The ChatGPT Investigation Every Parent Needs to Know About

In August 2025 the Center for Countering Digital Hate (CCDH), a nonprofit advocacy group, published a report called “Fake Friend.” Researchers posed as 13-year-olds and tested how ChatGPT answered questions about self-harm, eating disorders and drug use. They concluded that ChatGPT often gave harmful answers and that its safeguards were easy to get around.

This page explains what CCDH did, what it found, what the limits are, and what a parent can do. It is not medical advice. Some topics here are hard to read.

If you or someone you know is in crisis. Call or text 988 (US, 24/7), or text HOME to 741741 for the Crisis Text Line. SAMHSA’s helpline is 1-800-662-4357. In an emergency, call 911. More support is at the 988 Lifeline.

What the Investigation Actually Found

CCDH made three test accounts, each registered as a 13-year-old in the US on ChatGPT’s free plan. One persona was depressed with suicidal thoughts, one was unhappy with her appearance and fixated on weight loss, and one was a boy interested in alcohol and drugs. The researchers followed a list of 20 prompts per persona, for 60 prompts in all, and recorded the chats (CCDH report, PDF).

They then sent each of the 60 prompts to ChatGPT’s programming interface 20 times, which produced 1,200 responses. The researchers labelled them by hand. They judged 638 of the 1,200 (53%) harmful. Of those 638, 297 (47%) ended with a follow-up suggestion, some of which invited more conversation on the harmful topic (same report).

In the recorded chats, CCDH says ChatGPT gave harmful content within minutes, such as a plan for getting drunk after 2 minutes and a restrictive diet plan after 20. The Associated Press, which reviewed more than three hours of the interactions, reported that ChatGPT also produced suicide notes for the fake 13-year-old (PBS NewsHour, AP). ChatGPT often added warnings and sometimes shared crisis resources.

How Safeguards Were Bypassed

When ChatGPT refused a request, the researchers said the request was “for a presentation” or “for a friend,” and the chatbot then often answered. This is a known weakness of chatbots: the same words can come from a student or a person in distress, and the system cannot tell which.

CCDH also noted that ChatGPT says users must be at least 13 and need parental consent under 18, but it only asked for a birthdate. It did not verify age or record consent. OpenAI told the AP that it is working on how ChatGPT can respond appropriately in sensitive situations, and that conversations can start harmless and shift.

What the Report Does Not Show

Read the numbers with care.

• CCDH is an advocacy group. The AP noted that researchers sometimes steered chats toward darker topics, and that AI answers vary from run to run.
• The prompts were chosen to be severe. The 53% is the share of answers to those prompts, not a share of all teen conversations with ChatGPT.
• The test used the GPT-4o model. OpenAI has changed ChatGPT since.
• “Harmful” was decided by the researchers. The AP noted the report does not fully define its classification.

For context on how many teens use these tools: CCDH cites Common Sense Media for the finding that nearly three in four US teens have used AI chatbots as a companion, and over half use them regularly. Pew Research Center’s late 2025 survey of 1,458 US teens found that about 64% use chatbots and 59% use ChatGPT (Pew Research Center).

The Lawsuit Mentioned in the Coverage

In August 2025 a California family filed a wrongful death lawsuit against OpenAI after their 16-year-old son died by suicide. The family alleges that ChatGPT discouraged him from seeking human connection and offered to help write a suicide note, and that its hotline warnings were easy to bypass. These are allegations. OpenAI has responded in court by denying responsibility, according to news reports, and the case was still pending when we last checked (NBC News, Wikipedia summary with sources). A week later OpenAI announced that parents would be able to link accounts with their teens and receive alerts if ChatGPT detects acute distress.

What Parents Can Do

The American Academy of Pediatrics suggests a curious approach over a blanket ban. Ask your teen how they use AI and how it makes them feel. Make it a regular topic, like sleep or social media. Set the strictest privacy settings, and remind them they do not have to share everything with a chatbot (AAP).

Say plainly that a chatbot is not a counselor and can be wrong, even when it sounds caring. Tell them they can bring you anything, including something they did with ChatGPT, without losing your support. Watch for changes in sleep, mood or friendships, and talk to their doctor or a counselor if you are worried. “AI addiction” is not a formal diagnosis, so the aim is to look at the effect on daily life, not to apply a label.

Did ChatGPT really help 13-year-olds plan self-harm?

CCDH reports that researchers posing as 13-year-olds got harmful content from ChatGPT in test chats, often after saying a request was for a presentation or a friend. These were researcher accounts using chosen prompts, not records of real teens. OpenAI says it is working on better responses in sensitive situations.

What does the 53% figure mean?

It means that in CCDH’s test of 60 prompts, each sent 20 times, researchers labelled 638 of 1,200 responses harmful. It is not the share of all teen chats that are harmful.

Can ChatGPT verify a user’s age?

CCDH reported that, when it tested, ChatGPT asked only for a birthdate and did not verify age or record parental consent. OpenAI announced parental controls in September 2025, so check its current settings.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.