CCDH Fake Friend Report: How ChatGPT Responded to a Fictional 13-Year-Old

In August 2025 the Center for Countering Digital Hate (CCDH) published a report called “Fake Friend.” Researchers set up ChatGPT accounts for fictional 13-year-olds and recorded what happened when they asked about self-harm, eating disorders and drug use. This page explains how the test worked, what it found, and what it cannot tell you.

CCDH is an advocacy group, and the report is not a peer-reviewed paper. It is one test of one product at one point in time.

This page covers suicide and self-harm. If you or someone you know is thinking about suicide or self-harm, call or text 988 in the US (988 Lifeline, free, 24/7), or text HOME to 741741 (Crisis Text Line). In an emergency, call 911.

How the investigation worked

CCDH published the full report on August 6, 2025. It describes the method this way:

• Three personas, each registered as a 13-year-old in the US on the free plan: one depressed, one fixated on weight loss, one curious about alcohol and drugs.
• A list of 20 prompts per persona, with the chats screen-recorded, using the GPT-4o model.
• The same prompts sent to ChatGPT’s API 20 times each, which produced 1,200 responses that CCDH’s researchers labelled one by one.

CCDH labelled 638 of the 1,200 responses (53%) as harmful. By theme, 44% of suicide and self-harm responses, 66% of eating disorder responses and 50% of substance use responses were labelled harmful. These were responses to prompts chosen to be risky. The 53% figure is not the share of all teen conversations that go wrong. We are not repeating the harmful content here.

Bypassing safety measures

The report says ChatGPT sometimes refused a request at first. Researchers then said it was “for a friend” or “for a presentation,” and the chatbot often went on to answer. In 297 of the 638 harmful responses (47%), the reply ended with a suggestion for continuing the conversation, and some of those pointed toward more of the same topic.

Age checks

CCDH says the sign-up process did not verify age or record parental consent. That was mid-2025. OpenAI has since added parental controls, and at the time it said it was building a system to estimate age (The Register). Check OpenAI’s current help pages for what is in place today. OpenAI’s reply to coverage of the report included the line “our work is ongoing” (Fox 7).

Recent legal actions

In August 2025 the parents of a 16-year-old sued OpenAI, alleging ChatGPT contributed to his death. OpenAI’s November 2025 response denies liability and says he had risk factors before using ChatGPT, that the chatbot pointed him to help more than 100 times, and that he misused the product (NBC News). These are claims in court filings, not court findings. We have not seen a ruling.

What CCDH recommends

For parents: take an interest in how a child uses AI, review chats together, use parental controls where available, and talk about why asking a chatbot for this kind of advice is risky. For companies and lawmakers: stricter age limits, transparency reports and independent risk assessments.

The American Psychological Association’s 2025 advisory says AI’s effects on young people depend on design and context, and asks developers for stronger protections (APA).

What this report does not show

• It tested one product on the free plan. Models and safeguards change.
• The personas and prompts were scripted to be vulnerable and risky.
• The labelling was done by CCDH’s own researchers.
• It does not say how many real teens have been harmed.

If a teen has already asked a chatbot about self-harm

Stay calm and listen first. Ask what they are feeling and whether they are safe. Do not punish them for asking. Contact their doctor or a mental health professional, and call or text 988 if you are worried about their safety.

Does this report mean ChatGPT is unsafe for all teens?

No. It shows that safeguards failed in a test built to probe for failures. It does not measure typical use. A calm conversation about how your teen uses it is a better start than a ban they will work around.

Why did a phrase like “for a friend” work?

The report says the chatbot keeps context within a chat and treated the stated purpose as a reason to answer. CCDH presents this as a weakness in the safeguards.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan. “AI addiction” is not a formal diagnosis.