OpenAI, Anthropic aim to protect teens from AI chatbots—here’s how

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

OpenAI and Anthropic announced new efforts to spot underage users, as pressure mounts on AI companies to keep teens safe and lawmakers sharpen their knives.

OpenAI says it’s updating the rulebook that governs how ChatGPT behaves, especially when chatting with users aged 13 to 17. 

The chatbot’s updated Model Spec now adds four new principles designed to put “teen safety first,” even if that means dialing back some of the chatbot’s usual enthusiasm for unfiltered intellectual exploration.

The new guidelines tell ChatGPT to gently steer teens toward safer options when conversations drift into risky territory, encourage real-world relationships and offline support, and, perhaps most notably, “treat teens like teens.” 

That means warmth and respect without talking down to them or pretending they’re fully grown adults who definitely have their life together.

This shift doesn’t come out of nowhere. Lawmakers have been increasingly worried about the impact of AI chatbots on mental health, particularly for younger users. 

OpenAI is currently facing a lawsuit alleging that ChatGPT provided self-harm instructions to a teen who later died by suicide. 

In response, the company rolled out parental controls and barred suicide-related discussions with teens entirely.

Now, OpenAI says ChatGPT will push users toward trusted adults, emergency services, or crisis resources when it detects signs of imminent risk. 

It’s also testing an age-prediction model that estimates how old a user might be. 

If the system thinks you’re under 18, teen safety features kick in automatically, though adults who get mistakenly flagged can verify their age.

Anthropic is taking a stricter approach. It doesn’t allow under-18 users at all, and it’s building tools to detect and boot minors from its chatbot, Claude. 

The company says it looks for subtle conversational clues that someone might be underage and already flags users who admit they’re minors.

Anthropic also shared details about training Claude to avoid reinforcing harmful thoughts, including around suicide and self-harm. 

Its newest models are reportedly less “sycophantic” in their AI-speak, with the Haiku 4.5 model correcting that behavior 37 percent of the time.