AI Adoption Spillovers

The intervention is the random assignment of students to physical seats (workstations) within each test session. Seating determines how many peers are seated ahead of and visible to each student, and therefore how much a student is exposed to peers’ on-screen behavior, including AI use. The study is conducted in two contexts: sessions in which AI use is encouraged and sessions in which it is not permitted. The seating experiment covers three computerized tests — Language, Math, and Problem-Solving — where Problem-Solving is a single test administered in two sittings to two groups of students. Every workstation screen is recorded. (Students sign up per test and need not take all tests.)

2026-05-30

2026-06-06

The single outcome variable is a binary dummy — Any AI use during the test (1 if the student used any generative-AI tool at any point, 0 otherwise) — in the AI-endorsed sessions (Language, Math, Problem-Solving I and II). We pre-register two primary estimands on this dummy: (1) ITT — the effect of the randomly assigned number of peers seated ahead of you on the probability you use AI; and (2) TOT/IV — the effect of one additional in-front peer actually using AI on the probability you use AI, instrumenting the number of in-front peers using AI (within a pre-specified early-test window) with the number of students seated in front of you. The ITT is the reduced-form (policy) effect of seating; the TOT is the per-peer transmission effect.

AI use is measured from each workstation’s screen recording using image and video analysis. The variable equals 1 for any detected use and 0 otherwise; workstations with unusable recordings are treated as missing (not as zero). Both primary estimands (ITT, TOT) are estimated only for this binary outcome. For each estimate the identification comes from random seat assignment within session-room. We report p-values two ways for the same point estimate: heteroskedasticity-robust (student-clustered) standard errors, and randomization inference (re-drawing the within-room seat assignment). Randomization inference is only a way to compute p-values; it is not a separate identification strategy, and the point estimate is the same either way. Multiple-hypothesis testing across the two primary tests will be handled using standard practices.

1. Intensity of own AI use, measured only over the period after the in-front-peer window (e.g., if the window is 15 minutes, intensity is the share of the remaining 75 minutes during which AI was used), including a TOT/IV variant of the per-peer transmission effect with intensity as the outcome. 2. Time to first AI use. 3. Same-tool matching with in-front peers (descriptive mechanism check): conditional on adopting AI, are students whose in-front peers used ChatGPT more likely to use ChatGPT (rather than Gemini), and vice versa? 4. AI use in the AI-forbidden sessions — exploratory.

Secondary AI-use measures come from the same image and video analysis of the recordings. The tool-matching check uses an indicator (ChatGPT vs. Gemini) for the in-front peer and for the focal student. The TOT/IV (2SLS) estimand is primary for the binary outcome (see Primary Outcomes); its intensity-outcome version is reported here as a secondary variant. In the AI-forbidden sessions the same detection is applied to flag illicit use; effects are exploratory given near-zero expected base rates.

Students at four campuses take computer-based tests on two dates: Language and Math on May 30, 2026, and Problem-Solving (in two sittings/groups) on June 6, 2026. At two campuses AI use is encouraged; at two it is not permitted. Within each test session, students are randomly assigned to workstations. Because students chiefly observe peers seated ahead of them, random seating creates exogenous variation in exposure to peers’ (potential) AI use. Screen recordings provide objective, individual-level measures of AI use. The primary analysis estimates the intention-to-treat effect of random seat-based peer exposure on the probability of own AI use in the AI-endorsed sessions.

Not available

Randomization done in the office by a computer: a pseudo-random permutation (documented seed/script) assigns students present at a session to workstations. Randomization is stratified by session-room (campus × test × date) and is re-drawn independently for each session.

The individual student’s seat (workstation) within a session is the unit of randomization.

No

Randomization is at the individual seat level, so there are no randomization clusters in the usual sense. For reference, the seating experiment spans 4 campuses and 16 sittings (4 per campus: Language, Math, and Problem-Solving twice), of which 8 are AI-endorsed and 8 AI-forbidden. For inference, observations are clustered by student; because students sign up for different subsets of tests, the number of distinct students is set at enrollment (bounded by per-campus capacity 36/40/30/25).

Up to 304 student × test observations in the primary (AI-endorsed) seating sample, and up to 524 seating observations in total (304 AI-endorsed + 220 AI-forbidden). Each campus contributes 4 sittings (Language, Math, and Problem-Solving twice). These counts are upper bounds: realized N will be lower in any session in which fewer than capacity-many enrolled students show up.

Treatment is a continuous, randomly assigned exposure (number of peers seated ahead), not discrete arms, so sample size is reported by design cell (each campus runs 4 sittings: Language, Math, and Problem-Solving twice). Numbers below are upper bounds; realized N in each cell depends on attendance.

The calculations below are for the ITT (the effect of randomly assigned peers-ahead exposure on Pr(AI use)); they do not characterize power for the TOT/IV, which depends on the realized first-stage strength. Primary outcome: any AI use (binary; SD = sqrt(p(1−p)) at base rate p). Assumptions: two-sided alpha = 0.05, power = 0.80, primary N = 304 AI-endorsed student×test observations at full attendance, session-room fixed effects, p-values reported both ways (heteroskedasticity-robust/student-clustered, and randomization inference). The power calculation uses a single expected room configuration — a 6×6 grid of 36 fully-occupied workstations; actual rooms may differ. Under it the number of peers seated ahead has SD ≈ 10. We use the full N ≈ 304 for the headline and report a conservative within-student-clustering bound (effective N ≈ 190) alongside. Low- vs. high-exposure contrast (risk difference, pp): • Headline (N=304): baseline 10% → 11.6 pp; 20% → 14.1 pp; 30% → 15.4 pp; 40% → 15.9 pp; 50% → 15.7 pp. • Conservative bound: 15.2 / 18.1 / 19.5 / 19.9 / 19.5 pp respectively. Continuous exposure (LPM slope, N=304, SD≈10 under the assumed room): MDE ≈ 6.4–7.9 pp per +1 SD of peer exposure (about 0.6–0.8 pp per additional peer ahead), roughly invariant to the base rate over 20–40%; ≈ 9 pp per SD under the conservative bound. Headline: the study can detect about a 15 percentage-point increase in P(AI use) for a low-vs-high contrast at a 30% baseline (≈19 pp under the conservative clustering bound), or about 7–8 pp per one-SD increase in continuous peer exposure (≈9 pp conservative). It is powered for moderate-to-large contagion effects, not small ones. Secondary AI-forbidden sample (up to N=220): ~11.5 pp at a 5% baseline, but expected floor effects make it exploratory.