determine whether a contract can be canceled,' and 'collecting flagged contracts.'
Mr. Lavingia had the opportunity to speak with Elon Musk at the DOGE general meeting, where he asked if he could open source the code he wrote, and received the answer, 'Yes.' Therefore, Mr. Lavingia has published the code for the AI tool he created to review the Department of Veterans Affairs contract on GitHub.
GitHub - slavingia/va
https://github.com/slavingia/va
About three weeks after joining DOGE, Lavingia returned to his home in New York and began working remotely. During that time, he believed Musk's words that he wanted to maximize DOGE's transparency, and gave an interview to the online media Fast Company. In it, Lavingia gave a fairly frank impression, saying, 'To be honest, the government wasn't as inefficient as I thought it would be. I expected it to be easier to get results.'
Thanks to DOGE, Gumroad's founder has a second job with the VA - Fast Company
https://www.fastcompany.com/91330297/doge-sahil-lavignia-gumroad
On May 9, 2025, the day after this interview was published, Lovingia was suddenly forced to rescind his access to DOGE and kicked out of DOGE. As a result, Lovingia said that although he was glad to have the opportunity to write code for the federal government, he was disappointed that he was unable to achieve the results he had initially hoped for, such as improving the UX for veterans' disability pension applications and automating application processing.
In a blog post, Lavingia explains what he did with DOGE.
DOGE Days
https://sahillavingia.com/doge
After being fired from DOGE, Lavingia told foreign media outlet NPR in an interview, 'I didn't feel that the federal government was full of waste, fraud or abuse of power.' 'The government has been under constant scrutiny for decades, so I'll be honest with you, personally, I was surprised at how efficient the government is. That doesn't mean that the government can't be made even more efficient. You can get rid of paper and faxes, but that doesn't necessarily mean fraud, waste or abuse. It just means that there is room to modernize and improve the U.S. federal government to be more 21st century-ready.'
Former engineer shares his experience working for DOGE : NPR
https://www.npr.org/2025/06/02/nx-s1-5417994/former-doge-engineer-shares-his-experience-working-for-the-cost-cutting-unit
ProPublica reported that they had six experts examine the code for Lavingia's AI tool and the contracts flagged by the AI tool, which they obtained from a source, and all of them pointed out flaws in the script.
For example, the AI tool often misinterpreted contracts and inflated their value, misreading them and inflating contract amounts. The AI tool determined that more than 1,000 contracts were worth $34 million, but some contracts were worth only a few thousand dollars.
The AI tool was also instructed to only review the first 10,000 characters (about 2,500 words) of the contract, which Lovingear explains is because the VA only had access to an older AI that it had already contracted with, and that AI could only process 10,000 characters.
In addition, the AI tool was instructed not to classify contracts related to 'direct patient care' as unnecessary, but even this judgment was questionable, with some contracts related to the maintenance of critical equipment used to position patients during treatment being classified as contracts that could be eliminated.
'I think mistakes were made, that's for sure. Mistakes happen all the time. I would never recommend that someone run my code and follow it exactly,' Lovingia told ProPublica. He argued that terminating a contract without scrutinizing the output of an AI tool would be like driving into a lake because Google Maps told you to 'keep going.'
The AI tool determined that more than 2,000 contracts were 'cancellable,' but it is unclear how many contracts were actually canceled or are expected to be canceled in the future. The VA said it canceled 600 contracts overall, but the breakdown is largely unknown. The VA has 24 contracts on the DOGE reduction list so far, including a 'contract for the maintenance of gene sequence analyzers to help improve cancer treatment,' a 'contract for blood sample analysis to support VA research projects,' and a 'contract for the provision of additional tools to measure and improve the care provided by nurses.'
This incident has also become a hot topic on the social news site Hacker News.
Doge Developed Error-Prone AI Tool to 'Munch' Veterans Affairs Contracts | Hacker News
https://news.ycombinator.com/item?id=44199887
One user commented , 'If you knew it was wrong, why did you write the code? You were just having fun?' Another said , 'People in Silicon Valley are out of touch with society as a whole.'
in Note, Posted by log1h_ik