Let's put AI to the test for voter research

One of my proudest professional accomplishments is serving on the original staff team that launched ATPE’s Teach the Vote in 2012. In the past 14 years, more than 1.2 million visitors have used the site to research candidates’ public education stances and stay up to date on the latest public education news.

I’ve traditionally relied on Teach the Vote and the League of Women Voters’ excellent website vote411.org to do my ballot research, but when a friend mentioned the other day she was using ChatGPT to prepare to vote, I started to wonder if could AI enhance my own research. I approach AI with a healthy dose of skepticism; I’ve seen plenty of hallucinations, and it concerns me that people might be led astray in their voting choices by bad AI or bad prompting. So, as a public service, I decided to conduct my own experiment and report the results here on Teach the Vote.

Developing a prompt to research November 2026 candidates on ChatGPT

At the moment, my current preferred AI tool is ChatGPT. I’ve dabbled in Claude and Perplexity, and I’m sure this would also work in Gemini and Copilot. But for the sake of transparency, my experiment was conducted in my personal paid ChatGPT Plus instance (given that this was my real voting research).

My process actually started outside ChatGPT. With the ease of using Teach the Vote, I feel very comfortable with my public education-related knowledge of the candidates for statewide races, my House and Senate districts, and my district on the State Board of Education. My knowledge gaps are related to the more local races, which I know are incredibly consequential but much harder to research. 

My initial prompt was this:

ChatGPT asked me to upload a sample ballot, directing me to a specific page on my county elections website, where I was able to download a PDF of the exact ballot I will receive (this may not be available for all of Texas’ 254 counties). 

I felt good about the prompt ChatGPT suggested. Here’s what I liked:

• The use of the phrase “factual, evenhanded”
• The guardrails related to identifying my exact ballot
• The directions not to ask for unnecessary personal information
• The direction to flag discrepancies and unresolved questions
• The mix of sources, including ATPE’s Teach the Vote
• Clear labeling of candidate-produced materials and the acknowledgement that nonpartisan organizations (such as ATPE) may still advocate particular policies
• Verification of relevant votes for incumbents
• Directions not to recommend whom I should vote for or assume my views

I uploaded my ballot, and ChatGPT then added ballot-specific instruction to its suggested prompt, grouping my ballot into seven batches for research. I added two directions to the final prompt:

• If a race is not specifically tied to public education, please explain what the office does, and share resources for further research.
• Please explain each ballot proposition and the pros and cons of each. Also list the supporters and opponents for each issue.

Then, I worked through each batch of candidates/ballot questions one by one. I could see that the information I was gathering was going to be useful, but it wasn’t in an easy-to-digest format, so I asked this: “One last question: Can you put the final prompt, the results of each batch, and a bibliography of cited sources into a PDF I can download and read?”

ChatGPT said yes, though I did have to copy and paste some of the chat results from early batches back into the tool. It took about 10 minutes to produce a 49-page downloadable PDF, which I scrolled through on my iPad and read while marking up a printout of the sample ballot PDF that I could take to the polls with me. (Remember: You can’t use mobile devices at polling places.) I feel more comfortable about voting in the down-ballot races than I ever have.

Lessons learned from this AI voting research experiment

• A 49-page PDF of voting research is, to put it mildly, “extra.” You may not want to be quite that comprehensive. It might be better to use this type of prompt for a smaller group of candidates, not your entire ballot, particularly if you’ve already researched candidates on Teach the Vote. (Remember: I was creating a proof of concept.)
• A free instance of an AI tool will likely not be able to produce this in-depth of research. My colleague Monty Exter had provided another interesting prompt, specifically focused on campaign finance information. I tried it in another free instance of ChatGPT; the tool got through about half of the research before shutting me off for the day. Exter did later find that a fair amount of the information his prompt sought can already be found on The Texas Tribune’s candidate guides. However, those do not cover all races. 
• This wasn’t a quick process. It took me about an hour to work through the prompting, download the ballot, run all the batches, and wait for the PDF creation. But it was a worthwhile endeavor—and a good reminder that it’s important to put time and care into any use of AI. The only way to achieve good outputs is to put effort into good inputs. As I like to say: Use your real brain first.

Will ChatGPT “endorse” candidates? Thankfully, no

After I had made my own ballot selections, I decided to see if ChatGPT would actually “endorse” candidates. The good news: It won’t!

I asked: “Based on what you know about me and the candidate research PDF you made me, please recommend candidates for each race in the candidate research brief. Please provide your justification for the recommendation.”

ChatGPT said:

I followed up with: “Why can’t you tailor voting advice?”

I then asked: “Could you evaluate/identify which candidate would be more commonly considered public education friendly?”

I uploaded the brief.

And I did get a nice chart of “public education-friendly” candidates. It aligned with the work my real brain had done. Take that for what you will.

Using ATPE’s AI assistant, Judy, to research candidates

ATPE’s own AI assistant, Judy, is trained on Teach the Vote content. While Teach the Vote is already extremely easy to use, Judy offers another way to access the candidate information compiled by ATPE. Here are some prompts that proved successful with Judy:

• Help me compare the candidates for lieutenant governor. 
• What are the major education stances of Dan Patrick and Vikki Goodwin?
• Please give me a side-by-side comparison of the candidates for governor in a short, take-to-the-ballot-box format.
• Who are the candidates for SBOE District 5, and what are their responses to the ATPE Candidate Survey?

If you’re an ATPE member, be sure to log in to the member instance of Judy to do your research you can enjoy unlimited use.

No matter how you choose to research candidates, remember that we must be PRESENT TO WIN at the polls if we want to see good results for public education in the 2027 legislative session. Go vote!