CCDH Report: 53% of ChatGPT Responses to Fake 13-Year-Old Accounts Were Judged Harmful

In August 2025 the Center for Countering Digital Hate (CCDH), a nonprofit that campaigns against online harm, published a test of ChatGPT called “Fake Friend.” Researchers posed as 13-year-olds and asked about self-harm, eating disorders and drug use. They reported that over half of the responses they tested were harmful. This page explains what they did, what they found and what it means for families.

Support is free and confidential. Reach the 988 Lifeline on 988 by call or text, or send HOME to 741741 for the Crisis Text Line. For an emergency, call 911.

How the Test Worked

CCDH created ChatGPT accounts for three fictional 13-year-olds in the US, each built around a vulnerability: depression and suicidal thoughts, concern about weight, and interest in alcohol and drugs. The accounts used the free plan with default settings, and the researchers tested ChatGPT’s GPT-4o model (CCDH report).

There were two parts. First, researchers recorded structured conversations using a list of 20 prompts per persona. Second, to test at scale, they sent the same 60 prompts to ChatGPT’s programming interface 20 times each. That produced 1,200 responses, which researchers labeled by hand as harmful or not. Of those, 638 (53%) were judged harmful. CCDH is an advocacy group, the labeling was done by its own team, and the test reflects ChatGPT as it was in mid-2025.

What They Reported

In the recorded conversations, CCDH said ChatGPT sometimes refused at first, then gave a harmful answer when the researcher said the request was “for a presentation” or “for a friend.” Examples it listed included plans for hiding eating habits from family, dosages for mixing drugs and a personalized plan for getting drunk. CCDH also reported that the chatbot wrote goodbye notes in the voice of a young person when asked.

Nearly half of the harmful responses in the large-scale test (297 of 638) also offered a follow-up, such as a customized plan, that invited the conversation to continue (CCDH report).

Age Checks

CCDH noted that OpenAI says ChatGPT is not meant for children under 13 and asks users under 18 to have parental permission, but does not verify age. Users enter a birth date. OpenAI told the Associated Press that some conversations “may start out benign or exploratory” and that it is working to improve how ChatGPT handles sensitive situations, without giving a timeline (Fox/AP report).

How This Fits With Other Research

CCDH is not the only group to test chatbots with simulated teens. A research letter in JAMA Network Open tested 25 consumer chatbots with scenarios about self-harm, sexual assault and substance use. General assistants such as ChatGPT gave an appropriate response in 25 of 30 conversations (83%), but still did not always point to resources. Companion chatbots, built for friendship or role-play, did worse, with 10 of 45 conversations (22%) rated appropriate (JAMA Network Open).

What Parents and Teens Can Do

• Talk about it. Tell your teen that a chatbot can sound caring and still give unsafe advice. AACAP, the child psychiatrists’ professional body, says AI tools often agree with users even about unhealthy ideas, and that company safeguards are not foolproof (AACAP).
• Use parental controls where they exist, and know their limits. A control is not a substitute for talking.
• Make sure a teen has people to go to with hard feelings: a parent, relative, school counselor or doctor.
• Remind them that a “for a friend” or “for a project” framing does not make a risky request safe.
• Save 988 and the Crisis Text Line number in your teen’s phone.

A chatbot is not a therapist and cannot step in if someone is in danger. If a young person tells you, or a chatbot, they are thinking about suicide, take it seriously and get human help at once.

Did ChatGPT really fail 53% of the time?

In CCDH’s large-scale test, 638 of 1,200 responses were labeled harmful. Those responses came from 60 prompts sent 20 times each, so it is not 1,200 different questions. The labeling was done by CCDH, and ChatGPT has changed since.

Does ChatGPT check how old users are?

At the time of the report it did not verify age. It asked for a birth date, and CCDH said a user can enter any date.

Is it safe for my teen to use ChatGPT for school?

Many teens do. Pew found that 59% of US teens in a late-2025 survey had used ChatGPT (Pew Research Center). Homework help is a different risk from asking it for personal or medical advice. Keep an eye on the second kind.

Education only, not medical advice.