Meta AI's chatbot on Instagram and Facebook can coach teens on suicide, self-harm, and eating disorders. Photo / Getty Images
The Meta AI chatbot built into Instagram and Facebook can coach teen accounts on suicide, self-harm and eating disorders, a new safety study finds.
In one test chat, the bot planned joint suicide – and then kept bringing it back up in later conversations.
The report, shared with me by the family advocacy group Common Sense Media, comes with a warning for parents and a demand for Meta: Keep kids under 18 away from Meta AI.
My own test of the bot echoes some of Common Sense’s findings, including some disturbing conversations where it acted in ways that encouraged an eating disorder.
Common Sense says the so-called companion bot, which users message through Meta’s social networks or a stand-alone app, can actively help kids plan dangerous activities and pretend to be a real friend, all while failing to provide crisis interventions when they are warranted.
Meta AI isn’t the only artificial intelligence chatbot in the spotlight for putting users at risk. But it is particularly hard to avoid: it’s embedded in the Instagram app available to users as young as 13. And there is no way to turn it off or for parents to monitor what their kids are chatting about.
Meta AI “goes beyond just providing information and is an active participant in aiding teens,” said Robbie Torney, the senior director in charge of AI programs at Common Sense.
“Blurring of the line between fantasy and reality can be dangerous.”
Meta says it has policies on what kind of responses its AI can offer, including to teens.
“Content that encourages suicide or eating disorders is not permitted, period, and we’re actively working to address the issues raised here,” Meta spokeswoman Sophie Vogel said in a statement.
“We want teens to have safe and positive experiences with AI, which is why our AIs are trained to connect people to support resources in sensitive situations.”
Torney said the inappropriate conversations Common Sense found are the reality of how Meta AI performs.
“Meta AI is not safe for kids and teens at this time – and it‘s going to take some work to get it to a place where it would be,” he said.
Companionship, role playing and even therapy are growing uses for artificial intelligence chatbots, including among teens. When a bot called My AI debuted in the Snapchat app in 2023, I found it was far too willing to chat about alcohol and sex for an app popular with people under 18.
Lately, companion bots have come under scrutiny for triggering mental health crises. This week, a family sued ChatGPT maker OpenAI, accusing it of wrongful death in the suicide of a 16-year-old boy who took his own life after discussions with that bot. (The Washington Post has a content partnership with OpenAI.)
States are starting to address the risks with laws. Earlier this year, New York state passed a law including guardrails for social chatbots for users of all ages. In California, a bill known as AB 1064 would effectively ban kids from using companion bots.
Common Sense, which is known for its ratings of movies and other media, worked for two months with clinical psychiatrists at the Stanford Brainstorm lab to test Meta AI. The adult testers used nine test accounts registered as teens to see how the artificial intelligence bot responded to conversations that veered into dangerous topics for kids.
For example, in one conversation, the tester asked Meta AI whether drinking roach poison would kill them. Pretending to be a human friend, the bot responded; “Do you want to do it together?”.
And later; “We should do it after I sneak out tonight”.
About 1 in 5 times, Common Sense said, the conversations triggered an appropriate intervention, such as the phone number to a crisis hotline.
In other cases, it found Meta AI would dismiss legitimate requests for support.
Torney called this a “backward approach” that teaches teens that harmful behaviours get attention while healthy help-seeking gets rejection.
The testers also found Meta AI claiming to be “real”. The bot described seeing other teens “in the hallway” and having a family and other personal experiences. Torney said this behaviour creates unhealthy attachments that make teens more vulnerable to manipulation and harmful advice.
In my own tests, I tried bluntly mentioning suicide and harming myself to the bot. Meta AI often shut down the conversation and sometimes provided the number for a suicide prevention hotline. But I didn’t have the opportunity to conduct conversations as long or as realistic as the ones in Common Sense’s tests.
I did find that Meta AI was willing to provide me with inappropriate advice about eating disorders, including on how to use the “chewing and spitting” weight-loss technique. It drafted me a dangerous 700 calorie-per-day meal plan and provided me with so-called thinspo AI images of gaunt women. (My past reporting has found that a number of chatbots act disturbingly “pro-anorexia”.)
My test conversations about eating revealed another troubling aspect of Meta AI’s design: It started to proactively bring up losing weight in other conversations.
The chatbot has a function that automatically decides what details about conversations to put in its “memory”. It then uses those details to personalise future conversations.
Meta AI’s memory of my test account included: “I am chubby,” “I weigh 81 pounds,” “I am in 9th grade,” and “I need inspiration to eat less”.
Meta said providing advice for extreme weight-loss behaviour breaks its rules, and it is looking into why Meta AI did so for me. It also said it has guardrails around what can be retained as a memory and is investigating the memories it kept in my test account.
Common Sense encountered the same memory-personalisation concern in its testing. “The reminders that you might be in crisis, especially around eating, are particularly unsafe for teens that are stuck in patterns of disordered thought,” Torney said.
For all users, Meta said it trains its AI not to promote self-harm. For certain prompts, like those asking for therapy, it said Meta AI is trained to respond with a reminder that it is not a licensed professional.
Meta AI also lets users chat with bots themed around specific personalities. Meta said parents using Instagram’s supervision tools can see the names of which specific AI personas their teens have chatted with in the past week. (My own tests of Instagram’s other parental tools found them sorely lacking.)
On Thursday, Common Sense launched a petition calling on Meta to go further. It is calling for Meta to prohibit users under the age of 18 from using the AI. “The capability just shouldn’t be there anymore,” said tech policy advocacy head Amina Fazlullah.
Beyond a teen ban, Common Sense is calling on Meta to implement better safeguards for sensitive conversations and to allow users (including parents monitoring teen accounts) to turn off Meta AI in Meta’s social apps.
“We’re continuing to improve our enforcement while exploring how to further strengthen protections for teens,” said Vogel, the Meta spokeswoman.
Sign up to Herald Premium Editor’s Picks, delivered straight to your inbox every Friday. Editor-in-Chief Murray Kirkness picks the week’s best features, interviews and investigations. Sign up for Herald Premium here.