Propublica is a nonprofit newsroom that investigates power abuse. Sign up and receive the biggest story as soon as it’s published.
When the Trump administration prepared to cancel contracts at the Veterans Affairs Bureau this year, authorities turned to software engineers with no medical or government experience to guide them.
Engineers working in the government’s Ministry of Efficiency have quickly built artificial intelligence tools to identify which services from private companies are not essential. He labeled those contracts as “muddy.”
This code used an outdated and inexpensive AI model to produce results with obvious mistakes. For example, they hallucinated the scale of contracts and frequently misinterpreted them to inflate their value. They concluded that it was worth $34 million each, but some actually only $35,000.
Doge AI Tools has flagged more than 2,000 contracts for “mushing.” It is unclear how many people will be cancelled or if they are on track. The Trump administration’s decision on the Virginia contract was largely a black box. VA uses contractors for a variety of reasons, including supporting hospitals, research and other services aimed at caring for veterans.
VA officials say they have killed nearly 600 contracts overall. Congressional Democrats are pushing VA leaders concrete details of what was cancelled without success.
We have identified at least 20 of our previously cancelled Doge lists. Some of the cancelled contracts maintained gene sequencing devices used to develop better cancer treatments. The other was for blood sample analysis in support of the VA research project. The other was to provide additional tools to measure and improve the care nurses provide.
Propublica has obtained flagged codes and contracts from sources and shared them with half a dozen AI and procurement experts. Everything said the script was flawed. Many criticized the concept of using AI to guide budget cuts in VAs.
Cary Coglianese, a professor of law and political science at the University of Pennsylvania who studies government use and regulations on artificial intelligence, said he was troubled by the use of these generic, large-scale language models, or LLMS. “I don’t think there’s a lot of credibility in something as complex and involved as this,” he said.
Sahil Lavingia, a programmer who enlisted in Doge, later admitted a flaw in the code as it was run by Elon Musk.
“I think a mistake was made,” said Levisia, who had worked at Dozia for nearly two months. “The mistakes were certainly true. The mistakes were always made. I would never recommend someone run my code and do what it says. It’s like the ‘Office’ episode of Steve Carell driving to the lake, as Google Maps says it drives to the lake. ”
Lavingia had previously spoken about his time at Doge, but this was the first time his work has been examined in detail, and the first time he has publicly described his process to a specific line of code.
Lavingia has nearly 15 years of experience as a software engineer and entrepreneur, but has no formal training in AI. He worked easily on Pinterest before starting Gumroad, a small e-commerce company that nearly collapsed in 2015. According to his personal blog post, Lavingia was floating the company by “replacing all manual processes with automated processes.”
Sahil Lavingia at his office in Brooklyn Credit: Ben Skiller for Propobrica
Lavingia didn’t have much time to immerse herself in how she handled veteran care between the time the VA wrote the tool March 17th until the next day. However, his experience with his own company was in line with the direction of the Trump administration. This has accepted the use of AI through government, streamlining operations and saving money.
Lavingia said the quick timeline for Trump’s February executive order was given to institutions for 30 days to complete contracts and grant reviews, and was too short to do the job manually. “That’s not possible – you have 90,000 contracts,” he said. “Unless you’ve written the code, but that’s not possible in practice.”
Old-fashioned, Lavingia said she completed the first version of the contract driving tool on the second day of work. He told Propublica that he downloaded his VA contract to his laptop for analysis during the first week.
VA spokesman Pete Casperowitz praised Doge’s work in a statement to ProPublica. “As far as we know, this kind of review has never been done before, but we are pleased to set this common-sense precedent,” he said.
He said the VA has looked at all 76,000 contracts to ensure each will benefit veterans and is making good use of taxpayer money. The decision to cancel or reduce the size of the contract comes after multiple reviews by VA employees, including agency contract experts and senior staff, he writes.
Kasperowicz said the VA would not cancel the contract for work to serve veterans or that the agency would not be able to do its own without a contingency plan. He added that contracts “have the ability of a VA to carry out its own, including waste, duplication, or services” will normally terminate.
Trump officials say they are working towards a “goal” of cutting around half a million workers from the VA’s workforce. Most employees work in one of the VA’s 170 hospitals and one of nearly 1,200 clinics.
The VA says it will avoid cutting contracts that directly affect care from fear that it will harm veterans. Propublica recently reported that relatively small cuts in government agencies have already put veteran care at risk.
The VA suggests that Lavingia’s code is a plan, but also does not explain plans to move services within the company while reducing staff.
read more
To help the public understand the flaws in Doge’s AI use, we’ve broken down how Lavingia’s rapid production works in another story.
Within AI, Doge prompts are used to “munk” contracts related to veteran health
Many people within the VA told Propublica that the process of considering a contract was very unclear and they couldn’t even see who made the ultimate decision to kill a particular contract. After the “mushroom” script selected a list of contracts, Lavingia said it would pass it to others who decide what to cancel and what to keep. He said the contract ended “without human reviews.”
“I just delivered it [list of contracts] To the VA employees,” he said.
VA staff told Propublica when Doge identified a contract that would be cancelled earlier this year, it sometimes doesn’t give employees time to justify maintaining services before Lavingia is brought in. One recalls that he was given only a few hours. The staff asked not to name them as they were afraid of losing their job to speak to reporters.
According to one internal email prior to Lavingia’s AI analysis, staff members had to respond with no more than 255 characters. This was shy about the 280 character limit on Musk’s X social media platform.
VA Mail tells staff that Doge covers the contracts that it must be limited to 255 characters. Credit: Get by Propublica
When he began his contract analysis for Doge, Lavingia said he was facing technical limitations. According to experts consulted by ProPublica, at least some of the errors generated by his code can be tracked to using older versions of the OpenAI models available through the VA. This model cannot solve complex tasks.
Furthermore, the fundamental instructions of the tool were deeply flawed. According to the record, Lavingia programms the AI system to make complex judgments based on the first few pages of each contract (the first 2,500 words), with only sparse summary information.
“AI is absolutely the wrong tool,” said Waldo Jaquith, a former Obama appointee who has contracts with the Department of Treasury. “AI gives you a compelling, look-up answer that’s often wrong. There needs to be people who are the job to do this.”
Lavingia’s prompts did not include the context of how the VA would work, which contracts were essential, or contracts required by federal law. This has led AI to determine the core parts of the agency’s own contract procurement system.
At the heart of Lavingia’s prompt is to eliminate contracts associated with “direct patient care.”
Next, based on these criteria, we assess whether the contract is “muddy.”
…
– Level 0: Direct patient care (e.g. bedside nurse) – Munch not possible
– Level 1: Necessary consultants not surrounded – Can’t be cleaned up
– Level 2+: Multiple layers removed from Veteran Care – One-handed
– Contracts (DEI) initiatives related to “diversity, equity, and inclusion” – Munchable
– A service that can be easily replaced by an in-house W2 employee – Munchable
Experts say that such an approach doesn’t address the reality that the work that doctors and nurses have done to care for hospital veterans is possible only with important support around them.
Lavingia’s system used AI to extract details such as contract numbers and “total contract value.” This caused a avoidable error in which AI returned the wrong dollar value when multiples were found in the contract. Experts said the correct information is readily available from public databases.
Lavingia admitted that errors came from this approach, but stated that these errors were later fixed by VA staff.
In late March, Lavingia published a version of the “munkable” script on his GitHub account, inviting others to use and improve it, he told Propublica. “If the whole federal government could use this script and anyone in the public could see this is VA thinking of cutting contracts, that would have been cool.”
According to a post on his blog, this was done with Musk’s approval before he left Doge. “When he asked the room about improving Doge’s public perception, I asked if the code I was writing could be open sourced,” Lavingia said. “He said ‘Yes.’ That was in line with Doge’s goal of maximum transparency. ”
That openness may have ultimately led to Rabinvia’s firing. Lavingia confirmed that she has finished with Doge after interviewing Fast Company Magazine about her work with the department. A VA spokesman declined to comment on Lavingia’s firing.
VA officials have refused to say whether they will continue to use the “risqué” tool in the future. But the administration can deploy AI to help institutions replace their employees. Documents previously obtained by Propublica show that Doge officials, proposed in March, will consolidate their benefits claims division by relying more on AI.
And government contractors are paying attention. After Lavingia posted the code, he said he heard it from people trying to understand how to flow money.
“I got some DMs from the VA contractor who asked me when I saw this code,” he said. “They were trying to make sure their contracts weren’t cut, or they learned why they were cut.
“After all, humans end a contract, but it’s helpful to see Doge, Trump, or the head of an agency thinking about how to munch on them. Transparency is good.”
If there is information about AI misuse or abuse within government agencies, Brandon Roberts is an investigative journalist for the News Applications Team and has extensive experience analyzing it using artificial intelligence. He can be contacted by signal @brandonrobertz.01 or email [email protected].
If you have any information about VAs you need to know, please contact reporter Vernal Coleman by signal, Vcoleman91.99, or email. [email protected]and via Eric Umansky on Signal, Ericumansky.04, or via email; [email protected].
