EPA CIO: AI pilots span agency, but experts must validate outputs

The Environmental Protection Agency is experimenting with artificial intelligence across nearly every function, but its chief information officer insists that only domain experts should handle the technology's most consequential applications.

CIO Carter Farmer told a FedInsider webinar that while EPA has piloted AI for tasks ranging from reviewing public comments to analyzing large scientific datasets, the agency requires subject matter experts to validate all outputs from high-stakes work.

"Something we tell our staff quite regularly is if you're not an expert in the subject matter you're using AI for, you probably shouldn't be using AI because it can be very convincingly wrong," Farmer said, according to FedScoop, which first reported his remarks. "If you're not an expert at that, validating those outputs is very hard."

Farmer added that effective AI use requires genuine skill in prompt engineering—not simply typing search-engine-style questions and expecting quality results. Understanding how AI systems work on the backend helps users frame requests more effectively, he noted.

Low-risk applications drive daily adoption

The bulk of EPA's current AI usage centers on low-stakes functions: drafting emails, creating presentations, and similar administrative tasks. Farmer said the agency has seen a significant jump in daily AI adoption after relaxing early guardrails, though strict internal controls remain in place to prevent staff from turning to unsecured public tools.

"If you don't give people the tools they need, you risk them going outside the bounds of what is available," he explained. "We want people to have the tools they need to do their jobs, but also put some guardrails around what's safe to use and what's not safe to use."

Why it matters

Farmer's approach reflects a pragmatic middle path for federal AI adoption: enable broad experimentation for routine work while reserving mission-critical applications for trained specialists. His acknowledgment that AI will produce errors—and that zero-tolerance expectations are counterproductive—signals a maturation in how agencies think about deploying the technology at scale. The strategy also addresses a common federal challenge: preventing shadow IT by providing approved tools that meet security requirements.

Realistic error tolerance and change management

Farmer emphasized that expecting 100% accuracy from AI is unrealistic and counterproductive. Holding AI to standards higher than human performance will stall progress and waste resources, he argued.

The hardest challenge, according to Farmer, is change management—getting the workforce educated and comfortable with AI tools. He encourages skeptical staff to learn about the technology, noting that education provides either confidence or better-informed criticism.

EPA has established a community of practice and a Teams channel where employees can discuss use cases and expedite approvals. The agency is also finalizing an AI acquisition policy focused on preventing data leakage and ensuring EPA information isn't used to train external models. Farmer said EPA has used AI itself to parse vendors' terms and conditions.

The agency is leaning heavily on the General Services Administration's AI offerings to clear initial security and acquisition hurdles, and Farmer advised other agencies to reuse existing solutions rather than starting from scratch.

"Anything you're probably trying to solve in your agency has probably already been solved 10 to 20 thousand times everywhere else in the world," he said.

These details were first reported by K. Sophie Will at FedScoop.