AI-supported Skills Elicitation, Personalized Information, and Youth Unemployment in South Africa

We implement a two-part, individual-level randomized intervention on SAYouth.mobi (in partnership with Harambee) targeting unemployed South African youth (18–35).
1. AI-supported skills elicitation & CV builder (vs. status quo and a static skills-picking form):
Half of participants are offered an AI conversation (embedded in our Compass tool) that helps them surface prior experience (including informal work), translate it into a clearer skills profile, and generate a CV; the other half completes a control online task and proceed with business-as-usual platform resources during the study window. We track clicks and usage from the chatbot and links it provides.
2. Personalized labour-market information (cross-randomized):
Participants are also randomized to one of two concise, personalized messages that provide information on the labour demand of their elicited skills.
• Pull-only (“good-news only”): highlights one of the user’s skills that is above their portfolio average in current vacancy demand.
• Push+Pull (“good-news + bad-news”): shows the same good-news line, plus an added “bad-news” line about the user's skill areas with below-average demand. This isolates whether adding unpleasant facts discourages action despite identical recommended next steps. We hold message length constant across arms.

Design notes and delivery: Recruitment and consent occur on-platform; up to ~4,000 users are enrolled. Messages are delivered in-app and/or by link and are brief, quantitative, and auditable (shares/counts derived from tagged vacancies). Participant incentives (not tied to treatment assignment): participants in both groups receive R30 to complete the online task to offset internet costs; those who complete the phone survey receive extra R50.

2025-11-24

2025-12-23

These may still be reduced in the final PAP (uploaded before endline)
Primary treatment: job interviews/callbacks, job offers, employment (at ~3 months), earnings (at ~3 months), confidence in skill communication
Cross-randomization: information demand/avoidance, level of job search effort, variance in job search effort, job search directedness, posterior beliefs about demand for one’s skills

These may still be reduced in the final PAP (uploaded before endline)
Primary treatment: time use on unpaid work, employment quality, job satisfaction (adjusted cantril ladder), reservation wage, aspirations
Cross-randomization: motivation, locus of control, recall of demand for one’s skills, jobmarket-specific self-efficacy, short CES-D, Substitution away from non-recommended skill areas (applications to “X” falling vs. “Y” rising); reporting total vs. reallocated search,

We conduct an individual-level randomized evaluation with unemployed South African youth (ages 18–35) on the national employment platform SAYouth.mobi, in partnership with Harambee Youth Employment Accelerator. The goal is to test scalable, low-cost ways of helping jobseekers identify and communicate their skills, update beliefs about their labour-market prospects, and act on better information
The experiment follows a 2 × 2 factorial design with two interventions cross-randomized at the individual level:

1. AI-supported skills elicitation & CV builder (vs. status quo).
Half of participants are offered access to an AI conversation that helps them recall prior experiences (including informal work), translate them into clearly expressed skills, and generate a CV. The comparison group completes a control online task (to make effort levels and compensation equal) and continues to use standard SAYouth.mobi tools during the same period.

2. Personalized labour-market information
Participants receive short personalized messages summarizing how their skills relate to current job demand. Both versions recommend the same next step (e.g., explore or apply for opportunities in line with highly demanded skill areas), but one presents only positive information, while the other combines positive and negative facts. This variation allows us to test whether adding “bad news” discourages or sharpens engagement and belief updating, given the more complete set of information. The design ensures that all participants are exposed to realistic, actionable information.

Participants are randomly assigned by the platform at the time of onboarding. Recruitment, consent, and delivery occur online within SAYouth.mobi and at the start of the online task/conversation with the AI tool. We plan to enroll roughly 4,000 participants, with follow-up data collected from platform usage logs and phone surveys at ~3 months after onboarding.

Primary outcomes include employment measures, behavioural measures (click-throughs and job applications), and belief updating about job prospects. Secondary outcomes include self-confidence, well-being, and other downstream employment indicators.

Not available

online after consent/accepting participation

individual

No

NA

4000

2000 treatment, 1800 business-as-usual control, 200 static-form control

The MDE is expressed at the individual level and standardized (units = standard deviations). Under our pre-analysis assumptions (two-sided α = 0.05, power = 0.80, equal allocation across arms, individual-level randomization, planned sample ≈ 4,000), the standardized minimum detectable effect is 0.089 standard deviations. In other words, we are powered to detect a treatment effect equal to 0.089 × SD of the outcome (0.089 × √[p(1−p)] for binary outcomes).