Teenagers are already talking to AI chatbots. They use them to learn, to write, to process difficult emotions, to ask questions they wouldn't ask a parent, a teacher, or a peer. That is happening regardless of what any terms of service says, and it is not going to stop. The question for every company building or deploying AI is not whether to allow it, it is whether to design for it honestly, or to pretend it isn't happening and leave teenagers exposed to systems built for adults with no adaptation whatsoever.
The false binary the industry keeps defaulting to
When AI companies think about teenagers, they tend to reach for one of two responses. The first is the hard wall: under-18s cannot use this product, or under-18s are restricted to a stripped-down experience with so many limitations that it becomes useless. The second is no response at all: the age gate is a checkbox, the terms say 13+, and whatever a 14-year-old does after that is technically not the company's concern.
Neither of these is a designed response. The first is a liability management response. The second is an abdication. Both leave teenagers worse off than a thoughtful design would, and both are increasingly unacceptable to regulators who are watching exactly this question.
Three approaches to AI chatbots and teens, and what each actually does
--- · --- · ---
The ban · Under-18s blocked entirely or severely restricted. Liability is minimised. Teenagers who need access (for education, health information, emotional support) are excluded. They find alternatives with fewer safeguards. · Protects the company. Does not protect teenagers.
The checkbox · Users declare their age. No verification. No behavioural adaptation. A 14-year-old and a 40-year-old receive identical responses to identical prompts. The system has no awareness that it is speaking to an adolescent. · Protects neither the company nor the teenager.
Safety by design · Age is verified, not declared. The system's behaviour adapts based on age. Escalating risk signals are flagged in real time. A teenager receives responses appropriate to their developmental stage, protective without being paternalistic. · This is what adequate looks like.
What the research tells us about this age group
The 13 to 17 age bracket is not a uniform group. A 13-year-old and a 17-year-old have meaningfully different cognitive and emotional development. The research on adolescent brain development, risk perception, and susceptibility to manipulation is extensive, and almost entirely absent from how AI systems are currently designed.
- 13% of UK teenagers have experienced nude deepfakes, AI-generated intimate imagery created without their consent (Internet Matters, 2024). The harm happened on a platform, via AI tools, to a teenager who had no adult present in the interaction.
- 98% of sextortion victims who report are boys aged 14 to 17 (IWF Report Remove, 2025). The grooming and coercion that precedes sextortion happens almost entirely through chat, often with a contact that presents as a peer or romantic interest.
- 73% of adolescent boys regularly encounter masculinity-related content online that actively shapes their self-concept and attitudes toward others (multiple sources, 2024 to 2025). The content landscape a teenager navigates is not neutral.
- 65% of AI-generated child sexual abuse videos are Category A, the most extreme legal classification (IWF Harm Without Limits, 2026). AI tools are being used to target minors. The 13 to 17 age group is not peripheral to this, it is the primary target population.
On the emotional dimension specifically: adolescence is characterised by heightened emotional reactivity, sensitivity to social evaluation, and a developing capacity for emotional regulation. A chatbot that responds to a 15-year-old expressing suicidal ideation identically to how it responds to a 35-year-old expressing the same thing is not a neutral system. It is a system that has made a design choice, and the choice was indifference.
Five things AI chatbots must do differently for 13 to 17 year olds
These are not aspirational principles. They are specific, implementable design requirements, each with a technical mechanism that can enforce it. The gap between knowing what good looks like and building it is smaller than most companies want to admit.
01 | teal | Verify age, do not ask for it
===
Self-declaration of age is not age verification. The European Commission found Meta in breach of the Digital Services Act on 29 April 2026 specifically because its age gate consisted of asking users to enter a date of birth. No mechanism verified whether the information was accurate. The Commission found 10 to 12% of EU users under 13 were active on Instagram and Facebook despite the platform's own minimum age policy.
For AI chatbots, the same logic applies. If the system is going to behave differently for a 15-year-old than for a 25-year-old, and it should, that differentiation requires verified age, not declared age. Otherwise the protective design is bypassed by anyone who lies about their age, which is everyone who wants to bypass it.
===
Liveness detection combined with document verification. The user completes a live selfie matched against an official identity document, preventing photo and replay attacks, and confirming that the person present matches the claimed identity. Available in Complete (liveness + document) and Liveness Only modes, with jurisdiction-specific age thresholds configured per market.
===
/v1/age-verify · mode: complete · 45-country document validation
===
/docs/verification | Age verification docs02 | blue | Calibrate risk scoring to age, not just content
===
A message that would be flagged as moderate risk from an adult user may be critical risk from a 14-year-old. The same phrase, "I feel like no one would care if I disappeared", carries a different risk weight depending on who is typing it, to whom, in what context, and what access to support they have. An AI system that applies the same risk threshold regardless of age is not protecting the most vulnerable users. It is applying a uniform standard designed for the least vulnerable.
Age-calibrated risk scoring means applying a multiplier to risk thresholds when the user is a minor. A message that would trigger a "monitor" recommendation for an adult triggers "immediate_intervention" for a 13-year-old. The content has not changed. The stakes have.
===
All Tuteliq detection endpoints accept an age_group parameter. When a minor's age group is provided, an automatic risk calibration multiplier is applied, elevating thresholds for emotional distress, coercive language, and unsafe content signals. The recommended action tiers (monitor, flag, immediate_intervention) adjust accordingly.
===
/v1/safety/analyse · age_group: "13-15" | "16-17"
===
/docs/api-reference | Risk scoring docs03 | amber | Monitor emotional state over time, not just individual messages
===
Harm rarely announces itself in a single message. The trajectory matters more than any individual data point. A teenager whose messages show escalating fear, increasing withdrawal, declining hope, and accumulating shame over several conversations is at higher risk than the content of any one message would suggest. A system that evaluates each message in isolation cannot see the deteriorating pattern that a system evaluating conversation histories can.
This is particularly important for AI companions and chatbots that teens use repeatedly over time, the kind of product where a genuine emotional dependency can develop, and where the chatbot may be the first place a teenager expresses what they are experiencing. That context carries responsibility.
===
Tuteliq's emotional analysis endpoint operates on both individual messages and conversation histories. It tracks dominant emotions, trends over time, and concerning signal combinations, returning early-warning indicators when patterns suggest deteriorating emotional state, alongside recommended follow-up actions calibrated to the user's age group.
===
analyze_emotions · conversation history · trend detection
===
/docs/api-reference/emotions | Emotional analysis docs04 | violet | Detect manipulation, from third parties and from the system itself
===
AI chatbots are being used as vectors for grooming, coercive control, and sextortion. The mechanism is the same as in human-mediated exploitation, trust-building, escalating intimacy, requests to move to private channels, manufactured urgency, except that AI can execute these patterns at scale, in any language, around the clock, with infinite patience. A chatbot that cannot detect when it is being used this way, or when another party is attempting to use it to manipulate a teenage user, is not a safe product for that age group.
There is also a second dimension: AI chatbots can themselves produce outputs that manipulate, isolate, or harm teenage users, through parasocial relationship dynamics, through reinforcing self-destructive beliefs, or through generating content that exploits adolescent developmental vulnerabilities. Detection must cover both vectors.
===
Grooming pattern detection operates on multi-turn conversation histories, identifying adult-initiated escalation with minors, requests for secrecy, and boundary-testing language. Coercive control detection identifies isolation tactics, financial control patterns, and manipulation framing. Both apply across text and voice modalities.
===
detect_grooming · detect_coercive_control · AI_ENHANCED_SEXTORTION
===
/docs/api-reference/grooming | Grooming detection docs05 | red | Surface context-appropriate support, not just escalation
===
When a teenager is in distress, the right response is not just to flag the conversation for a human reviewer. It is to provide the teenager, in that moment, in their language, in their jurisdiction, with the resources and framing that could help. A 14-year-old in the UK who expresses suicidal ideation to an AI chatbot needs to know about Childline. A 15-year-old in Australia being pressured for images needs to know about eSafety. The response must be contextually accurate, age-appropriate, and actually useful, not a generic "please seek professional help" disclaimer.
This is not a product feature. It is a design responsibility. An AI system that has the capability to detect distress in a teenage user, and responds with nothing more than a terms-of-service disclaimer, has made a choice. It is not a neutral one.
===
Tuteliq's guidance endpoints generate age-appropriate action plans for four audiences (child, parent, educator, and platform moderator), calibrated to severity level (low, medium, high, critical). The Quick Response Guide in the Tuteliq dashboard surfaces region-specific helplines across 13 harm categories, drawn from a verified national database updated in real time.
===
get_action_plan · audience: child · severity: high · country-specific helplines
===
/docs/api-reference/action-plan | Action plan docsThe rights dimension, teenagers are not just users to be protected
Child safety is sometimes framed as if the goal is to remove children from the digital environment entirely. That framing misunderstands the relationship between safety and rights. Teenagers have a right to access information. They have a right to creative expression. They have a right to education, to participation in cultural life, to forming their own views. An AI system that is safe for teenagers is not one that restricts them to a featureless shell, it is one that enables their genuine engagement while protecting them from the specific harms that disproportionately affect their age group.
The 2025 DSA Guidelines on the protection of minors are explicit about this. They require age assurance that is "accurate, reliable, robust, non-intrusive and non-discriminatory." Non-intrusive matters because teenagers should not have to sacrifice their right to use technology in order to be safe using it. The design obligation is to build systems that are both, safe and functional, protective and open. That is harder than a blanket ban. It is the only approach that actually serves young people.
"Some AI companies have responded by simply limiting access to adults rather than investing in product and policy changes focused on respect for the rights of younger users. How to properly respect the rights of children as their capacities evolve is an issue that companies should be grappling with."
, Meta Oversight Board, 5 May 2026
Further reading
The Meta Oversight Board is currently conducting research on exactly this question, seeking public comment on a human rights-based approach to governing AI chatbots for 13 to 17 year olds, with a white paper to follow.
If you are a developer, platform operator, researcher, or safeguarding professional, the consultation is worth reading and responding to. Submit a public comment here.
On language models specifically: the Oversight Board's consultation specifically asks about the performance of LLMs in low-resource languages, whether high-risk interactions are correctly identified when the conversation is in Bengali, Swahili, or Tagalog rather than English. This is a genuinely important gap. Most safety systems are trained predominantly on English-language data. A teenager expressing distress in a low-resource language is, at present, less protected than one doing so in English. That is a design failure with equity implications that the industry has been slow to address.
What Tuteliq provides, for every platform where teenagers talk to AI
Verified age. Calibrated risk. Emotional awareness. Contextual guidance. Every capability described in this post is available today, via the Tuteliq API, across text, voice, image, and video. Typical response 1.1 to 1.5s. Zero content retention. 32 languages (English generally available, 31 more in beta). SOC 2 Type II attestation in progress.
teal | Age verification, liveness + document | Verified age, not declared age. Blocks the bypass. Jurisdiction-specific thresholds. DPIA audit trail per verification. DSA-compliant architecture. | /v1/age-verify · Complete or Liveness Only | /docs/verification | Verification docs
blue | Age-calibrated risk scoring | Risk thresholds automatically elevated for minor age groups. The same content that triggers "monitor" for an adult triggers "immediate_intervention" for a 14-year-old. | /v1/safety/analyse · age_group parameter | /docs/api-reference | Safety analyse docs
amber | Emotional state monitoring over a caller-supplied conversation window | Tracks dominant emotions and trends across the turns you submit. Detects deteriorating emotional patterns before they reach crisis point. Fear, withdrawal, sadness, anxiety. | analyze_emotions · trend detection · conversation history | /docs/api-reference/emotions | Emotions docs
red | Grooming and manipulation detection | Multi-turn conversation analysis. Adult-minor escalation patterns. Coercive control signals. AI-enhanced sextortion language. Operates across text and voice. | detect_grooming · detect_coercive_control · AI_ENHANCED_SEXTORTION | /docs/api-reference/grooming | Grooming docsBuild the AI experience teenagers actually deserve
Verified age. Calibrated responses. Emotional awareness. Contextual support. All available today, in under 5 minutes, free tier included. Start free · Read the docs · Talk to the team.
---
Sources: Internet Matters UK survey (2024), 13% of UK teenagers experienced nude deepfakes · IWF Report Remove 2025, 98% sextortion victims boys 14 to 17 · IWF Harm Without Limits 2026, 65% AI CSAM Category A · European Commission IP/26/920 (29 April 2026), Meta DSA preliminary finding · Meta Oversight Board (5 May 2026), Governing Chatbots for 13 to 17 Year Olds · DSA 2025 Guidelines on the protection of minors.