"Rewiring the Code: AI-Powered Gender Role Models and Gender Attitudes - Evidence from a School Randomized Controlled Trial"

We run a class-level, three-arm RCT in lower- and upper-secondary schools to evaluate whether exposure to STEM role models affects students’ school-track intentions, interest in STEM, and gender attitudes. Classes are randomly assigned to:

1. Female Role-Model videos,
2. Male Role-Model videos, or
3. Neutral control videos (orientation content without role models).

All arms view short videos in class and complete a brief survey and the Implicit Association Test (IAT) - Gender Science immediately after.

We conduct a class-level, three-arm randomized controlled trial (RCT) in lower- and upper-secondary schools to evaluate how exposure to gendered STEM role models influences students’ school-track intentions, interest in STEM, and gender attitudes.

Classes are randomly assigned to one of three groups:

1. Female Role-Model videos,
2. Male Role-Model videos, or
3. Neutral control videos presenting orientation content without role models.

All students watch a short video sequence (approximately 15 minutes total) during a single class session (about 70 minutes), followed by an immediate post-intervention survey and an Implicit Association Test (IAT) Gender-Science. The intervention materials are available in both Italian and German, and sessions are conducted using headphones in computer labs to minimise spillovers.

The videos feature four young professionals working in different STEM fields. For treatment groups, the same individuals appear in AI-edited versions that vary only by the apparent gender of the speaker, allowing us to isolate the causal effect of role-model gender on students’ perceptions and aspirations.

Immediately post-viewing, students complete: (i) questions on intended next school track and future plans/aspirations, STEM interest, classroom behaviour, peers/teachers relationship, explicit STEM & gender attitudes; (ii) a Gender-Science IAT.

Parental consent is obtained in advance; students use headphones to avoid spillovers.
A short debrief is provided at the end of the session; follow-up administrative linkage (where permitted) is planned to observe realised track choices.

2024-10-21

2026-01-31

1. Choice of next school track (intention immediately post-treatment; realised choice if available).

2. STEM interest/aspirations (e.g., likelihood of choosing STEM track/subjects; interest in STEM activities/careers).

3. Gender attitudes (explicit scales and Gender-Science IAT D-score).

1. Next school track choice

Definition: Students’ intended or actual choice of upper-secondary school program after middle school (or next educational step for older cohorts).

Measurement:
(a) Immediate intention collected post-intervention through a fLikert question listing all provincial school tracks (e.g., liceo scientifico, tecnico tecnologico, linguistico, etc.).
(b) Realised enrolment obtained, where possible, by linking administrative data on actual school registrations.

Construction: Outcomes include (i) a binary indicator for choosing a STEM-oriented track and (ii) a categorical or multinomial variable showing the distribution across all school-track types.

2. STEM interest and aspirations

Definition: Students’ interest in STEM subjects, motivation to continue studying them, and long-term aspirations toward STEM-related careers.

Measurement: Based on harmonised survey items capturing:

a. Enjoyment of or interest in STEM subjects (e.g., “I like maths/science classes”).
b. Intention to pursue advanced STEM courses.
c. Aspiration to have a STEM-related profession.

Construction: Each item is standardised (z-score), and if multiple items load on the same construct, they are combined into an index using inverse-covariance weights (Anderson, 2008).
The resulting index represents an overall measure of STEM interest/aspirations, where higher values indicate stronger interest or motivation toward STEM pathways.


3. Gender attitudes

Definition: Students’ explicit and implicit beliefs about gender roles and abilities in STEM fields.

Measurement:
(a) Explicit attitudes: Agreement with statements such as “Boys are naturally better at science than girls” or “Both men and women can be good engineers.” Responses are combined into a composite index reflecting gender-stereotypical views.
(b) Implicit attitudes: Measured through the Gender–Science Implicit Association Test (IAT), producing a D-score that captures the strength of implicit associations between gender and science.

Construction:

- Explicit indices are computed as standardised (z-scored) averages across items, directionally aligned so that higher values indicate less bias (more egalitarian views).
- IAT D-scores are computed following the pooled-standard-deviation algorithm, excluding error-prone or extreme trials per standard guidelines.
- All continuous outcomes are standardised within study wave to facilitate interpretation and comparability.

1. STEM self-efficacy and perceived fit/belonging in STEM.

2. Ability to detect AI-edited content.

3. Perceptions of scientists’ personal attributes (engagement, self-confidence, and competitiveness) across AI-manipulated male/female role-model versions.

4. Perceived returns and costs of STEM tracks.

5. Beliefs about others’ expectations (descriptive/normative norms).

6. Parents' influence on child's next school track choice.

Each domain is measured with pre-specified items; indices will be constructed (z-scores, directionally aligned, inverse-covariance weighted) and analysed alongside individual components. Binary behaviours (e.g., “plans to attend open day”) are coded as 0/1.

1. STEM self-efficacy and perceived fit/belonging in STEM

This measures how confident students feel about performing well in science, technology, engineering, and mathematics (STEM) subjects and whether they feel they “belong” in those fields.

Self-efficacy refers to beliefs such as “I am good at solving science problems” or “I can succeed in a STEM career.”

Belonging/fit captures whether students see STEM as “for people like me.”
These constructs are measured with Likert-scale questions and combined into a standardised index (e.g., mean=0, SD=1).

2. Ability to detect AI-edited content

This outcome captures the detection ability: whether students realise that some of the videos were AI-edited (i.e., the role model’s gender was swapped). Students are asked if they noticed any AI manipulation or if they believe the videos were digitally altered.

3. Perceptions of scientists’ personal attributes (engagement, self-confidence, and competitiveness) across AI-manipulated male/female role-model versions.

After watching each video, students rate each scientist on three 5-point Likert scales capturing perceived engagement, self-confidence, and competitiveness. The same role models appear in both a male and a female version (AI gender-swapped). By comparing average ratings between the male and female versions of the same scientist, we assess whether gender presentation affects perceived competence or likeability.
This provides an implicit evaluation measure of gender bias, complementary to explicit attitude and IAT-based measures.

4. Perceived returns and costs of STEM tracks

This outcome measures students’ beliefs about the advantages and disadvantages of choosing a STEM-oriented educational path.

Perceived returns include expected job opportunities, salaries, or prestige of STEM careers.

Perceived costs include difficulty, workload, or fear of failure. These perceptions are key predictors of school-track choices and may shift after exposure to role models.

5. Beliefs about others’ expectations (descriptive and normative norms)

This outcome explores social norms related to gender and education:

Descriptive norms = what students think others typically do (e.g., “Most girls in my class are not interested in technology”).

Normative norms = what students think others expect them to do (e.g., “My friends think it’s strange if a girl studies engineering”).
These beliefs capture perceived social pressure and are important mediators between gender stereotypes and individual aspirations.

6. Parents’ influence on child’s next school track choice

This measures the perceived role of parents in students’ educational decision-making.
Students report whether their parents influence their decisions. Additional analysis is performed also based on parents’ level of education and occupation.
The outcome helps understand how family guidance interacts with role-model exposure in shaping aspirations.

The study is a cluster-randomised, parallel-arm RCT implemented at the class level in lower- and upper-secondary schools. The intervention tests whether exposure to gendered STEM role models influences students’ school-track intentions, interest in STEM, and gender attitudes.

Classes are randomly assigned to one of three conditions:

1. Female Role-Model videos,
2. Male Role-Model videos, or
3. Neutral control videos presenting orientation content without role models.

Randomisation is stratified by school grade and language track (Italian/German) to ensure balance and limit cross-class contamination. The intervention and immediate post-survey are conducted during a single class period, typically in the computer lab or classroom using headphones.

Randomisation is conducted at the class level (cluster randomisation). We stratify by school system (Italian-speaking schools vs. German-speaking schools) and then randomise within school × grade. Thus, within each participating school, classes in the same grade are allocated independently across the three arms, while overall treatment shares are balanced across the two school systems.

Classes are assigned 1:1:1 to:

1. Female Role-Model videos,
2. Male Role-Model videos,
3. Neutral Control videos (orientation content without role models).

Allocation is implemented centrally via reproducible computer randomisation (fixed seed; allocation lists stored). Sessions last ~70 minutes and comprise:

1. Video sequence (≈15 min) showing four local STEM professionals. In treatment arms, videos are AI-edited so the same individual appears as either a woman or a man (only gender presentation varies).
2. Post-survey on next school-track intentions, STEM interest/aspirations, and explicit gender attitudes, plus background covariates.
3. Gender-Science IAT.

Delivery is standardised: researchers invigilate with uniform instructions; students use headphones/IT rooms to reduce cross-class spillovers. All materials are bilingual (Italian/German); translations were harmonised ex-ante to preserve item meaning. Participation requires parental consent; data are anonymised. Where permitted, post-hoc linkage to administrative records will capture realised enrolment.

Analytical plan (pre-specified): Intention-to-Treat (ITT) estimates with SEs clustered at the class level (school-level clustering in robustness checks). Main models use OLS/logit with blocking (system × school × grade) fixed effects and pre-specified covariates (e.g., student gender, baseline STEM interest/achievement if available, school type). Multiple-hypothesis adjustments are applied within outcome families (track choice; STEM interest/aspirations; gender attitudes). Prespecified heterogeneity: student gender, baseline STEM interest, teacher gender, school type, and school system (Italian vs. German); we also report system-specific estimates and test for treatment–system interactions.

A short debrief is provided after participation. Ethical approvals/authorisations were obtained from the relevant authorities in each school system.

Randomization done in office by a computer.

Classes

Yes

100 classes

1,300 students

30 classes control, 35 classes treatment arm 1 and 35 classes treatment arm 2

Assumptions for pairwise comparisons (each treatment vs control), two-sided α=0.05, power=0.80, cluster randomisation: Arms / clusters: ≈25 control classes; ≈25 Female-RM classes; ≈25 Male-RM classes. Students per arm: 370; avg class size (m): 15. ICC: 0.05 (conservative for attitudinal/choice outcomes). Design effect: DE = 1 + (m−1)·ICC = 1 + 14·0.05 = 1.70. Effective N (students) per arm: 370 / 1.70 ≈ 218. Based on these, the MDE ≈ 0.27 SD (Cohen’s d) for standardised continuous outcomes (each treatment vs control). For a binary outcome with baseline p = 0.50 (SD = 0.50), this corresponds to roughly 13.5 percentage points. Treatment vs treatment (Female-RM vs Male-RM): MDE is similarly ≈ 0.27 SD at α=0.05. Family-wise error control (Bonferroni over 3 tests, α≈0.0167): MDE ≈ 0.31 SD (≈ 15.5 pp at p=0.50). Sensitivity: If ICC = 0.03 (m=15 ⇒ DE = 1 + 14·0.03 = 1.42), MDE ≈ 0.24 SD. If m differs, DE changes accordingly (larger m ⇒ larger DE ⇒ higher MDE; smaller m ⇒ lower MDE). We will report realised ICCs and class sizes and re-compute detectable effects ex post.