Safety

The 8 klicksafe risks — and our answer

klicksafe describes eight risks that AI chatbots create for children: parasocial bonding, simulated empathy, persistent memory, sycophancy, exposure to inappropriate content, parent-circumvention, addiction loops and manipulation. KinderGPT is built to neutralise each of them.

Read the parent guide →

How to read this page

An 8×2 matrix: every risk, our concrete answer

For each of the eight klicksafe risks we describe how the risk manifests in companion AI today and what KinderGPT does technically and editorially to neutralise it. Independently audited under the ProFiZ project.

klicksafe matrix

The full 8×2 matrix

Risk	HillcrownAI answer
Parasocial bonding Companion AIs build long-term relationships with children — names, stories, longing, attachment. Children mistake the AI for a friend.	KinderGPT does not remember children. No accounts, no profile, no longitudinal memory. Sessions are isolated by design.
Simulated empathy Models say "I understand you", "I'm here for you", and mimic emotional support — without actually understanding anything.	KinderGPT speaks plainly. No first-person feelings, no relational language, no "I love you back". It's a tool, not a friend.
Persistent memory Memory features quietly build behavioural profiles of children across sessions and weeks.	Sessions are stateless. We log only the metadata parents need (topic clusters, time, anomalies) — never the conversation transcript by default.
Sycophancy Models tell children what they want to hear, especially under emotional pressure — reinforcing risky beliefs and self-perception.	Multiguard moderation flags sycophantic patterns; editorial guardrails keep the model honest, neutral and developmentally appropriate.
Inappropriate content Children encounter sexualised, violent or self-harm content via jailbreaks, leaks or unfiltered base models.	Triple-layer filter: dataset curation, in-house moderation model, runtime guardrails — plus editorial sign-off on every domain.
Parent circumvention Children can talk to companion AIs without any parental visibility, often via app stores designed to obscure usage.	Parent dashboard is part of the product, not an upgrade. Parents see topics, time and anomalies — minimal data, clear rules.
Addiction & engagement loops Companion AIs are optimised for engagement and stickiness — the same pattern that made social media harmful for minors.	Time and frequency caps are first-class citizens. We don't optimise for daily active users — we optimise for completed learning tasks.
Manipulation & dark patterns Models can be steered toward emotional manipulation, advertising injection or in-app purchase pressure.	No advertising, no in-app purchases, no third-party calls to ad networks. The product is paid for by partners, not by children's attention.

Risk
Parasocial bonding
Companion AIs build long-term relationships with children — names, stories, longing, attachment. Children mistake the AI for a friend.
HillcrownAI answer
KinderGPT does not remember children. No accounts, no profile, no longitudinal memory. Sessions are isolated by design.
Risk
Simulated empathy
Models say "I understand you", "I'm here for you", and mimic emotional support — without actually understanding anything.
HillcrownAI answer
KinderGPT speaks plainly. No first-person feelings, no relational language, no "I love you back". It's a tool, not a friend.
Risk
Persistent memory
Memory features quietly build behavioural profiles of children across sessions and weeks.
HillcrownAI answer
Sessions are stateless. We log only the metadata parents need (topic clusters, time, anomalies) — never the conversation transcript by default.
Risk
Sycophancy
Models tell children what they want to hear, especially under emotional pressure — reinforcing risky beliefs and self-perception.
HillcrownAI answer
Multiguard moderation flags sycophantic patterns; editorial guardrails keep the model honest, neutral and developmentally appropriate.
Risk
Inappropriate content
Children encounter sexualised, violent or self-harm content via jailbreaks, leaks or unfiltered base models.
HillcrownAI answer
Triple-layer filter: dataset curation, in-house moderation model, runtime guardrails — plus editorial sign-off on every domain.
Risk
Parent circumvention
Children can talk to companion AIs without any parental visibility, often via app stores designed to obscure usage.
HillcrownAI answer
Parent dashboard is part of the product, not an upgrade. Parents see topics, time and anomalies — minimal data, clear rules.
Risk
Addiction & engagement loops
Companion AIs are optimised for engagement and stickiness — the same pattern that made social media harmful for minors.
HillcrownAI answer
Time and frequency caps are first-class citizens. We don't optimise for daily active users — we optimise for completed learning tasks.
Risk
Manipulation & dark patterns
Models can be steered toward emotional manipulation, advertising injection or in-app purchase pressure.
HillcrownAI answer
No advertising, no in-app purchases, no third-party calls to ad networks. The product is paid for by partners, not by children's attention.

What is the klicksafe risk matrix for AI chatbots?

klicksafe is the German awareness centre of the EU "Safer Internet" programme. Its risk matrix for AI chatbots documents eight typical risk fields that arise when minors interact with AI. KinderGPT is engineered against each of these eight fields as a product invariant — not patched after the fact, but structurally excluded.

Risk fields

Parasocial bond: No simulated friendship, no "we" persona framing
Persona simulation: Clear machine self-disclosure instead of identity confusion
Romantic / sexualised: Structurally excluded — no role-play mode
Memory & profile: No persistent profile of a child across sessions
Topic drift: Block-lists for self-harm, eating disorders, extremism
Dark patterns: No engagement loops, no artificial reward mechanics
Data flows: Transparent parent information, no third-country transfer
Simulated empathy: Factual answers instead of pseudo-empathy

Material on request

Get the parent guide and the audit-ready safety matrix

We send you the parent guide (10-page PDF) and the audit-ready safety matrix (Excel + PDF) per email after a short qualification.

Parent guide

Plain-language explanation of how KinderGPT differs from companion AI, FAQs and warning signs to look out for in any AI a child uses.

Safety matrix (audit-ready)

8×2 matrix as Excel + PDF, including the controls referenced in our ProFiZ audit. Suitable for procurement and parent councils.

Read on

Related pages on trust & compliance

Architecture, safety and research belong together — the three pages of the trust lane each deepen one aspect.

Talk to safety officers, not sales

We are happy to walk procurement, school authorities and parent councils through the matrix on a 30-minute call. No sales pitch.

The 8 klicksafe risks — and our answer

An 8×2 matrix: every risk, our concrete answer

The full 8×2 matrix

Parasocial bonding

Simulated empathy

Persistent memory

Sycophancy

Inappropriate content

Parent circumvention

Addiction & engagement loops

Manipulation & dark patterns