Campaign AI content rose in a U.S. election study, while disclosures were scarce

Measuring detectable AI, not proving authorship

The researchers assembled candidate information for federal, state and selected mayoral elections from 2023 to 2025. They collected campaign website material, posts from campaign X accounts and archived political emails, recovering some unavailable websites through the Internet Archive.

The resulting content dataset contained 29,996 text samples and 104,562 images. Those totals do not mean every candidate had usable material: the paper reports campaign magnitude scores for 7,144 candidates in its campaign-characteristics analysis before further exclusions.

For text, the team used the commercial detector Pangram, version 3, counting samples labeled either AI-generated or mixed human–AI as detected AI. They excluded non-English text and samples shorter than 70 tokens. For website detection, they sampled a contiguous 60% of each page’s body text to reduce costs.

For images, they used GPT-5.4-mini. On a balanced validation set of 340 images drawn from an earlier political-image study, its classifications agreed with the existing labels 78.3% of the time. Disagreements more often involved missing images previously labeled AI than flagging images previously labeled non-AI.

The authors also calculated a campaign-level “magnitude” score between zero and one. It combines text-length-weighted AI-assistance scores with image classifications, weighting the two content types by their numbers of samples. This is a constructed indicator—not a direct measurement of the fraction of campaign work performed by AI.

Rising detection, with differences across channels

Overall, the detectors flagged 7.6% of text samples and 6.1% of images. These are sample-level rates, not percentages of candidates using AI or messages seen by voters.

Average magnitude generally rose over the observation period, although statistically significant increases varied by office and interval. State House, State Senate and U.S. House campaigns showed significant increases from 2024 to 2025, but not from 2023 to 2024. Mayoral campaigns showed the reverse timing: a significant increase in the earlier interval only.

Channel comparisons also depended on which candidates were included. Among 130 candidates with material from websites, X and email, average magnitude was highest on X, followed by email and websites. A larger comparison of 1,535 candidates with both website and X material likewise found higher average magnitude on X. These within-candidate comparisons differ from simply comparing detection rates across all collected samples.

Few disclosures in the material reviewed

Searching website text and manually checking matches produced just one explicit AI-use disclosure. Separately, the authors reviewed 505 randomly sampled X posts containing images flagged as AI-generated. They found no AI disclosure in the images or accompanying captions, including platform labels.

That finding does not establish widespread lawbreaking. Many election laws concern deceptive synthetic media rather than all AI-assisted content, and the study did not determine whether individual messages met those legal definitions.

Campaign expense records offered another partial view. The researchers matched 129 candidates with available content to reports of AI-related spending. Such spending can support outreach or logistics rather than content generation, while free tools leave no purchase record. Financial disclosure and content detection therefore measure different things.

Challengers differed from incumbents; regulation results were mixed

A regression analysis reported lower AI magnitude for incumbents than challengers after accounting for other campaign characteristics. The estimated difference was 0.014 on the zero-to-one scale, about 20% lower relative to challengers. Smaller associations appeared for party, funding and office level. These observational relationships do not explain why campaigns chose particular tools.

A separate analysis accounting for state and election-year differences found no statistically detectable overall association between state AI election laws and content magnitude. Federal candidates showed a modest negative association, but state and municipal candidates did not. The authors explicitly caution that this does not establish that laws caused behavioral changes.

Important limits and preprint status

Detectors can both miss AI-assisted material and falsely flag human work, particularly with short texts or changing generation methods. Website images were collected only for candidates whose text had already been flagged, limiting interpretation of that image sample. Coverage was retrospective and incomplete; video and audio were not analyzed, and disclosure checks could miss dynamically displayed website material.

The manuscript also contains reporting inconsistencies: its main dataset is described as covering 2023–2025, while the regulation appendix describes 2023–2026 and includes an additional office category. These discrepancies warrant clarification.

This briefing concerns arXiv version 1, posted November 8, 2026, a preprint whose supplied record does not establish completed peer review. Its strongest contribution is a measurement approach and evidence of scarce visible disclosure—not a definitive census of campaign AI use. The authors propose better platform-data access and itemized AI spending reports; the study does not test whether those policies would improve transparency or trust.