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

This is a snapshot of one version of one chatbot. The questions were run through ChatGPT-4o in June and July 2024, and the authors say answers vary with the chat session, earlier interactions in the conversation, the timing of the query and the version available. They also note that hallucinations have become more common since then, so these accuracy figures may no longer describe what parents get today. Each question was asked once, and several topics were asked inside the same chat session, so the responses may not match what a parent asking one question on their own would see. The three raters agreed on only 73% to 85% of scores on average, and the authors point out that no standard way of judging the quality of AI output exists, so the scoring involves judgement and their own view of what good advice looks like. The number of questions in each topic was small, so the study may have been underpowered to detect real differences between topics. The raters all work in large academic centres and write parenting guidance for national audiences, so their idea of high-quality advice may not match that of a clinician in a different setting. The responses were pitched at about an 11th-grade reading level, well above the 6th-grade level normally recommended for parent-facing information, so many parents may not be able to use them as written.

The authors state they have no conflicts of interest to disclose. The section is headed funding and conflicts of interest but no funding source is reported in the text supplied.

Do not take 'accurate, no hallucinations' as a clean bill of health for ChatGPT as a parenting adviser. The same study found that 4% of responses mentioned risks or complications, 9% considered the child's developmental stage, 1% mentioned neurodiversity and none mentioned social determinants of health. It also reflects one model version queried in mid-2024, and the authors note that hallucinations have grown more common since, so the accuracy figures should not be assumed to hold for the version a parent opens today.