EPA piloting AI on ‘everything,’ but experts still needed for highest levels

The Environmental Protection Agency has run artificial intelligence pilots on “everything,” but its chief information officer only wants subject matter experts to be using the technology at a high level.

During a FedInsider webinar last week, CIO Carter Farmer said while the agency has piloted AI to review public comments and analyze large scientific datasets, he still wants experts to review outputs.

“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,” he said. “If you’re not an expert at that, validating those outputs is very hard.”

Another reason why using AI can require more experience: “Prompt engineering is a real skill,” Farmer said, and proper use is rarely plug-and-play.

“The amount of people who are just throwing generic Google search questions into AI and expecting it to come out with this amazing output, it happens a lot,” he said. “Having to learn how AI works — and how the back end of it actually works — is very helpful in how to think about how you should be using this tool.”

But the agency’s daily use of AI is less high-stakes. 

Farmer said the EPA’s biggest focus currently is using AI for “low-level” or “low-risk functions” like email drafting and creating presentations. The agency has seen a “huge” jump in daily AI usage because of fewer guardrails than before, when the technology was first starting out, he said.

“If you don’t give people the tools they need, you risk them going outside the bounds of what is available,” he said. “We don’t want people to go use some public tool — we have very strict guardrails around their AI tools internally. So 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.”

Regardless of the AI use level, Farmer said it will produce errors, and zero tolerance for error is unrealistic. Expecting 100% accuracy and waiting too long means “nothing is going to be good enough, so you’re going to spend a lot of money and effort and getting nowhere,” he said. 

“That’s never a feat we’ve ever accomplished in the history of humanity,” he said. “Holding AI to a level higher than us is not a way to move forward. We have to be able to take risks, understand within certain guardrails what is acceptable, what’s not, and be able to move forward.” 

The hardest challenge of it all is “change management” or getting the workforce up to speed on AI, Farmer said, because AI education and implementation requires “a lot of time, effort and resources.”

“We tell people: If you’re not confident or have reservations about AI, the best way to approach something you don’t understand or trust is to educate yourself on that thing. Especially if you have negative feelings toward it — even more a reason why you should educate yourself,” he said. “Even if it doesn’t change your mind, you’ll have a better argument at the end of the day.”

The EPA has a community of practice and a Teams chat where people can discuss use cases and help fast track it for approval.

Farmer said the EPA is currently finalizing a “really solid” AI acquisition policy, as it is concerned about data leakage and the risk of EPA data being used to train models. The agency has been using AI to try and understand the AI’s terms and conditions. “The best way to learn about AI is to ask AI,” Farmer said.

The EPA is also “leaning into pretty heavily” on the General Services Administration’s AI offerings, Farmer noted.

“That is really helping get us across the starting line — maybe not the finish line of getting through some of those hurdles from an acquisition standpoint, a security standpoint and not have us start at zero,” he said.

Farmer said the Federal CIO Council is reusing work that is already being done across government on AI, and advised others to lean into older versions of AI if appropriate.

“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. “There is no need for you to go out there and be a trailblazer on something that has already been done 20 times by different agencies.”