Skip to content

LEGAL

AI Transparency (Art. 50 AI Act)

KindChat's use of AI — transparency under Art. 50 of the AI Act.

Last updated: September 2026 | Art. 50 AI Act applies from 2 August 2026.

KindChat uses AI exclusively for moderation and to protect users. The AI classifies content — it does not generate new content (no text, images, videos or voices). The lighter transparency obligations of Art. 50(1) AI Act (interaction transparency) therefore apply, not the marking/watermarking obligations for generative AI (Art. 50(2)).

1. What the AI does at KindChat

  • Automated content checks for harmful/unwanted content (insults, hate, threats, scams, phishing, unwanted nude images, bullying, etc.).
  • Checks performed primarily directly on your device (on-device), before a message is encrypted and sent.
  • In rare cases of doubt (high uncertainty of the local AI or a legitimate user report), an additional, strictly purpose-bound check on EU servers.
  • E2EE remains unaffected; message contents are not analysed for advertising, profiling or training purposes.

Result for you: protection from harmful content without the provider reading your private messages.

2. Where and how to find the AI transparency information

2.1 In the app (Art. 50(1))

During onboarding, a clear notice:

KindChat uses automated procedures including artificial intelligence to detect harmful content and protect you. The check takes place primarily on your device.

At the point of effect (blur/warning/block): a short notice "Automated moderation decision" + a direct button/link to report/complain (the DSA complaint function). In the settings: a detailed explanation with a reference to this page and the Terms of Use.

2.2 On the website (kindchat.app)

  • this section (linked from the footer, Terms of Use, Privacy Policy);
  • this subpage "AI Transparency & Art. 50 AI Act";
  • a reference in the Terms of Use (clause 3) and the Privacy Policy.

3. Limits of the AI protection features

No automated system is perfect. KindChat substantially reduces the risk of harmful content but cannot fully eliminate it: harmful content may go undetected (false negative), harmless content may be flagged in error (false positive). The protection features do not replace your own caution. You retain full control: view, report, block the sender. No quantified detection rates (no-quota rule).

4. Human oversight and the right to complain (DSA + Art. 14 AI Act)

  • the right to report harmful/wrongly flagged content in the app;
  • a complaint against moderation decisions via the integrated DSA complaint function;
  • in appropriate cases, review by a natural person.

Safety-critical cases are not decided conclusively by automation alone.

5. Care Circle feature — consent-based and transparent

The optional "Care Circle" (e.g. for families) helps manage contacts and protective measures. Only with explicit consent that can be revoked at any time. No one (neither the provider nor other people in the Care Circle) gains insight into the actual message contents. All actions taken on behalf of a protected person are visible to that person. KindChat is not a surveillance tool.

6. What KindChat AI does not do (and never will)

  • No generation of content (text, images, deepfakes, voices) — no Art. 50(2) obligations.
  • No reading/storing of message contents for advertising/profiling purposes.
  • No emotion recognition or biometric categorisation (Art. 50(3) not affected).
  • No "guaranteed" security — only ever "risk reduction".

7. Language-dependent detection quality

Quality can vary by language (training data, linguistic specifics). Continuous improvement for all supported languages, but no guarantee of equivalent quality in every language.

8. Technical documentation / Model Card

A public, privacy- and security-conscious version of the technical documentation of the moderation AI (Model Card) can be provided on request or via a protected link. It contains no details that would allow moderation to be circumvented.

Model used for the additional check. The additional check on our EU servers uses Meta’s Llama Guard 3, which we run ourselves. Built with Llama. “Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.” Licence text: llama.com/llama3_1/license.

9. Contact & feedback on the AI

For questions about the use of AI, a specific moderation decision or a complaint: use the report/complaint function in the app or email hi@kindchat.app.

Legal context: This transparency information refers to Art. 50(1) of Regulation (EU) 2024/1689 (the AI Act) for AI systems with limited risk (content moderation in a consumer messenger). KindChat does not fall under the high-risk categories of Annex III of the AI Act, as long as moderation is not used in the sensitive contexts named there (education, employment, law enforcement, etc.).

sonrisa ventures GmbH | https://kindchat.app | Juni 2026

← Back to home