What happens when research institutions move from talking about AI to actually building with it?
You start confronting the practical stuff: how to find the right expertise, how to keep AI from confidently making things up, how to make research accessible to AI systems, and how much staff time these projects actually require.
Teal Media brought together digital leaders from The Brookings Institution, KFF, the National Council on Teacher Quality, and Pew Research Center to share what they’re learning as they put AI into practice.
You Don’t Have to Be an AI Expert to Get Started
Ashley Kincaid, Chief of Strategy & Operations at NCTQ, offered a practical approach to finding and vetting AI expertise.
Start by asking your network for recommendations, including people in adjacent fields who may broaden your view of the options. Once proposals come in, AI can help compare where they agree and differ, then draft a preliminary scope of work based on what you’ve learned.
And don’t stop vetting after you’ve selected a partner. Bring in unaffiliated experts to review the work and provide an independent perspective.
Teach AI When to Say “I Don’t Know”
Lisa Petrovich, Brookings’ Director of Website Strategy and Operations, is preparing to launch a tool with a deliberately narrow job. It helps people explore a database that tracks countries’ financial ties with the rest of the world back to 1970. The bot answers questions in plain English and builds charts from the underlying numbers.
Brookings needed to reduce the risk of the bot inventing answers, misinterpreting the data, or presenting outside research as if it were its own. So it trained the bot to say “I don’t know” rather than guess.
Don’t Let AI Decide What Your Data Means
That’s the advice from Melinda Wuellner, Senior Director, Digital and Creative at KFF.
Some of KFF’s most valuable findings live in its data visualizations, where they aren’t as easily discoverable by AI search engines. Written descriptions of those charts, known as alt text, can make those insights accessible, but describing a complex chart without oversimplifying it takes time.
So KFF developed guidelines and trained the research teams on the value of alt text for AI discoverability, in addition to accessibility. They also developed a detailed prompt intended to help the research teams create alt text for complex data viz quickly. KFF’s experts review the drafts before they’re published. The process keeps experts in control without asking them to write every description from scratch.
Build for AI Search and Find Uses You Didn’t Expect
At Pew Research Center, Head of Engineering Seth Rubenstein focused on making the organization’s research easier for large language models (LLMs), the technology behind tools like ChatGPT, to read.
Pew invested heavily in creating a behind-the-scenes version of its website content in Markdown, a simple text-based format that gives LLMs a cleaner way to understand and retrieve every paragraph, table, and chart.
Making the content easier for LLMs to use opened up other possibilities, too. Pew can now use that same content to draft social posts or turn a 20-minute read into a four-minute audio overview, a format it didn’t previously produce.
Pew is making the code open source so other nonprofits and research organizations can build on what it created. Check it out on PRCs GitHub: http://www.pewresearch.org/opensource
Budget for the Work Behind the AI
The AI tool isn’t the only thing you’re paying for. Kincaid pointed to the J-curve of AI productivity: organizations may see an initial dip as staff spend time implementing, testing, and learning to work with AI before the benefits show up. That staff time belongs in the budget, too.
NCTQ saw that firsthand while developing two AI tools, one to code teacher policy data and another to answer questions about that data. Testing involved six people, and the principal researcher spent more than a third of her time on the project. Altogether, the work consumed the equivalent of 1.4 full-time staff members.
At KFF, a multimedia staff member spent time fine-tuning a text-to-speech tool on nuances like intonation. The AI sometimes struggled to pronounce authors’ names, so KFF ultimately decided to record the authors saying their own names.
And the costs don’t stop at launch. Petrovich cautioned organizations to budget for hosting, usage, security, and legal review, plus the staff responsible for maintaining the tool and checking its accuracy over time.
Ultimately, AI Is a Tool, Not a Strategy
None of these organizations approached AI as a simple plug-and-play solution. They started with a specific problem, kept experts involved, built in ways for the technology to fail safely, and accounted for the people and infrastructure required to maintain it.
That’s perhaps the biggest lesson for nonprofits considering AI: the technology may be getting easier to use, but using it well still requires strategy, expertise, and ongoing investment. The organizations that get the most from AI won’t necessarily be the ones using the most advanced tools. They’ll be the ones that are thoughtful about where AI can actually help, where humans need to stay in control, and what it takes to make the whole thing work.
