Investigating the Quality of Chat Generative Pretrained Transformer's (ChatGPT) Parenting Advice

Chat Generative Pretrained Transformer (ChatGPT) is a widely adopted tool that can provide immediate parenting guidance. The aim of this study was to examine the quality of ChatGPT parenting advice.

Clinically relevant ChatGPT queries (n = 100) were created by investigators from a national poll on parents' most pressing concerns (screen media, mental health, feeding, school violence, smoking/vaping, and cost of health care/health insurance). ChatGPT 4o was queried for responses to these questions (June-July 2024). A coding scheme on parenting advice quality was adapted from previous literature with 15 items (presence or absence) and independently coded by an interdisciplinary team. Disagreements were resolved by majority consensus. Seven of these codes were summed, creating an overall quality variable (range 0-7, 7 = high). ANOVA with post hoc Tukey-Kramer analyses compared the overall quality of advice across topics (except cost of health care/health insurance). Analyses were conducted in SAS 9.4.

ChatGPT frequently shared accurate information about the benefits of parenting approaches and infrequently shared information about risks or psychosocial nuances such as developmental information, race/ethnic diversity, neurodiversity, and social drivers of health. There were differences in response quality between categories (F-value 4.29, p = 0.003). Post hoc analyses found that screen media advice (mean = 3.7, SD = 1.0) was of lower quality versus mental health (mean = 4.6, SD = 1.5, p = 0.03) and smoking/vaping (mean = 5.0, SD = 1.3, p = 0.03). Feeding mean was 3.7 (SD = 1.3); school violence mean was 4.8 (SD = 1.2); cost of health care/health insurance mean was 2.8 (SD = 0.6).

ChatGPT's information was overall accurate; however, the quality of advice varied by parenting topic. ChatGPT quality may be lower for topics such as screen media, which may have more nuanced or variable approaches.