Correcting Asymmetry of Information: How AI Persona Design Shapes Higher Education Enrollment Intentions

The intervention consists of a structured interaction with an AI-driven educational chatbot ("Academic Pathway Assistant," powered by Google Gemini via Vertex AI) designed to assist secondary school students, with a particular focus on those from migrant backgrounds. Participants are randomly assigned to one of five groups, each targeting a specific dimension of the informational wedge:

Arm 1 — Neutral/Factual (Active Control). A purely factual, reactive information assistant that delivers country-specific higher education facts without emotional framing. Directly targets the informational wedge at its root by correcting misperceptions about costs, pathways, and returns. Tone: polite, concise, and direct. No emojis, chitchat, or emotional validation. Closes with: "Does this answer your question?"

Arm 2 — Engaging/Enthusiastic (Treatment 1). A high-energy student ambassador presenting the same factual information framed as exciting discoveries ("Did you know?"). Designed to lower the psychological cost of the interaction and prevent attention-driven attrition. Uses exclamation points, emojis, and opportunity framing. Closes with an upbeat question about interests.

Arm 3 — Empathetic/Supportive (Treatment 2). A reassuring educational counselor that validates the student's emotions and anxieties before providing factual information. Each response begins with approximately two sentences of emotional support, then presents country-specific facts. Closes with: "How does that make you feel?"

Arm 4 — Relatable Mentor/Storyteller (Treatment 3). A near-peer mentor (older student persona) wrapping factual information in brief micro-stories about a "friend" or peer. May invent peer names and emotions but must not invent institutional details (e.g., grant amounts, deadlines). The story's resolution must be a real, country-specific fact. Closes with a connection-building question.

Arm 5 — Summer Activities (Placebo Control). A friendly summer activities guide discussing fun and interesting leisure activities specific to the student's country (outdoor activities, festivals, cultural events, local traditions, travel, sports, seasonal hobbies). This arm deliberately excludes all higher education content. It retains identical conversational flow constraints (response length, language matching, variation rules) but does not receive the educational knowledge base.

2026-06-01

2027-06-30

Intention to Enroll in Higher Education.

Measured as a continuous 0–100 percentage scale: "How likely are you to enroll in higher education?" Assessed identically in the pre-test questionnaire (before chatbot interaction) and the post-test questionnaire (after a minimum of 30 conversational turns).

The primary estimand is the between-arm difference in post-test enrollment likelihood, controlling for pre-test enrollment likelihood, baseline covariates, and country fixed effects (ANCOVA specification):

Y_is = α + Σ_k β_k T^k_is + γ Y^pre_is + X'_is δ + θ_s + ε_is

Where:
- Y_is = post-intervention enrollment probability for student i in country stratum s
- T^k_is = treatment arm dummies (reference category: Arm 1 for persona comparisons; Arm 5 for information provision test)
- Y^pre_is = baseline enrollment probability
- X_is = controls (age, gender, migrant generation, parental education, track type, counseling receipt)
- θ_s = country fixed effects
- Standard errors clustered at the country-stratum level

1. Academic Self-Efficacy
2. Revealed Preference (Information Seeking)
3. Future Planning Concreteness
4. Perceived AI Core Style (Manipulation Check)
5. AI Trust and Perceived Helpfulness
6. Behavioral Engagement (Attrition and Interaction Intensity)
7. Conversational Engagement Depth

1. Academic Self-Efficacy. Composite score of 3 Likert-scale items (1–7), measured pre- and post-intervention:
- "I believe I can understand even the most difficult material presented in higher education courses."
- "I am confident I could do an excellent job on assignments and tests in higher education."
- "I believe I have what it takes to succeed in a challenging higher education program."

Treatment effect: between-arm difference in post-test composite, controlling for pre-test composite. Primary mechanism variable for Arm 3 (self-efficacy scaffolding).

2. Revealed Preference (Information Seeking). Binary outcome (1/0) indicating whether the student opted in to receive a "Future Student Newsletter" at the end of the post-test survey. A costly signal of continued interest in higher education information, complementing the self-reported enrollment intention.

3. Future Planning Concreteness. Composite of post-test indicators capturing whether the informational wedge has been narrowed:
- Plans clarity (Likert 1–7): "How much clearer are your future educational plans after this conversation?" Key mechanism variable — RCT 1 found warmer personas significantly reduced plan clarity (p < 0.05).
- Information-seeking confidence (Likert 1–7): "How confident do you now feel about finding the information you need to make decisions about your future education?"
- Concrete next step (open text): "What is one concrete next step you plan to take regarding your future education?" Coded ordinally by specificity.
- Intention type shift: Pre-post change in stated intention type (University / College / Not planning / Don't know).
- Field-of-interest shift: Pre-post change in stated field of knowledge, area of interest, and program of interest.

4. Perceived AI Core Style (Manipulation Check). Semantic differential scales (1–7) validating that participants perceived the intended persona characteristics:
- Personality warmth: Cold & Unfriendly (1) to Warm & Friendly (7)
- Perceived competence: Incompetent (1) to Competent (7)
- Perceived objectivity: "The AI seemed objective and data-based" (Strongly Disagree to Strongly Agree)
- Perceived supportiveness: "The AI made me feel more capable of pursuing my goals" (Strongly Disagree to Strongly Agree)
- Perceived relatability: "The AI used relatable stories or examples from people like me" (Strongly Disagree to Strongly Agree)

Expected pattern: Arm 1 highest on objectivity; Arm 2 on warmth; Arm 3 on supportiveness; Arm 4 on relatability. Arm 5 provides baseline warmth absent educational content.

5. AI Trust and Perceived Helpfulness. Two Likert items (1–7):
- "How helpful was the AI assistant in supporting your educational planning?" (1 = Not helpful at all, 7 = Very helpful)
- "How trustworthy did you find the information provided by the AI assistant?" (1 = Not trustworthy at all, 7 = Very trustworthy)

Tests whether persona framing affects trust in AI-delivered information — a prerequisite for information to update subjective beliefs and correct the informational wedge.

6. Behavioral Engagement (Attrition and Interaction Intensity). Computed from automatically logged interaction data:
- Post-test completion rate (attrition): Binary indicator. Analyzed as both an outcome (which arm retains more students?) and a threat to internal validity (selective attrition). Replication target: RCT 1 found the Friendly Guide reduced attrition by 13.5 pp.
- Total conversational turns: Count of user–bot exchanges per session beyond the 30-turn minimum.
- Continued chatting after post-test: Binary indicator of choosing to continue after survey completion.
- Average user message length: Mean character count of user messages (engagement depth proxy).
- Average bot message length: Mean character count of bot responses (fidelity check on conversational flow constraints).

7. Conversational Engagement Depth. Derived from full-text chat logs stored in BigQuery:
- Session duration: Minutes between first and last logged interaction timestamp.
- User question diversity: Distinct topics raised by the user (coded via NLP topic modeling or manual annotation of a subsample).
- Sentiment trajectory: Change in user message sentiment from early to late turns (automated sentiment analysis).

This is a five-armed between-subjects Randomized Controlled Trial (RCT) grounded in the Human Capital Model under information asymmetry. It replicates and extends AEARCTR-0017212 with a placebo control arm and pre-registered hypotheses. Participating individuals from schools and NGOs are randomly assigned to one of five experimental groups:

1. Arm 1 — Neutral/Factual (Active Control): factual information assistant targeting the informational wedge through rational persuasion.
2. Arm 2 — Engaging/Enthusiastic (Treatment 1): enthusiastic guide testing positive affect priming and cognitive search cost reduction.
3. Arm 3 — Empathetic/Supportive (Treatment 2): empathetic counselor targeting self-efficacy via psychological scaffolding.
4. Arm 4 — Relatable Mentor/Storyteller (Treatment 3): near-peer mentor targeting underestimated benefits via narrative persuasion.
5. Arm 5 — Summer Activities (Placebo Control): leisure conversation guide isolating the effect of information provision from AI interaction.

Data is collected via pre- and post-interaction surveys embedded in the chat application. The post-test is triggered after a minimum of 30 conversational turns. All chat messages (user and bot) are logged with timestamps.

Not available

Computerized randomization using client-side stratified block randomization. Within each country stratum, participants are assigned using balanced permuted blocks of size 5, guaranteeing equal allocation across all five arms within every 5 consecutive participants per country.

Individual (stratified by country).

No

Not applicable (individual-level randomization, stratified by country).

525 students (105/arm), derived from d = 0.50 (medium), inflated for 40% attrition.

105 participants per arm

The primary outcome is self-reported enrollment probability (0–100 scale). Power calculations follow Duflo, Glennerster & Kremer (2007) for OLS estimation with a two-sided test at 80% power. The pilot (N = 94) yielded a pooled baseline SD of 22 pp and a pre-post correlation of 0.80–0.83 (R² = 0.90), implying a residual SD of approximately 13.4 pp after partialling out baseline enrollment and covariates. This RCT adopts a conservative pre-post correlation of 0.80. Planned sample. For a small-to-medium effect (d = 0.30; 6.7 pp), 63 analytic students per arm are required, yielding 315 total across five arms. Adjusting for 40% attrition gives a recruitment target of 525 students. At this sample size, the study is powered for any single pairwise comparison at α = 0.05. With Bonferroni correction for four planned comparisons (α = 0.0125), the MDE rises to 8.0 pp (d = 0.36); for all ten pairwise comparisons (α = 0.005), to 8.7 pp (d = 0.39). A small effect of d = 0.20 (4.5 pp) would require 142 per arm and is beyond the scope of this study. Pooled comparison (H1). Testing Arms 1–4 jointly against Arm 5 (information provision), with 252 treated versus 63 control after attrition, yields an MDE of 5.3 pp (d = 0.24), falling between small and small-to-medium. Heterogeneity interaction (H4). The MDE for the Treatment × Patience interaction is 13.4 pp at a 50/50 patience split and 13.7 pp at 60/40 — both below the pilot interaction of 14.54 pp. At the pilot's observed 90/10 split, the MDE rises to 22.4 pp and this test would be underpowered.