Taxpayer Behavior in the Age of AI: A Field Experiment on Property Tax Appeals (Holz, Perez-Truglia, Simon & Zentner, 2026)

Summary

"Taxpayer Behavior in the Age of AI: A Field Experiment on Property Tax Appeals," by Justin E. Holz (University of Michigan), Ricardo Perez-Truglia (UCLA), Andrew Simon (University of Virginia), and Alejandro Zentner (UT Dallas), is NBER Working Paper 35632 (August 2026, revised). It is a randomized field experiment on whether an AI chatbot changes who successfully appeals a property-tax assessment — a demand-side companion to the wiki's existing coverage of who exploits the appeals process, and a timely 2026 data point on AI-assisted tax administration.

The Experiment

The authors recruited 645 households in Dallas County, Texas (2026 average property tax bill ~$7,900) who visited a website offering personalized property information, filing instructions, and supporting evidence for a property-tax appeal, after 45,200 postcard invitations were mailed. Half the households were randomly assigned an otherwise identical version of the site that also included an AI chatbot capable of answering questions and tailoring guidance to the household's specific case; the chatbot arm was further sub-randomized between a machine-toned "TaxBot" persona and a more personal "Marina" persona.

The Finding: A Large Effect, Unevenly Distributed

Chatbot take-up was high — 78% of assigned households initiated a conversation. Access to the chatbot raised the probability of filing an appeal without a human agent from 41.4% to 50.5%, a 9.1-percentage-point increase (p=0.021; 10.5pp with controls, p=0.006) — nearly double the 4.98pp effect of a prior mailed step-by-step guide intervention (Nathan et al. 2025) in the same setting. Clickstream and conversation-transcript analysis suggests the chatbot's value was less about answering mechanical questions and more about helping households exercise judgment — for example, interpreting which comparable nearby homes best supported their specific appeal case; chatbot-assigned households spent more time on the site (17.3 vs 12.1 minutes) and shifted toward uploading custom evidence rather than relying solely on pre-generated comparison reports.

The paper's central caution: the filing-rate effect was smaller among less-educated homeowners, homeowners with lower-valued properties, and racial or ethnic minority homeowners — a pattern consistent, though not every dimension individually reached statistical significance, across all three dimensions, which the authors read as suggestive evidence that chatbot assistance widened rather than narrowed existing disparities in appeals access. A companion survey cited in the paper found 40.0% of Texas homeowners who filed an appeal unassisted had already used a general-purpose AI tool like ChatGPT.

Relation to the Georgist Case

The wiki already documents unequal access to the property-tax appeals process as a channel through which sophisticated, well-resourced actors extract disproportionate value — see McCanless's Memphis finding that institutional single-family-rental landlords appeal at roughly six times the rate of homeowners. This paper adds the demand side of the same problem from a different angle: even a well-designed intervention explicitly built to lower the cost of appealing — free, personalized, AI-assisted — can still reproduce or widen the underlying inequality if the technology's marginal value is itself unevenly distributed by education, income (proxied by home value), and race. It is a caution for any assessment-quality reform, including LVT implementations, that assumes making appeals cheaper and more accessible is sufficient to equalize who actually uses them.

Nuances and Limits

• Small analysis sample (645 households) drawn from 45,200 invitations, and the subgroup-disparity findings are, in the authors' own words, "large in magnitude but imprecisely estimated and therefore not always statistically significant" individually — the paper's claim rests on consistency across three separate dimensions (education, home value, race/ethnicity) rather than any single subgroup result being definitive on its own.
• A single US county and tax system. Texas property-tax appeals (protests) are a specific administrative process; the finding that AI assistance can widen access disparities may or may not generalize to other appeals systems or other AI-assisted government processes.
• A-claim. Full text (introduction, experimental design, headline results, and mechanism sections) read directly from the NBER working paper PDF, not reconstructed from an abstract.

Bears On

• Research: McCanless: Property Tax Appeals and Rentier Capitalism — the supply-side companion finding: which owners already exploit the appeals process, versus this paper's demand-side finding on who benefits from a tool meant to broaden access.
• Objection: Land value can't be assessed accurately — bears on assessment-and-appeals administrative design generally.

See Also

• McCanless: Constructing Public Rent — Property Tax Appeals and Rentier Capitalism
• Narrative: Land and the Black-White Wealth Gap
• Objection: Land Value Can't Be Assessed Accurately
• Mass Appraisal Methods

Sources

1. Justin E. Holz, Ricardo Perez-Truglia, Andrew Simon & Alejandro Zentner (2026), "Taxpayer Behavior in the Age of AI: A Field Experiment on Property Tax Appeals," NBER Working Paper 35632, August 2026 (revised). Preregistered: AEA RCT Registry AEARCTR-0018430. nber.org/papers/w35632 — full text (introduction, experimental design, headline results, and mechanism sections) read directly from the working-paper PDF, 2026-09-02 — used for the Dallas County experimental design, the 41.4%/50.5% filing-rate finding, the comparison to Nathan et al. (2025), the mechanism evidence (time-on-site, feature use, transcript analysis), and the subgroup-disparity finding (A-claim; full text read and cross-checked against a second independent extraction of the same PDF).