Consumer Choices over AI Assistant Services

This study employs an online survey experiment targeting U.S. adults to examine preferences for hypothetical AI assistant services through a series of discrete choice tasks. The experiment uses two layers of randomization. First, participants are randomly assigned to choice environments that differ in (i) whether a public-sector AI assistant is available alongside private alternatives and (ii) how that public-sector option is described. Second, across tasks, the attributes of each AI service -- such as provider type, performance, access conditions, and monthly fees -- are randomized.

2025-12-11

2025-12-18

- Adoption propensity: A binary indicator for whether a participant selects any AI assistant or opts out (chooses "None") in each choice task.
- Provider choice shares: The distribution of choices across different provider types across experimental conditions.
- Market substitution effects: Changes in the choice shares of private providers when a public option is available in the choice set.

We use the data from the discrete choice experiment to construct several related endpoints. First, we measure overall market participation as the probability of choosing any AI assistant rather than the "None" option. Second, within choice sets where a public-sector provider is present, we measure its take-up rate and analyze how this probability varies with service attributes. Third, we summarize substitution patterns across provider types using observed choice shares. These results are then combined with a structural demand model to recover preference parameters and implied willingness-to-pay for specific service characteristics.

The study is a randomized survey experiment in a sample of U.S. adults.

- Screening: Participants first complete a short screening module that asks about prior use of, or interest in, AI assistants. Only those who indicate some prior use or interest are invited to proceed, so that the experiment focuses on individuals for whom AI assistant adoption decisions are relevant.

- Randomization: Eligible participants are randomly assigned to one of two experimental conditions that differ in the composition of the AI services shown in the choice tasks (with vs. without a public-sector provider), and, within the public-sector condition, to one of two framings of that provider.

- Choice tasks: Each respondent completes a series of discrete choice tasks (conjoint questions) in which they choose between several AI assistant profiles or an explicit "None" option. Attribute levels (e.g., price, performance, access conditions, provider type) vary across profiles and tasks according to a pre-generated experimental design.

- Survey module: A separate survey section collects covariates such as institutional trust, privacy attitudes, attitudes toward AI, fiscal attitudes, and demographics.

Not available

Randomization will be implemented within the online survey platform using built-in randomization routines and pre-generated design files:

- Respondents are randomly assigned with equal probability to one of two experimental conditions (with vs. without a public-sector provider in the choice set).

- Within the condition that includes a public-sector provider, respondents are further randomized with equal probability to one of two descriptions of that provider.

- For each experimental condition, we use pre-generated discrete choice designs in which task order, alternative positions, and attribute combinations are randomized ex ante subject to pre-specified constraints (e.g., no duplicate or strictly dominated profiles within a task). Respondents are then randomly assigned to one design block within their condition.

- The order of the survey modules and the DCE is randomized.

All randomization is done by computer; the research team does not observe or modify treatment assignment during data collection.

Individual (assignment to experimental condition) and choice‑task level (for attribute profiles within each respondent).

No

Not clustered.

Approximately 5,000 respondents (yielding 50,000 choice observations, as each respondent completes 10 tasks).

- Condition A (choice sets with only privately provided AI assistants): approximately 2,500 respondents.

- Condition B (choice sets that also include a publicly provided AI assistant): approximately 2,500 respondents.

- Within Condition B, approximately 1,250 respondents are assigned to each public-provider description arm.