AI has now arrived at the Treasury Department.
Sam Corcos, a former startup leader and Department of Government Efficiency affiliate now serving as chief information officer at the Treasury Department, appears to have approved spending at least $1.5 million on up to 3,000 licenses for ChatGPT, the OpenAI platform, federal spending records show. The agency has obligations to spend $1.5 million on the services, and has already outlaid more than $500,000 for the technology, those records show.
Fast Company obtained a user agreement showing that Treasury is allowing employees to use ChatGPT for “authorized” mission purposes. Such purposes include using the technology, in certain circumstances, with what’s known as “controlled unclassified information,” a government designation that’s given to information that isn’t classified, but still requires some safeguarding. The expanded use of the tool comes amid growing pressure on federal agencies to adopt artificial intelligence systems, which advocates say can increase efficiency and cut down on excess bureaucracy.
In this case, the rules laid out in the user agreement include strong limits on how AI systems might be used—particularly, for example, with regards to personally identifiable information, market-sensitive economic information, and federal tax data. The rules also forbid Treasury staffers from trying to tamper with or evade an AI chatbot’s security measures without express authorization. Employees aren’t supposed to use the output of an AI system without a human reviewing that work, or obfuscate the role AI played in making a particular product, according to the user agreement. A violation of these rules could lead to someone being fired, the agreement states.
One former Treasury official said department staff are probably using the tech on heavy lifting for tasks that would normally take a long time. Tony Arcadi, the official that Sam Corcos replaced, tells Fast Company that there were myriad use cases that could benefit from the technology, including automating administrative work. “Done correctly and with robust controls, LLMs could be a force multiplier for intelligence, operations, finance, enforcement, and public engagement,” he says. The agency had previously invested in a smaller cache of ChatGPT licenses.
The Treasury Department and OpenAI did not respond to a request for comment. Still, in September, the agency released a compliance plan focused on promoting the use of AI, as well as a strategy spelling out its approach to the technology.
Amid the move to speed up the use of AI throughout the government, including the military’s new GenAI.mil tool, there’s still the serious risk of government officials putting too much faith in the far from faultless technology.
For example, it seems like a recent report from the Department of Health and Human services may have been created using artificial intelligence—and included fake citations. Federal clerks have used ChatGPT and Perplexity and have ended up including misquotes and other errors in documents.