Vanguard Open submissions close September 25Enter now →

Research

Evidence first. Then the argument.

Everything AI Vanguard asks of schools and districts rests on primary research: what students told us, what teachers told us, and now a preregistered look at whether student voice changes policy at all.

Current study · UnderwayPolicy survey · 447 students · 2025Teacher pilot · 10 educators · 2026

Current study · Preregistered

When Does Student Voice Change AI Policy?

Led by Avery Updike, AI Vanguard's founder, as an independent preregistered study.

Records collection and coding underway

Protocol fixed · September 2026

Research question

What conditions are associated with consequential student participation in generative-AI policymaking in U.S. school districts?

01What methods do students use to participate in district AI policymaking?
02Which aspects of participation correlate with a traceable response to student recommendations?
03How does a student recommendation move from its origin to adoption, modification, rejection, acknowledgment, or an unknown disposition?

Districts in the fixed sample

Largest U.S. systems, set before coding

Public-records requests filed

Sept 3 – 18, 2026

Independent coders

Agreement reported as Cohen's kappa

0–4

Disposition scale

Unknown through substantially adopted

The sample

Twelve districts, chosen by someone else.

The twelve largest U.S. school districts examined by Liang, Hu, Ren, and Trinidad (2026) in Educational Policy. The sample was fixed before any outcome coding, so it could not be chosen to fit a story.

These are large, well-documented systems. Findings will describe them, not all U.S. districts. Public-records requests filed under state open-records laws with twenty large school systems, the twelve in the sample among them, between September 3 and September 18, 2026.

01New York City Public Schools
02Los Angeles Unified School District
03Miami-Dade County Public Schools
04Chicago Public Schools
05Clark County School District
06Broward County Public Schools
07Hillsborough County Public Schools
08Orange County Public Schools
09Palm Beach County School District
10Houston Independent School District
11Gwinnett County Public Schools
12Fairfax County Public Schools

What counts as evidence

A press release is not a consequence.

A recommendation episode: a specific concern, proposal, or recommendation attributable to students, traced through the district's process to its last documented step.

A traceable consequence is coded only for 2, 3, or 4, and only when the recommendation's origin is known, the input preceded the decision, a response is documented, and evidence links the two. A reasoned rejection counts as a consequence; being ignored does not.

Disposition of each student recommendation

0Disposition unknown or untraceable
1Acknowledged, no serious response established
2Considered and rejected with reasoning
3Adopted in modified or partial form
4Substantially adopted

Strength of the evidence behind it

AA contemporaneous recommendation, a record of the discussion, and a comparison between draft and decision.
BAn official district document directly linking a decision to student input, with less detail on the process.
CA retrospective account or media report corroborated by another source.
DParticipant lists, survey announcements, or task-force descriptions: evidence of access, not of consequence.

Preregistration

Sample, coding rules, outcome measure, evidence levels, reliability thresholds, and stopping rules were fixed in a written protocol before outcome coding began. Any deviation is logged with its date, reason, and effect on the analysis.

Reliability

Two independent coders code every eligible episode from the same codebook and sources without seeing each other's work. Agreement is reported with Cohen's kappa, weighted for the ordinal disposition scale. If raw agreement falls below 80% or kappa below 0.60, coding stops, the unclear rule is revised, and affected episodes are recoded.

Researcher position

The study's lead founded AI Vanguard. The protocol names that as a possible source of bias and builds in controls: a fixed sample, an outcome measure that counts reasoned rejection, unknowns kept unknown, contradictory evidence retained, a second independent coder, and AI Vanguard's own materials held to the same evidentiary standard as any other record. Findings that show no student influence will be reported.

What happens next. Descriptive findings and up to four process-tracing case studies will be published on this page when the preregistered analysis is complete.

Sample source: Liang, S., Hu, Y., Ren, X., and Trinidad, J. E. (2026). Governing GenAI through Redefinition, Regulation, and Innovation: Policy Responses in the Largest US School Districts. Educational Policy. DOI ↗

2025 Student AI Policy Survey

447 students. 6 schools. One clear message.

Our first research cycle, run through campus representatives at partner schools. The raw data behind the policy conversations we bring to districts, and the evidence base for the 2026 policy brief.

First cycle · Fall 2025

Aug 25 – Dec 17, 2025

84%

Use AI for schoolwork

Frequently or occasionally

87%

Find AI helpful

Very or somewhat

90%

Say AI is acceptable

With guidance or for specific tasks

74%

Want schools to teach responsible use

The dominant policy preference

What students want schools to do

The dominant preference isn't a ban. It's guidance.

Only 4% of students called AI use outright cheating. 74% asked schools to teach responsible use, and 35% explicitly asked to be involved in shaping the policies themselves, the core justification for our representative model.

01
Teach students how to use AI responsibly
74%
02
Involve students in shaping AI rules and policies
35%
03
Create stricter rules to limit misuse
26%
04
Allow free use with minimal restrictions
24%
05
Do nothing; AI use should be up to individuals
14%
06
AI should complete most of the work
5%

Methodology

447 responses collected via Google Form, Downey (55%), Cerritos (41%), Cypress (4%) representing the bulk of responses. Multi-select questions total more than 100%. Raw data retained by AI Vanguard; percentages are computed from every submitted response, not a sample.

2026 Teacher pilot

The other side of the desk.

A companion pilot survey of 10 educators in January 2026. Small sample by design: a read on where teachers sit before we broaden distribution. The findings hang together with the 447-student data.

Pilot cycle · Jan 2026

Jan 12 – Jan 15, 2026

80%

Feel pressure to integrate AI

Yes or somewhat

80%

Suspect frequent unauthorized student use

Rated 4 or 5 out of 5

30%

Accuracy on a 3-paragraph detection quiz

Worse than random chance (33%)

0/10

Correctly identified the student-written paragraph

Every teacher called it AI or AI-assisted

The gap this exposes

Pressure to adopt. Confident detection. Wrong answers.

The survey asked teachers to classify three unlabeled paragraphs as student-written, AI-assisted, or AI-generated. Teachers averaged 30% accuracy, below the 33% you'd expect from guessing. Every teacher misidentified the paragraph that was actually written by a student.

Actually: Student with AI assistance

6 of 10 teachers correctly identified

60%

Actually: Entirely AI

3 of 10 teachers correctly identified

30%

Actually: Entirely student

0 of 10 teachers correctly identified

Rep-led field study

Can teachers actually tell?

A separate qualitative study, distinct from the teacher pilot. An AI Vanguard representative asked five teachers to compare an AI-assisted paper against a student's original work, first blind, then with the source revealed, and recorded how their grading shifted.

Teachers surveyed (blind then revealed)5

Preferred AI version — both before and after reveal2

Preferred student work — grades shifted toward student after reveal3

AI Perception Change

Artificial Intelligence versus the human mind. Both are capable of anything. However, it is commonly asked if the two can be compared when it comes to writing. With AI on its way to becoming every student's shortcut to success, millions of papers are written with its help daily. The voices of students themselves slowly fade, but can their teachers tell? Research was conducted in which several teachers were asked to compare an AI assisted paper with a student's original work. Without knowing one was written by AI, the teachers gave grades and explained which they preferred and why. After revealing one was AI, the teachers were then asked if their perceptions on the papers had altered. Out of the five teachers, two teachers were on the same page, while with different thoughts, the other three were aligned. All were surprised, but after the realization set in, they could easily tell the difference between both papers. The two teachers with similar thoughts preferred the AI version through and through as seen through their given grades. Before the reveal, with slight hesitation they chose the AI written paper believing it was easier to read and straight to the point. They had agreed the student's original work had more citations and went into detail, but overall the human written paper was not on the same level as the digital brain for them. After finding out that one was AI, their perceptions had not changed as they believed the AI version was more effective as it was easier to read and straight to the point. Looking at the other three teachers, from the start, they had chosen the student-written paper as they could hear the student's unique voice through the writing. They asserted that the AI paper was a solid essay with just the facts but that's what had made it unmemorable and boring. After the unveiling, their grades for the AI paper dropped and the student's grade rose. They had come to the conclusion that the student's paper was significantly more lively and memorable because of the risk-taking and figurative language.

Written by Saiya Bhakta, AI Vanguard student representative

The research became six asks.

Our 2026 policy brief translates the student and teacher findings into six concrete recommendations for schools and districts. Districts, researchers, and reporters who want the underlying data can ask for it.

Read the policy brief Join as a student researcher