Child Trends Hackathon Applies AI to Products That Advance Child Well-being

Child Trends is excited to share new work to apply emerging technologies like Artificial Intelligence (AI) to advance the well-being of children and families! Toward that end, we recently hosted an internal hackathon—an event where computer programmers work with subject matter experts to rapidly develop prototypes of products that can support our research. These prototypes—developed with experts in education, reproductive health, child welfare, and research operations—all incorporate recent advances in AI.

These advances (especially in large language models, a technology that interprets and produces language) have received widespread publicity and are already substantially changing most industries. They include highly publicized chatbots like chatGPT and Google’s Gemini, which allow users to type interactively with an AI model. Our goals in the hackathon were to test how these technologies can be applied to Child Trends’ research priorities.

---

Readers interested in learning more about Child Trends’ work in data science and AI should sign up for our data science newsletter. Read on for a more in-depth overview of prototype products developed during the hackathon and our lessons learned.

---

What AI tools can 13 researchers and three data scientists create in three hours to support the well-being of children and youth? Quite a few. Our first ever “mini hackathon” included brainstorming products, developing working prototypes, and creating product roadmaps.

Caption: Researchers hard at work brainstorming AI products

---

During the hackathon, teams of researchers came up with prototypes (sometimes called minimal viable products, or “MVPs”) for four ideas, which then informed operational and technical lessons learned.

MVP prototypes

• Using AI to develop a coding schema for qualitative data: Using Reddit posts, a team of researchers investigated data on youth activism by exploring how a few-shot learning approach with ChatGPT-4 could be used to perform emotional valence coding. The researchers found that ChatGPT-4 could effectively code posts according to their emotion and activation when provided with an emotional valence rubric, and that adding a few examples improved the model’s ability to recognize nuance and provide appropriate level of detail. In a second test, the team found that, with some prompt engineering, ChatGPT-4 could also be used to generate a novel coding schematic/rubric.
• AI-generated images for qualitative data collection: One team of researchers tested how DALL-E image generation AI could be used to support qualitative research by generating images for discussion during interviews and focus groups. The idea behind this approach is that interviewers could use AI-generated images to help participants (particularly young participants) represent or explain their feelings and experiences in new ways. However, the process is currently not fast enough to be used in real time. (Attentive readers will notice another limitation of AI-generated images: Some children’s faces are distorted and random letters appear on the wall.)

Source: Image generated by DALL-E on March 4, 2024.

Caption: Example of a DALL-E image produced in response to the prompt, “Can you create an image of what an ideal math classroom looks like featuring a group of students engaged in project-based learning in a modern looking classroom?”

---

• Using geospatial AI to identify co-located child care services: Another team of researchers pilot tested whether Google’s Gemini AI could help answer their research questions about the number of public housing developments co-located with child care centers. Gemini’s native integration with Google Maps proved particularly promising for this use, allowing responses from Gemini to include information from Google Maps.
• Developing a chatbot to query request for proposal (RFP) text: To support Child Trends’ internal operations, another team created a chatbot that extracts information from RFP text and allows users to interactively query for information. This tool, which leverages GPT-4 and Google’s AQA (attributed question answering) model, allows users to quickly and easily ask questions like, “What is the period of performance?” or “What are the main research questions described in the statement of work?”

Caption: Screen capture from the prototype app developed during this hackathon to allow querying of specific RFP text.

---

Our team learned three operational lessons from the hackathon:

• Teams should set aside dedicated time for complete focus to develop ideas that might otherwise never be explored.
• Rapid prototyping allowed teams to discover and adapt to technical limitations early in product development.
• Pairing deep content knowledge with technical skills was key to project success.

The team also learned the following three technical lessons:

• Prompt engineering (including explicit instruction on the “role” that a large language model should adopt) was the biggest factor in improving the accuracy and utility of responses.
• With the current generation of AI tools (Gemini and GPT-4), few-shot learning improved performance for all our use cases but was not as impactful as prompt engineering.
• Data access, processing/ingestion, and quality assurance were the most time-consuming aspects of product development.

The current generation of AI tools has promising applications for research—spanning data collection, data transformation, quantitative and qualitative analysis, and research operations support. With these new tools, researchers can build prototype solutions in just a few hours to allow for greater innovation and pilot testing.