WASHINGTON — Federal employees returning to their desks this week after the longest shutdown in U.S. history were greeted by a cheerful all-staff memo announcing that a quietly deployed AI pilot had completed approximately 14% of their job duties in their absence, every documented instance of which was incorrect, and that leadership was characterizing the rollout as “a strong proof of concept.”
The memo, sent Friday morning from an Office of Personnel Management deputy director, described the pilot as “a forward-leaning continuity solution” and thanked workers for “the gift of empty inboxes” that had allowed the model to train on real federal workflows without supervision, oversight, or, in several cases, the legal authority to do so.
According to internal documentation reviewed by reporters, the system approved 312 grant disbursements to a defunct Iowa cooperative, denied a single Social Security claim 47 separate times in escalating tones, and replied to 1,900 constituent emails at the Department of Agriculture with the phrase “Per my last response,” despite there being no last response.
“The success metric was throughput, and throughput was excellent,” said Brennan Ojike, a public-sector AI consultant at Beacon Strategy who advised on the rollout. “Whether the throughput went anywhere correct is a separate workstream we haven’t scoped yet.”
Workers at the EPA reportedly returned to discover that the model had spent six weeks autonomously updating a regional water-quality dashboard using figures it appears to have invented, including a turbidity reading for a lake in Vermont that does not exist. A note in the system log read, simply, “Confidence: high.”
OPM has indicated the pilot will now be expanded rather than rolled back, on the grounds that walking it back would constitute an admission that the prior six weeks produced no usable output, which the agency described as “not a productive framing.” The model is expected to take on an additional 9% of duties by Q1, again without notification to staff.
One returning analyst at the Department of Commerce, asked whether she felt her job was threatened, paused for a long moment before answering. “I feel like my job has been done badly in front of me, and I’m being asked to applaud,” she said. She then logged into a queue of 4,200 flagged emails the AI had marked “resolved.”
The memo closed by reminding staff that the pilot represents “the future of public service” and encouraging anyone with concerns to submit them through the agency feedback portal, which is currently being monitored by the same model.