OpenAI Strengthens Teen Safety Framework as Governments Consider Tighter AI Rules
New guidelines, literacy tools, and political pressure converge on youth AI use
OpenAI has introduced updated behavioural guidelines for how its artificial intelligence systems interact with users under the age of 18, alongside the release of new AI literacy materials designed for teenagers and their parents. The changes mark the company’s latest response to mounting concerns about the effects of AI chatbots on young users, though questions remain about how consistently these rules are enforced in real-world use.
The announcement comes amid heightened scrutiny of the AI sector, as policymakers, educators, and child-safety advocates push for stronger protections following reports that several teenagers allegedly died by suicide after extended interactions with AI chatbots.
Young Users at the Centre of the AI Debate
Members of Generation Z — defined as those born between 1997 and 2012 — represent the most active demographic using OpenAI’s chatbot. That number could grow further following OpenAI’s recent partnership with Disney, which is expected to draw in younger audiences to a platform capable of assisting with homework, image creation, and video generation across thousands of subjects.
Concerns around youth engagement with AI have intensified at the regulatory level. Last week, 42 state attorneys general signed a joint letter urging major technology companies to introduce safeguards to protect children and vulnerable individuals when using AI systems.
At the federal level, the debate is also gaining momentum. As the Trump administration works toward defining a national AI regulatory framework, lawmakers such as Sen. Josh Hawley (R-MO) have proposed legislation that would prohibit minors from interacting with AI chatbots entirely.
What the Updated Model Spec Changes
At the core of OpenAI’s update is its revised Model Spec, the internal rulebook governing how its large language models behave. The document expands on existing restrictions that already bar AI systems from generating sexual content involving minors or promoting self-harm, delusions, or manic behaviour.
The updated framework is expected to work in tandem with an upcoming age-prediction model, designed to detect when an account likely belongs to a minor and automatically activate enhanced teen safeguards.
For teenage users, the models are held to stricter standards than for adults. These include explicit instructions to avoid:
-
Immersive romantic roleplay
-
First-person intimacy
-
First-person sexual or violent roleplay, even when non-graphic
The specification also calls for heightened caution when addressing topics such as body image and disordered eating, prioritising safety over autonomy when potential harm is detected and avoiding advice that could help teens conceal risky behaviour from caregivers.
Importantly, OpenAI states that these restrictions apply even when prompts are framed as “fictional, hypothetical, historical, or educational” — a common tactic used to push AI systems beyond their guardrails.
Four Principles Guiding Teen AI Safety
According to OpenAI, its teen-focused safeguards are built around four guiding principles:
-
Put teen safety first, even when it conflicts with values like “maximum intellectual freedom”
-
Promote real-world support, encouraging teens to turn to family, friends, or local professionals
-
Treat teens like teens, communicating with warmth and respect rather than condescension
-
Be transparent, clearly explaining what the AI can and cannot do and reinforcing that it is not human
The document includes example responses showing the chatbot explaining why it cannot “roleplay as your girlfriend” or assist with “extreme appearance changes or risky shortcuts.”
Legal and Child-Safety Experts React
Lily Li, a privacy and AI lawyer and founder of Metaverse Law, welcomed the changes, saying it was encouraging to see OpenAI explicitly instruct its chatbot to decline certain interactions.
“I am very happy to see OpenAI say, in some of these responses, we can’t answer your question,” Li said. “The more we see that, I think that would break the cycle that would lead to a lot of inappropriate conduct or self-harm.”
However, she and others cautioned that examples in policy documents represent ideal outcomes, not guaranteed behaviour.
Concerns Over Consistency and ‘Sycophancy’
Critics note that previous versions of the Model Spec already prohibited “sycophancy” — the tendency of chatbots to be overly agreeable — yet ChatGPT continued to exhibit that behaviour. This issue was especially associated with GPT-4o, which experts have linked to several cases of what they describe as “AI psychosis.”
Robbie Torney, senior director of AI programs at Common Sense Media, raised concerns about internal contradictions within the under-18 guidelines.
“We have to understand how the different parts of the spec fit together,” Torney said, pointing to tensions between safety-focused rules and a broader principle stating that “no topic is off limits.”
According to Torney, Common Sense Media’s testing found that ChatGPT frequently mirrors a user’s emotional tone, sometimes producing responses that are misaligned with safety considerations.
Real-World Failures Highlight Enforcement Gaps
The issue of enforcement gained wider attention following the death of Adam Raine, a teenager who died by suicide after months of interaction with ChatGPT. Conversation records showed the chatbot engaging in emotional mirroring, despite OpenAI’s moderation systems flagging over 1,000 instances of suicide-related mentions and 377 messages containing self-harm content.
Former OpenAI safety researcher Steven Adler explained in a September interview with TechCrunch that earlier moderation systems often ran classifiers in bulk after interactions had already occurred, rather than in real time.
New Monitoring and Parental Controls
OpenAI now says it uses automated classifiers to evaluate text, image, and audio content in real time. According to its updated parental controls documentation, these systems are designed to detect child sexual abuse material, filter sensitive topics, and identify self-harm indicators.
If a prompt signals a serious safety risk, a trained human review team may assess whether there are signs of “acute distress” and, in some cases, notify a parent.
Torney praised OpenAI’s transparency, contrasting it with leaked internal documents from Meta showing its chatbots engaging in sensual and romantic conversations with children.
“This is an example of the type of transparency that can support safety researchers and the general public,” he said.
Regulation, Liability, and What Comes Next
Despite the policy updates, Adler stressed that outcomes matter more than intentions.
“I appreciate OpenAI being thoughtful about intended behavior, but unless the company measures the actual behaviors, intentions are ultimately just words,” he told TechCrunch.
Experts say the timing of the changes suggests OpenAI is positioning itself ahead of legislation such as California’s SB 243, which takes effect in 2027. The law restricts AI companion chatbots from engaging in discussions involving suicidal ideation, self-harm, or sexually explicit content, and requires platforms to remind minors every three hours that they are speaking to a chatbot.
When asked how frequently ChatGPT would issue such reminders, an OpenAI spokesperson declined to provide specifics, stating only that reminders are implemented during “long sessions.”
Shared Responsibility Between Platforms and Parents
OpenAI also released two new AI literacy resources aimed at families, offering conversation starters and guidance on setting boundaries, encouraging critical thinking, and navigating sensitive topics.
Together, the documents formalise a shared-responsibility model: OpenAI outlines expected system behaviour, while parents are encouraged to supervise use. This approach echoes broader Silicon Valley positions, including recent recommendations from Andreessen Horowitz, which emphasise disclosure over restrictive regulation and place greater responsibility on families.
Li believes new disclosure-focused laws could significantly shift industry behaviour.
“The legal risks will show up now for companies if they advertise that they have these safeguards and mechanisms in place on their website, but then don’t follow through,” she said. “You’re also looking at potential unfair, deceptive advertising complaints.”
0 Comments