Artificial Intelligence and Tax Behavior: A Field Experiment

We mailed postcards to a random sample of households that pay property taxes and could legally reduce their taxes by filing an appeal.

The postcards directed recipients to a website we designed to help them decide whether to appeal and, if so, how to do it. Households were randomly assigned to different versions of the website: one provided only information, such as supporting evidence and a step-by-step guide, while another also included a chatbot. And within the chatbot condition, we randomize between a more human-like version ('Marina') and a less human-like version ('TaxBot').

2026-05-01

2026-05-15

The main outcome of interest is whether the household filed a tax appeal on its own; the 2026 filing deadline is May 15. For households assigned to the chatbot treatment, a key secondary outcome is chatbot usage, including whether they interacted with it at all and, if so, what questions they asked.

Given the nature of the setting and the timing of the protest process, we expect the intervention to affect primarily whether households file a direct protest. For this reason, we will focus on direct protests as the primary outcome. We also plan to use direct protests in prior years as falsification, or placebo, outcomes in the spirit of an event-study analysis.

The rich administrative data will also allow us to examine additional outcomes. For example, we can estimate effects not only on whether households file an appeal, but also on whether their appeal is successful. We may also be able to obtain information on the Opinion of Value (for households who file an appeal), which could be potentially an outcome of interest. For households assigned to the chatbot treatment, we can also analyze chat transcripts to study the content and dynamics of their interactions.

Some households are randomly selected to receive a postcard with a URL directing them to a website. Screenshots of a sample postcard and a sample website are attached to this preregistration.

Among these households, we randomize whether the website provides information only or also includes a chatbot. Our hypothesis is that the chatbot may affect the decision to file an appeal by helping households determine whether filing this year is worthwhile and by assisting them in navigating the appeal process. Among households assigned to the chatbot condition, we randomize between two versions of the chatbot ('Marina' and 'TaxBot'). These versions differ in that one appears more human-like and is also prompted to interact in a more human-like manner.

An important feature of the design is that some households will need to be dropped from the study sample for reasons we cannot observe ex ante. Most importantly, the window to begin filing tax appeals opens on April 15 and closes on May 15. Because of delays in printing and delivering the postcards, however, households will begin receiving them only around the middle of that window (at the time of this preregistration, we do not yet know the exact delivery date). By then, at least two weeks will have passed, and a substantial share of households may already have submitted an appeal before receiving our postcard. Because their decision will already have been made, the website cannot plausibly affect it. For that reason, households that have already filed an appeal will be excluded from the analysis. We will also exclude other households as needed. For example, if administrative records indicate that (by the time of the postcard delivery) a household had already hired an agent to file on its behalf, we will exclude that household as well. We will exclude households whose postcards are returned to the sender. And some households may need to be excluded if their addresses are flagged as undeliverable or as having mail forwarding through the National Change of Address (NCOA) system.

There are several pretreatment characteristics that are likely to predict whether a household files an appeal. For example, households whose assessed values increased are probably more likely to appeal than households whose assessed values did not change. We may use these pretreatment characteristics both as control variables and to study treatment heterogeneity. We are also interested in additional heterogeneity analyses, such as whether the chatbot has different effects for higher- versus lower-income households or for Hispanic versus non-Hispanic households.

Randomization done in office by a computer

Household

No

61,474 households. However, as explained above in 'Experimental Design,' some of these households will need to be dropped ex post; for example, if they had already filed an appeal by the time our postcards were delivered.

61,474 households (again, some of these households will have to be excluded from the sample)

45,200 households are sent a postcard, while the remaining 16,274 households are not.

Among households that receive a postcard, we randomly assign them to one of two sub-treatments:
- 50% are assigned to a website without a chatbot.
- 50% are assigned to an otherwise identical website with a chatbot.

And among households assigned to the chatbot condition, we randomly assign them to one of two sub-treatments:
- 50% are assigned to a more human-like version ('Marina').
- 50% are assigned to a less human-like version ('Chatbot').