Large language models are a tool, not a replacement, for utility communications

Many industries have turned to large language models, or LLMs, to assist with communications. Models like ChatGPT, Gemini, and Claude can draft emails, social media content, and longer narratives with speed, authority, and ease.

But what may sound authoritative may not be accurate (because LLMs don’t guarantee accuracy) — and in the world of water, where public trust is paramount, that presents an important caveat for utilities.

“Whatever AI provides you must be validated and edited by a human. Just because it sounds fluent does not mean it is factual,” said Lindsay Hughes, founder of Tapped In Public Relations, during an AWWA panel discussion that shared tips for using AI assistance in water utility communications.

“It could sound confident and it could sound ready to post in public or send out to customers, but unless you verify that it’s accurate and it has your voice, it’s still in draft form,” she said during the webinar, titled “Applying AI in Water Utility Communications,” where drew more than 800 registrants.

To improve the quality of the outputs from LLMs, Hughes recommended that communicators spend a little upfront time to be as specific as possible about the inputs. Tell the LLM exactly what the situation is, what the task is, and who the audience is for the intended message, she said.

Communicators can also personalize the settings in an LLM to suit their utility’s brand and voice, said Rebecca Emery, founder of Seacoast AI. She recommended that communications professionals share the utility’s brand guide and any past communication that will help the LLM or agent understand the utility’s tone. Anytime something new is produced and shared publicly, be sure to share it with the LLM so its outputs stay aligned with the utility’s talking points, Emery added.

It also helps utility communicators combat what she called the “visibility gap,” which happens when LLMs provide answers to community members’ questions by culling information online. Sometimes, those responses include utility information; sometimes, they don’t.

“When AI answers a question [in a Google search] about your municipality or something happening in the community, we don’t typically get to know who is asking, what tools they’re using, what they’re asking, and how AI is answering on our behalf — whether those responses are accurate and trustworthy — and whether the person is takes any further action from there,” Emery said. “That makes it more challenging for us in communications. … But we can control the consistency of the online content that AI sees and reuses.”

Cayce Myers, communications professor at Virginia Tech, emphasized this point, noting that content creation has evolved from search engine optimization, or SEO, to generative engine optimization, or GEO — which refers to the AI summaries that search engines create or the responses people receive when they use LLMs. (This is also known as answer engine optimization, or AEO.)

“We as communicators and content producers have to understand this new reality,” Myers said. “There’s risk in this because there’s not total control” of the message.

He said Q&As, reviews, storytelling, and earned media resonate most with LLMs. “AI pulls so much from the news,” he added.

For more recommendations on LLM and AI use, including internal policies to guide it, view the webinar recording here.