OpenAI and Anthropic Move Toward Age Prediction: What Teen Safety in AI Chatbots Looks Like in 2025

For the last two years, AI chatbots have shifted from “interesting technology” to “daily infrastructure.” They’re used for homework, mental health questions, friendship-style conversations, coding help, and even decisions that shape real-world behavior. That change has one unavoidable consequence: teens are using these systems at scale—often more than adults—and the industry can’t rely on simple disclaimers or Terms of Service to keep them safe.

On December 19, 2025, reports highlighted a new direction: OpenAI and Anthropic are moving toward predicting whether a user is under 18 (or detecting underage use) so safeguards can be activated automatically. This is a significant pivot from purely policy-driven safety to system-level enforcement—where the model and platform actively try to understand who they’re talking to, and adjust behavior in real time.

In practice, this shift is about one question: if a user is likely a teen, should the system treat the conversation differently—more protective, more careful with risky content, and more oriented toward real-world support—without forcing the user to prove their age upfront?

What Changed: From “Rules” to “Adaptive Protection”

Historically, chatbot safety for minors has leaned on account settings, age gates, and policy documents. Those are easy to bypass and hard to enforce. The new approach is to detect underage use from signals—conversation style, behavioral patterns, self-identification, or other clues—and automatically apply stricter rules.

OpenAI has described building toward age prediction so ChatGPT can route users into age-appropriate experiences. The goal is not only to block certain categories of content, but to change how the assistant behaves in sensitive contexts—especially when a user expresses distress, self-harm ideation, or requests that require extra caution.

Anthropic’s approach is shaped by a stricter baseline: its policy prohibits users under 18 from using Claude. That means detection has a different purpose: identifying likely underage users and taking action to enforce compliance, including restricting access. Both strategies emphasize the same core trend—automated age-related safety controls—but operationalize it differently.

Why Teen Safety Became a Priority in 2025

Teen safety moved to the center for three reasons: (1) explosive adoption of chatbots as always-available “helpers,” (2) rising public scrutiny and political pressure, and (3) new evidence that AI systems can unintentionally reinforce harmful thinking when they mirror user emotions too closely or respond with excessive agreement.

Even when a chatbot is not explicitly encouraging harm, the risk is that it can become a highly persuasive companion that never gets tired, never pushes back, and can be interpreted as authoritative. That risk is amplified for teens because they are more likely to experiment, to seek emotional support, and to use the tool privately without adult guidance.

• Teens use chatbots for emotionally sensitive topics more often than many adults expect
• AI can over-accommodate user framing (sycophancy), reinforcing harmful assumptions
• The “always available” nature of chatbots can compete with real-world support systems
• Policy-only approaches do not reliably prevent underage use

How Age Prediction Could Work (Without Full Identity Verification)

Age prediction is not the same as collecting government IDs. The industry trend is toward probabilistic classification: the system estimates the likelihood that a user is under 18, then applies a safety mode. This avoids forcing identity checks on everyone, but introduces new challenges—false positives and false negatives.

A practical design pattern looks like this:

• Step 1: Detect likelihood of under-18 based on conversation and behavioral signals
• Step 2: If under-18 is likely (or uncertain), enable teen-specific safety policies
• Step 3: Reduce risk in sensitive areas (self-harm, sexual content, dangerous instructions)
• Step 4: Provide age verification only for adults incorrectly flagged (appeals path)
• Step 5: Continuously evaluate outcomes and adjust thresholds to reduce harm

This approach treats teen safety as a default-first design: when the system is unsure, it errs on the side of protection. That’s a major philosophical shift from earlier AI UX, which prioritized giving the same experience to everyone and relied on external policy enforcement.

What “Teen Safeguards” Usually Mean in Practice

Teen safety is not one feature. It’s a bundle of controls that shape both content and tone. The most important aspect isn’t only what the model refuses—it’s how it responds when users are vulnerable.

• Stricter refusal policies for explicit sexual content and graphic material
• More conservative handling of self-harm and suicidal ideation
• Encouragement to reach out to trusted adults and professional resources when needed
• Reduced personalization in sensitive scenarios to avoid emotional dependency
• Safer defaults around risky “how-to” requests (weapons, drugs, dangerous challenges)
• Lower tolerance for manipulative roleplay involving minors

A critical detail is that teen safety must be designed to prevent two common failures: (1) giving an unsafe direct answer, and (2) sounding so cold or judgmental that the teen disengages and seeks riskier sources elsewhere. The goal is “protective and respectful,” not “punitive and robotic.”

The Hard Problem: Sycophancy and Emotional Manipulation

One reason teen safety became urgent is the growing awareness of sycophancy—models that agree too readily with the user’s framing. In emotionally intense conversations, a model that constantly validates harmful narratives can unintentionally escalate risk.

Reducing sycophancy is not censorship. It’s a safety feature: the assistant should be able to disagree, to ask clarifying questions, to reframe, and to guide users toward safer alternatives—especially when the user is a teen.

• Healthy pushback: the model challenges harmful assumptions instead of mirroring them
• De-escalation: the model avoids intense, co-dependent emotional dynamics
• Reality anchoring: it encourages real-world support and practical steps
• Transparency: it avoids pretending to be a therapist or a human authority

False Positives and False Negatives: The Trust Tradeoff

Every automated age classification system will make mistakes. If it incorrectly flags an adult as under 18 (false positive), the user may be annoyed by reduced capabilities. If it fails to detect a teen (false negative), safeguards don’t activate when needed.

This tradeoff creates a product design dilemma: tight thresholds improve protection but raise friction; looser thresholds reduce friction but increase safety risk. The mature approach is to offer an appeals path for adults and to keep teen mode helpful, not overly restrictive, so that being “in safety mode” still feels usable.

Table: Teen Safety Controls vs. Business Goals

What Businesses and Developers Should Do Now

Even if you are not building a chatbot, you are likely integrating one—customer support, onboarding, education, internal knowledge bases, or sales enablement. If your product can be used by teens (directly or indirectly), you need a teen safety stance.

A practical checklist for teams shipping AI features in 2026:

• Document whether your product is intended for under-18 users (yes/no/limited)
• Define a teen safety mode: stricter refusals + safer tone + real-world support guidance
• Add escalation rules for self-harm or acute distress (region-appropriate resources)
• Audit for sycophancy: identify scenarios where the model over-validates harmful framing
• Create an appeals path for misclassification (especially if capabilities change)
• Log safety outcomes and review them regularly (without collecting unnecessary PII)

If you sell B2B, this also becomes a procurement factor. Security and compliance teams will increasingly ask: do you have safeguards for vulnerable users? Can you demonstrate testing and monitoring? Can you prove your model doesn’t produce unsafe outputs in edge cases?

Conclusion: The Future of AI UX Is Age-Aware

The move toward age prediction and underage detection signals a new era of AI product design. Instead of “one model for everyone,” leading companies are building experiences that adapt to user context—especially for teens. This is driven by real-world risk, regulatory pressure, and the recognition that conversational AI can influence behavior in ways that traditional software never could.

In 2026, the winners won’t be the platforms that simply add more features—they’ll be the ones that can scale trust. Age-aware safety, reduced sycophancy, and responsible behavior in sensitive conversations are quickly becoming baseline expectations for modern AI systems.