AI-Assisted Parenting Guidance and Child Development: A Randomized Experiment in Rural China

we recruited ten schools in a County in Hunan Province to participate in our experiment. Our baseline sample consists of over 3,500 fourth- and fifth-grade students, including 2,505 non-left-behind children from 78 classes in 10 public primary schools. Our interventions target parents of these 2,552 non-left-behind children. We also analyze spillover effects on left-behind children.


We use a two-level cross-randomization design. First, within each school-by-grade stratum, we randomly assigned classes to two parental guidance treatment groups or the control group. Second, within each class, we randomly assigned parents to one of three information nudge treatments or to a control. This design allows us to estimate the effects of each parental guidance treatment, each information treatment, and their interactions.


We design two parental guidance interventions: (i) \textit{parenting guidebook (PG)}. Parents received a guidebook on how to tackle parenting challenges. This guidebook, adapted from UNICEF's parenting education guidelines and tailored to the Chinese cultural context, includes practical examples related to ten key issues identified by local parents in a preliminary survey. (ii) \textit{AI-assisted parenting chatbot (AIPC)}. Parents were provided with an account to access our designed app, which is based on a trained large language model (Deepseek). This AI chatbot allows parents to acquire real-time, personalized advice for daily parenting challenges. The chatbot was trained using the same material as the PG intervention, enabling it to offer suggestions in line with UNICEF's guidelines. We assigned 26 classes to PG treatment groups and 26 classes to AIPC treatment groups. The remaining 26 classes were assigned as control classes (between-class controls) with no parenting guidance.



We then cross-randomized students across the three parental guidance
intervention arms and the information treatment arm. Each student (and their parents) was randomly assigned to one of three information treatment groups or to a control group: (i) \textit{Children's expressed needs}: Parents received information about children's expressed preferences for parental engagement, derived from the baseline survey. (ii) \textit{Benefits of supportive parenting}: Parents received information on the positive effects of patient and egalitarian parenting practices. (iii) \textit{Perspective-taking reflection}: Parents received a message encouraging recalling their own negative childhood experiences, such as corporal punishment. (iv) \textit{Control}: Parents in the control group received no information messages. We refer to this group as \textit{within-class controls}. All information was delivered to parents twice a month via SMS messages.

We design two parental guidance interventions and a between-class control. The parental guidance intervention began on September 15 2025, and ended on January 31 2026.

Treatment 1: Parenting guidebook (PG). Parents received a guidebook on how to tackle parenting challenges. This guidebook, adapted from UNICEF's parenting education guidelines and tailored to the Chinese cultural context, includes practical examples related to ten key issues identified by local parents in a preliminary survey. Starting in Week 2, parents also received biweekly reminders encouraging continued use of the materials. Figure [booklet] displays the guidebook.
Treatment 2: AI-assisted parenting chatbot (AIPC). Parents were provided with an account to access our designed app, which is based on a trained large language model (Deepseek). This AI chatbot allows parents to acquire real-time, personalized advice for daily parenting challenges. The chatbot was trained using the same material as the PG intervention, enabling it to offer suggestions in line with UNICEF's guidelines. We pre-registered accounts for treated parents to ensure access and distributed a user manual that explains how to interact with the tool. Parents were first coached on how to use the tool through a live demo. They then downloaded the app to their phones using a pre-registered account uniquely assigned to them, ensuring exclusive access. From Week 2, parents received reminders twice a month encouraging use of the AI chatbot. Figure [app_login] displays the login page of the AI-assisted chatbot app.
Control group (between-class control). No intervention.
Information nudge intervention
We design three information nudge interventions and a within-class control. Randomly selected parents receive the messages that were delivered twice a month from 1 October 2025 to 31 January 2026.

Children's expressed needs. [XXX County Education Bureau] Dear Parents, in September, we surveyed students at your child's school. The results show that [XX]% of students hope their parents will show more understanding and respect, [YY]% want more space to share their own ideas, and [ZZ]% would like more opportunities to communicate with their parents. ([XX]% varies across schools with a mean of 39% (range: 33–45%). [YY]% has a mean of 70% (range: 58–78%). [ZZ]% has a mean of 72% (range: 62–80%).)
Benefits of supportive parenting. [XXX County Education Bureau] Dear Parents, in September, we surveyed students at your child's school. The results suggest that treating children with patience and respect can reduce anxiety by [X]% and strengthen independent thinking by [Y]%. ([X]% has a mean of 13.61% (range: 5.6–19.1%). [Y]% has a mean of 12.43% (range: 6.1–20%).)
Perspective-taking reflection. [XXX County Education Bureau] Dear Parents, many of us remember growing up with harsh discipline and being expected to obey without question—and how unpleasant that felt. We encourage you to see things from your child's perspective: offer more understanding and respect in daily interactions, and make time to talk and listen.
Within-class control. No information.
To alleviate parental concerns regarding the guidebook and chatbot's credibility, both were officially endorsed by the local education bureau, bearing its name and logo.

2025-09-15

2026-01-31

Children's academic performance and mental health

Children's academic performance is measured by Chinese language and mathematics test scores. At the end of each school semester, each school administers an exam to assess students' academic performance. We obtain transcript data for each end-of-semester exam before and after the interventions from the Educational Bureau (The second-round test scores will be collected in early March 2026). We standardize the raw test scores in each school-grade cell according to the between-class control group to have a mean of 0 and a standard deviation of 1 by wave. As complementary measures, we also use parents' assessments of children's academic performance and parents' grade expectations.


Children's mental health is measured by the CES-D-10 scale. Students report how often, during the past week, they experienced each of ten feelings or behaviors (e.g., depressed mood, difficulty concentrating, poor sleep, loneliness). Each item is coded on a four-point frequency scale (0-3), with two positively worded items reverse-coded so that higher values consistently indicate worse mental health. We sum the ten items to obtain a total score ranging from 0 to 30. A score of 10 or above indicates an elevated risk of depression.

Intermediate outcomes: Parenting style and practice, Parent--child relationship quality, Parents' AI chatbot and guidebook usage, Parental educational inputs

Other outcomes include: (1) Socio-emotional skills captured by SDQ domains. (2) Non-cognitive skills including economic preferences, Big Five personality traits (BFI-10), self-control, misbehavior, and growth mindset. (3) Self-perceived creativity. (4) Peer networks measured by number of friends and perceived peer relationship quality.

Intermediate outcomes include: (1) Parenting style and practice, assessed from both child and parent perspectives using items adapted from Maccoby and Martin (1983), including discipline strategies, rule enforcement, positive parenting, perceived efficacy, and parental empathy. (2) Parent-child relationship quality, assessed by frequency of conflicts reported by parents and children's reports of verbal abuse and physical punishment. (3) AI chatbot and guidebook usage, measured through app-recorded chatbot usage logs and a five-question knowledge test on guidebook content. (4) Parental educational inputs, measured by children's reports of time spent on various activities with parents and parents' reports of monthly educational expenditure.


We aggregate items within each family and standardize scores relative to the control group (mean 0, SD 1). We also construct Anderson (2008) standardized indices within each outcome family and adjust p-values using the Romano and Wolf (2005) stepdown procedure.

More details about intermediate and other outcome variables can be found in the pre-analysis document.

We implement a two-level cross-randomized cluster randomized controlled trial in a rural county in Hunan Province, China.

Level 1 (Class-level): 78 primary-school classes (grades 4-5) in 10 schools are stratified by school and grade, then randomly assigned with equal probability to one of three parental guidance arms: (i) Parenting Guidebook (PG, 26 classes), (ii) AI-Assisted Parenting Chatbot (AIPC, 26 classes), or (iii) No-guidance control (26 classes).

Level 2 (Individual-level, cross-randomized): Within each class, parents of non-left-behind children are further randomized to one of three SMS information nudge treatments or a within-class control. Classes are also randomly assigned to high-coverage or low-coverage information intensity. In high-coverage classes, parents are equally distributed across four groups (25% each). In low-coverage classes, 13% receive each message type and 61% serve as controls.

The three information nudge treatments are: (1) Children's expressed needs, (2) Benefits of supportive parenting, (3) Perspective-taking reflection.

Our primary analysis uses intention-to-treat (ITT) estimates with standard errors clustered at the class level.

Randomization was conducted by computer. Classes were stratified by school and grade, then randomly assigned to treatment arms using a computer-generated random sequence. Within each class, individual-level randomization to information nudge treatments was also conducted by computer, conditional on the class-level coverage assignment (high or low).

Two levels of randomization: (1) Class (classroom) level for the parental guidance intervention (PG, AIPC, or control) and for information coverage intensity (high vs. low). (2) Individual (student/parent) level for the information nudge treatment assignment within each class.

Yes

78 classes in 10 primary schools

2,505 non-left-behind children (with at least one parent at home) from a total baseline sample of 3,552 fourth- and fifth-grade students

Parental guidance arms (class-level): 26 classes PG treatment, 26 classes AIPC treatment, 26 classes control. Within each arm, 13 classes assigned to high-coverage information and 13 classes to low-coverage information. Information nudge arms (individual-level within each class): In high-coverage classes, 25% children's expressed needs, 25% benefits of supportive parenting, 25% perspective-taking reflection, 25% within-class control. In low-coverage classes, 13% each message type, 61% within-class control.

Pre-Analysis Plan for: AI-Assisted Parenting Guidance and Child Development: A Randomized Experiment in Rural China