On May 29, 2025, the Department of the Treasury (Treasury) provided a briefing to Ways
and Means Committee staff and its Members’ personal staff to discuss the Department of
Government Efficiency’s (DOGE) work at Treasury. During this briefing, DOGE employees said
they were excited about the prospect of using AI to handle old technology systems. COBOL is a
legacy technology and people are no longer being trained to code in COBOL at school.
Therefore, the technology is becoming obsolete. Previously, efforts were made to modernize the
source code, but we were told the IRS is abandoning those efforts in favor of maintaining the
legacy systems.
Though DOGE stated they are “excited” to use AI to interact with COBOL, we would
like to better understand this approach to using AI, especially given the many known limitations
with AI conversion capabilities. A study released in July by Model Evaluation & Threat Research
found that AI’s ability to create source code is frequently inaccurate.1
often have to spend as much or more time checking and rewriting the code that AI systems
1See Are We in an AI Bubble? - The Atlantic, available at https://www.theatlantic.com/economy/archive/2025/09/aibubble-us-economy… (“The results of the March METR study, for example, were based on a “50 percent success rate,” meaning the AI system could reliably complete the task only half the time—making it essentially useless on its own.”).
Add new comment