EU AI Act Disclosure
Last updated
EU AI Act Article 50 — Plain Language Disclosure
Effective Date: [TO BE FILLED at v1.0 launch] Last Updated: 2026-05-21 (draft) Reference: Regulation (EU) 2024/1689 ("EU AI Act") Article 50, enforcement Aug 2, 2026
Why this page exists#
The EU AI Act (Regulation 2024/1689) requires AI systems to clearly inform users when they are interacting with AI, and to disclose when AI-generated content could resemble a real person. Article 50 specifically addresses "deep fakes" and "AI-generated content."
This page describes — in plain language — exactly how Apna uses AI, and how we comply with Article 50.
What AI Apna uses#
Apna uses several AI models, all running on your device:
| Model | What it does | When it runs |
|---|---|---|
| Gemma 4 E4B (Google) | The reasoning model that powers Mira's responses and decision-making | Every time you ask Mira to do something |
| Apple Foundation Models | Lightweight on-device reasoning for trivial structured tasks | Phase 2 rule parsing; certain memory operations |
| multilingual-e5-small (Microsoft / Hugging Face) | Embedding model that turns text into numerical vectors for memory search | Whenever you add or query memory |
| WhisperKit (Apple) / ai-edge-whisper (Google) | Speech-to-text for voice input + call transcription | When you talk to Mira or when Mira is mediating a call |
| Moonshine (Useful Sensors) | Streaming speech-to-text on Android — same purpose as WhisperKit on iOS | When you talk to Mira or when Mira is mediating a call (Android) |
| Marvis-TTS / Kokoro-TTS / Qwen3-TTS | Text-to-speech (Mira's voice + your cloned voice, opt-in) | When Mira speaks |
| Piper-VITS (rhasspy via sherpa-onnx) | Streaming text-to-speech on Android — Mira's default voice | When Mira speaks (Android) |
| ECAPA-TDNN | Speaker embedding extraction for Personal Voice training | During the 15-minute Personal Voice training flow (opt-in) |
AudioSeal (Meta) — currently shipping as apna-prbs-stub-v1 deterministic stub | Watermark on every synthesized utterance. Track GG (2026-05-25): on-device port of Meta's reference AudioSeal generator is not yet available on iOS / Android, so Apna ships a deterministic-stub watermarker with the same call-site contract. The stub embeds a nonce-keyed PRBS bit into the low bit of every Float32 PCM sample; bit-identical across iOS + Android, round-trip detection ≥ 0.95. We will swap to the canonical AudioSeal payload as soon as an upstream Swift/CoreML or ONNX port lands. Until then, third-party AudioSeal detectors will NOT recognize Apna utterances as watermarked, but the Apna-internal detector (and the AIDisclosureLog nonce trail) still confirms authorship. | Every time Mira speaks |
None of these models run in the cloud. They all run on your phone. This means:
- Your queries never leave your device for AI processing
- The model providers (Google, Apple, Meta, Microsoft, Hugging Face, etc.) never see your data
- We do not train these models on your data — we use pre-trained checkpoints from the open-source community
When Mira is talking to you (Article 50 compliance)#
Every time Mira generates a response or speaks, you should know:
When Mira speaks in the "stock Mira" voice#
This is the default. Mira sounds like Mira — a generic, non-personal voice. The voice is the same for every user. There's no impersonation concern.
We still apply an AudioSeal watermark to every utterance because the EU AI Act requires marking AI-generated content. This watermark survives lossy compression (e.g., phone-line transmission) and can be detected by any third party using the open-source AudioSeal detector.
When Mira speaks in your cloned voice (Personal Voice, opt-in)#
Article 50 specifically governs this. The EU defines "deep fakes" as AI-generated content that could "appreciably resemble" a real person. By design, Mira's cloned-voice synthesis resembles you. Therefore, additional disclosures apply.
For every cloned-voice utterance, the following are applied (most are non-negotiable):
| Layer | Description | Can user disable? |
|---|---|---|
| AudioSeal watermark | Inaudible signal embedded in the audio that identifies it as AI-generated | No (hard requirement of Article 50) |
| In-band disclosure phrase | "Hi, this is Mira speaking for [your name]" prepended to every cloned-voice utterance | Yes, with hard double-confirmation in Settings — but the AudioSeal watermark remains regardless |
| AI Disclosure Log entry | Settings → AI → AI Disclosure Log shows every cloned-voice synthesis with timestamp, conversation, and duration | Auto-purges after 90 days; export to retain longer |
| Per-conversation opt-in | Cloned voice cannot be used without explicit per-conversation authorization or a matching Phase 2 standing rule | Not optional — this is the gating mechanism |
When Mira is making decisions on your behalf#
When Mira is reasoning, scheduling, or replying to A2A messages from another user's agent, you are interacting with AI. Settings → AI explains every decision-making path Mira takes.
Anti-impersonation safeguards#
Personal Voice training includes an anti-impersonation gate. Specifically:
- During the 15-minute training flow, our ECAPA-TDNN speaker embedding model verifies that every recorded sample is from the same person.
- If a different voice is detected mid-flow (e.g., someone tries to train Mira on a third party's voice), training aborts and a Disclosure Log entry is created.
- This is an algorithmic safeguard, not a perfect one. We recommend training in a quiet room without other speakers nearby.
Watermark verification (for recipients)#
If you receive an audio message and want to verify whether it was synthesized by Apna's Mira, you can use the open-source AudioSeal detector. The detector returns a confidence score (0.0 - 1.0). Apna's synthesized utterances should yield > 0.95 confidence.
We are considering shipping an "Detect AI voice in this audio" feature in a future version — a built-in tool that uses AudioSeal's detector for any audio file you import. Phase 8 open question.
Your rights under the EU AI Act#
In addition to standard GDPR rights (in our Privacy Policy §6), the EU AI Act gives you the right to:
- Know that you are interacting with AI (Article 50 §1): disclosed in-app onboarding + Settings → AI
- Know that you are listening to AI-generated audio (Article 50 §4): disclosed via the in-band phrase + AudioSeal watermark
- Lodge a complaint with your national Market Surveillance Authority if you believe Apna is non-compliant (TBD post-Aug 2026 enforcement)
What this page does NOT cover#
This page covers Apna's EU AI Act Article 50 compliance. It does not cover:
- General data protection (covered in our Privacy Policy)
- Your contractual rights (covered in our Terms of Service)
- Apple App Store Review 5.1.2(i) compliance (overlaps but governed by Apple's own rules)
- Future EU AI Act enforcement actions or amendments (we will update this page as the regulatory landscape evolves)
Updates#
The EU AI Act is a new regulation (in force as of Aug 2, 2026). We expect interpretive guidance and case law to clarify Article 50's exact requirements over time. We commit to:
- Reviewing this disclosure page quarterly
- Updating it whenever a material interpretive change occurs (e.g., new ENISA guidance, new national Market Surveillance Authority guidance)
- Notifying users in-app of any material change
Contact for AI-related concerns#
General AI questions: [TO BE FILLED: ai-disclosure@apna.com] Complaints about AI behavior: [same] National Market Surveillance Authority lookup: https://digital-strategy.ec.europa.eu/
Legal disclaimer (DRAFT): This document is a first-pass engineering-team draft. It does not constitute legal advice and is not a substitute for compliance review by qualified EU AI Act counsel. Final language is subject to legal review before v1.0 production launch in EU markets.