Method

All statements from voter's pamphlets in the covered regions were analyzed using Pangram, a leading AI-detection model. Text was extracted through parsing the PDF or using a data API, depending on year and jurisdiction. When applicable, the statement and its rebuttal to the other side were analyzed together. Statements with fewer than 50 words were not analyzed. The raw score from Pangram was saved and used to generate the site.

The Control Set

The same detector was run against voter pamphlets from earlier years before ChatGPT existed. Voter pamphlets from three Washington counties (King, Pierce, Snohomish) as well as the statewide races were taken for four years of elections. For 2020 and 2022, both written before ChatGPT's release, none of the 479 statements were flagged as using AI. 2024 showed mild usage and 2026 showed a considerable leap. This shows that campaign statements do not seem to show a higher false positive rate than Pangram's baseline. That does not mean that 2024 and 2026 don't contain false positives.

Caveats

This site presents data from Pangram. We have softened the wording of the verdict to avoid implying it is absolute proof (e.g., "AI-generated" becomes "Appears AI"). Our work is limited to merging a dataset with a model and presenting the results in an easy-to-digest format. The model is less accurate on short text, which is the case for many campaign statements. Also, the model may be confused by the formal text that is common in candidate statements or by non-native speakers. Currently only Washington state has been analyzed.

Contact

If you have any questions or comments, email ballot@dataopen.org.