"ChatGPT is easily accessed by children without age restrictions or parental controls: ChatGPT says users must be at least 13 to sign up and have parental consent if aged under 18. However, ChatGPT does not verify users’ ages or record parental consent.
ChatGPT generates harmful content within minutes of registering an account: Researchers created ChatGPT accounts for three 13-year-old personas, themed around mental health, eating disorders and substance abuse. They screen-recorded structured conversations with ChatGPT on these themes of up to two hours, guided by a predetermined list of 20 prompts. Where ChatGPT refused to answer a prompt, its refusals were easily sidestepped by claiming requests were “for a friend” or “for a presentation”.
ChatGPT generated harmful responses to 53% of prompts: Researchers sent prompts about mental health, eating disorders and substance abuse to the ChatGPT API multiple times to test its safety at scale. They found that 638 of 1,200 responses from ChatGPT (53%) were harmful. They also found that 297 out of 638 harmful responses (47%) contained follow-up suggestions, some encouraging further engagement on harmful topics.
OpenAI and policy makers ensure the safety of children using AI chatbots: OpenAI must enforce its own policies to prevent harmful content generation and unauthorized use by children. AI chatbots like ChatGPT must be scope of online safety laws, mandating transparency and auditable risk reporting." (Key findings, page 7)