Over the course of the last week and a half, two government technology officials — one an artificial intelligence engineer and the second a former DOGE affiliate — proposed a flurry of changes to the code that underpins at least 600 government websites.
A government contractor posted a plea for “some communication please” from the two officials.
Artificial intelligence “appears to be the driver,” the contractor wrote last week on the federal web design system’s GitHub page — a developer platform used to share and manage code. “It’s scaring” the federal agencies who use the web design system to build their own sites, the contractor wrote.
“An accessibility proven design system,” they continued, “seems to be taking a backseat to the AI push.”
Trending
The official communication that followed only deepened fears about the extent of the Trump administration’s use of AI for government services. Ryan Parker, the AI engineer from the Treasury Department who is newly in charge of the government’s web design system, responded to the plea by saying he plans to fully automate accessibility testing on the government’s web design systems.
In other words, the Trump administration wants AI to conduct the tests that ensure that blind people and those with other disabilities can access the government websites that Americans use to sign up for benefits or find other critical information.
The outcry was immediate. The government’s own web page on Section 508, the legal requirement for government technology to be accessible, says that manual testing is necessary because some accessibility standards are “highly context-dependent and require human judgment to determine conformance.”
“No one’s been able to” automate accessibility testing, including large accessibility companies, one government technology employee told The Washington Sun. Changes made with “AI and the confidence of a coked-up teenager” could make the system — and websites that rely on it — less accessible, they said.
The idea is “dangerous,” said another government technology employee.
Both employees requested anonymity for fear of retribution. But backlash was apparent on the government GitHub page, where commenters raised concerns about removing human-run testing. In response, Parker and the former DOGE affiliate, Treasury Department Chief Information Officer Sam Corcos, appeared to be updating posts in real time and deleting comments that raised concerns.
Parker has edited his intentions for AI-powered accessibility several times, changing it from a goal to “automate the full range of accessibility tests needed” for the system to “automate as much accessibility testing as possible.” He also added — and removed — references to falling back on human testing as needed.
In another instance, Parker wrote: “We do not yet know whether every assessment can be automated reliably; demonstrating that reliability is part of the work.”
Laura Gerhardt, a civic technologist and engineer who previously implemented the web design system at several agencies, left comments on the system’s GitHub detailing concerns about its overhaul apparently being led by a single engineer.
Corcos responded with a single word — “wrong” — at 4:32 a.m. Wednesday, according to screenshots viewed by The Sun. Since then, Gerhardt’s comments have been deleted and her ability to make further ones has been blocked, Gerhardt told The Sun.
Amanda Kelly, a former General Services Administration employee, also left a comment asking why other notes were being deleted. Those comments, too, were deleted, she confirmed to The Sun.
“The U.S. Web Design System is responsibly integrating AI tools, human review, and community input to improve accessibility,” a GSA spokesperson told The Sun. “Incorporating AI will result in identifying barriers earlier, expanding testing coverage, and improving accessibility while we validate reliability.”
President Donald Trump has made reshaping the way government websites look a priority; last year, he ordered his new National Design Studio — staffed by many prominent former DOGE associates — to overhaul the U.S. Web Design System. Whether the National Design Studio is involved currently is unclear. But the head of that design team has promised to overhaul the government’s entire online presence.
Many of those early efforts of the design studio have run into accessibility issues, as The Washington Sun has previously reported.
For over a decade, the General Services Administration has maintained the U.S. Web Design System. In recent weeks, that responsibility has been shifted from a small team at GSA and open-source contributors to two people: Corcos and Parker.
Corcos was installed in several roles at GSA last month. Since then, the sole full-time federal employee tasked with maintaining the design system was put on administrative leave. Open-source contributors — added after the main contract for the system was cut earlier this year — have been booted.
Matt Henry, who used to work on the system at GSA until last fall, is now encouraging those who use the system to use the “pre-slop” version archived before all these changes were made.
Now, the prospect of AI taking over accessibility testing is magnifying those worries.
Automated tools can’t test if a person can understand and use a website, digital accessibility expert Nicolas Steenhout wrote in a LinkedIn blog about the idea. The tools can technically run a screen reader — software that helps blind people use websites by reading the text on them out loud or offering a braille display— but they can’t see if those programs work for real people.
“I have never seen an automated testing tool that can capture the full range of accessibility issues,” said Beth DeConinck, an accessibility specialist who’s been in the field for a decade and previously worked as a government contractor.
Although she uses automated tools, most of her work is manual, DeConinck said, and she frequently uses the federal design system as a gut check for her work at a private company.
“This is a valuable resource for everybody, and I would hate for it to be ruined,” she said.
Author
