Beliefs about AI Exposure and the Demand for Redistribution: An Information Experiment in a Japanese Panel

A randomly selected half of respondents is shown a short information screen at a fixed position in the questionnaire. The screen contains a truthful, published estimate, attributed on screen to its source, of the share of working people in Japan who are in jobs with high exposure to artificial intelligence (AI), together with a brief statement that the form of AI's impact differs across jobs. The control half sees no information screen. All other survey content is identical across the two groups. The exact wording, number, and source are documented in the hidden fields and the attached analysis plan, embargoed until trial completion.

2026-08-24

2026-11-30

(1) Support for reducing income differences between the rich and the poor (survey item policy22), measured on the survey's standard five-point oppose-support scale.
(2) Support for uniform benefits to all citizens or consumption-tax cuts financed by future tax increases (survey item policy10), same five-point scale.
Both items sit inside the panel's standard policy battery. The battery's format has been the same since 2020 - the same stem, the same five-point oppose-support scale and the same separate "don't know" option - and items that recur are worded identically across waves, but the set of items has changed as topics were added and retired: policy10 has been fielded in seven waves, policy22 in one.

The two items are co-primary: policy22 is the canonical redistribution question (closest to the cross-national literature); policy10 is the panel's longest-running redistribution item (waves 3-8 and 14). Both co-primary items were fielded in wave 14, so both have pre-treatment values available as controls; what distinguishes them is that policy10 has been fielded often enough for its within-person autocorrelation to be measured, while policy22 has not. Items are answered on a 1-5 scale (1 = oppose ... 5 = support) with a separate "don't know" option; "don't know" responses are treated as missing in the scale analyses, and their incidence is analyzed as a secondary outcome. Hypothesis tests are adjudicated on policy22, with policy10 reported alongside and interpreted in light of the following registered circumstance. In July 2026, before registration, the Government of Japan announced a temporary reduction of the consumption tax on food effective April 2027. The policy10 wording ("consumption-tax cuts financed by future tax increases") therefore refers to a policy that has moved from hypothetical to announced between waves 14 and 15, and policy22, whose wording is unaffected by the announcement, is expected to be the cleaner of the two co-primary outcomes.

(1) Support for strengthening taxation of high earners (item policy23), same five-point scale.
(2) Support for the promotion of artificial intelligence (AI) and big-data use (item policy8).
(3) Support for regulating the development and use of generative AI (item policy11), same five-point scale.
(4) Post-battery belief about the share of workers in jobs with high AI exposure, numeric 0-100 in the same format as the prior question (same-wave first stage).
(5) Wave-16 (early 2027) re-asks of the 0-100 belief question and the policy battery, conditional on the items being re-fielded in wave 16 as agreed with the survey organization in July 2026.
(6) Incidence of "don't know" responses on the primary outcome items.
(7) Belief about the number of foreign workers per 100 working people (numeric, 0-100).

Item (1) extends the redistribution outcome family: policy23 is a sharper tax-the-rich item. Item (2) is not a redistribution outcome but the mediating attitude: support for AI promotion itself, through which any effect on redistribution demand may run. Item (3) is the battery's second AI item and a plausible destination for the information; it is named here rather than left to the exploratory spillover family, but the study's hypothesis concerns compensation through redistribution and makes no prediction about regulation, so no direction is registered for it. Item (4) measures whether the information registered - the same-wave first stage; because the numeric format is fielded, it additionally supports a learning regression relating updating to the distance between the information and the respondent's prior. Item (5) measures persistence of belief updating and of any policy-preference effects about four months after treatment, with no re-supply of information. Item (6) tests whether information provision changes the propensity to answer the policy items at all. Item (7) is a belief about a topic unrelated to the information content, used to test whether the information moved beliefs about AI specifically rather than beliefs in general. The remaining policy items in the same battery, none of which concerns AI, are analyzed as exploratory spillover outcomes.

The experiment is embedded in wave 15 of the NIRA-Okubo Japanese national online panel (NIRA is the Nippon Institute for Research Advancement; fielded in the second half of 2026 by a commercial survey company; roughly 9,000 expected completes). Continuing panelists are invited first and are not screened on employment status, and newly recruited respondents, who fill the remaining cells of the survey company's stratified design, are in employment at recruitment. The analysis sample is restricted to respondents aged 18 and over, a restriction adopted for the university ethics review and applying identically across arms; it affects at most about 0.2 percent of respondents, judging by the six most recent waves. Early in the questionnaire, all respondents report a prior belief: out of 100 working people in Japan, how many are in jobs with high exposure to artificial intelligence (AI), as a numeric answer from 0 to 100, with a definition of "exposure to AI" shown to everyone. Respondents are randomly assigned (50/50, individual level) to an information group, which later sees a short truthful statement on the published estimated share (see the intervention description), or to a control group, which does not. The survey's standard policy-preference battery (about ten items, five-point oppose-support scale, format identical to all previous waves) contains the primary outcomes. After the battery, all respondents answer a belief question about the exposed share. Wave 16 (early 2027) re-asks the belief question and the battery with no information provision and no new randomization. The assignment indicator is stored permanently as a respondent-level variable. A separate, independently randomized survey module on an unrelated topic appears elsewhere in the same questionnaire and is registered separately.

Not available

Randomization is carried out by the survey company's platform at the individual respondent level, at the moment a respondent begins the questionnaire. Each respondent receives an independent random draw over the eight combinations of the two experiments fielded in this wave (two information conditions for this experiment, crossed with four conditions of a separate investment-framing experiment registered separately), each combination with probability one eighth. Assignment is therefore 50/50 for this experiment and independent of the other experiment by construction. The draw is made when the respondent enters the main survey, and a number is drawn for every respondent who enters, including those who later break off; the pattern does not change if a respondent pauses and resumes. The assignment indicator is stored permanently as a respondent-level variable. The assignment involves no blocking, no stratification and no quota balancing: group sizes are left to come out as chance determines.

Individual (respondent)

No

Approximately 9,000 (clusters are individuals; individual-level randomization).

Approximately 9,000 wave-15 completes (about 4,500 per arm); approximately 6,300 expected at the wave-16 follow-up (70 percent continuation, about 3,150 per arm). Respondents under 18, whom the panel admits and who make up about 0.2 percent of a wave, are excluded from the analysis sample.

2 arms: approximately 4,500 information, 4,500 control.
The wave also carries a separate four-arm experiment, registered separately, and the survey platform assigns respondents over the eight combinations of the two (2 x 4). Assignment to the two arms of this trial is therefore approximately half and half, and approximately one eighth of respondents falls in each of the eight combinations.

All inputs are measured on waves 1-14 of this panel. The primary outcomes are five-point items whose standard deviation, excluding "don't know" responses, is 1.06-1.10 for one co-primary item (seven waves) and 0.99 for the other (one wave). The within-person one-wave-apart R-squared of the same item is 0.10 for the redistribution outcome, which is the value used for covariate adjustment. Wave 15 refreshes the panel; wave 14 completed 8,215 interviews and the three previous refreshes following an unrefreshed wave grew the sample by 9, 10 and 14 percent, so power is planned at 9,000 completes (4,500 per arm), with 8,000 as the pessimistic case. Retention into wave 16 is set at 70 percent, deliberately just below the lowest of the thirteen observed transitions (the observed range is 70.5-85.5 percent): the average is not the right planning value because first-time respondents return at 61 percent against 78 percent for established panelists, and a refreshed wave carries more first-time respondents. Using MDE = 2.8 x SD x sqrt(2/n per arm) at 80 percent power and a 5 percent two-sided test: (i) simple treated-control difference: 5.9 percent of a standard deviation, that is 0.058 to 0.065 scale points depending on the item; (ii) conditioning on the lagged (wave-14) outcome and covariates: approximately 5.6 percent; (iii) wave-16 follow-up at approximately 3,150 per arm: approximately 7.1 percent; (iv) interaction of treatment with the standardized prior gap (balanced arms): approximately 5.9 percent per one standard deviation of the gap.