AI Use in U.S. Local Government Communications: Prevalence and Reach on Social Media

Generative AI has made polished social-media copy nearly free to produce, and a plausible worry is that cash-strapped local governments adopting it will pay for the savings with reach: residents may read machine-drafted posts as less authentic and engage with them less. We test that worry directly, within account, across 767,131 Facebook posts from roughly 4,900 U.S. city and county governments in all 50 states and DC. After we control for length, theme, writing register, and formality, posts that read as more machine-like draw 2–3% fewer likes and comments per standard deviation of machine-likeness, and no reliably different share count. The penalty is real but small, which means local governments can likely use AI to help reach residents online at much lower staff cost, without paying a major price in engagement. Reaching that conclusion first requires measuring how much local-government writing is already machine-generated, which is harder than it looks: the short, formal posts typical of government social media are the worst case for machine-text detection, and both a zero-shot detector (Binoculars) and a lexical “AI-style” checklist mistake ordinary formal human writing for AI output. We therefore calibrate the detector against a human baseline learned from 868 pre-ChatGPT posts, which the raw detector already flags at 3.8%, about the rate it flags posts published since (3.7%). After calibration, the flagged share of later posts exceeds that baseline by less than one percentage point, not statistically distinguishable from zero given the size of the baseline sample. The lexical “AI-style” proxy is similarly unreliable: it flags 39% of posts but also 22% of posts written before ChatGPT existed. At the account level, a small group of offices has high detector flag rates, partly reflecting short posts; population, revenue, and staffing explain little of the variation. The broader measurement lesson is that estimates of machine-written content in short, formal text need a human baseline from the same kind of text.