AI Act self-assessment

VersionDateAuthorStatus
0.12026-09-02Daniel RothmannDraft

Introduction

Eigin is a local-first personal AI app, described in the compliance overview.

This document is our self-assessment against Regulation (EU) 2024/1689, the AI Act, as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744). It sets out how Eigin is scoped under the Act, which obligations apply, and how each is met. It is not legal advice.

We rely on the Commission's AI Act FAQ, the guidelines on the AI system definition, the Article 50 FAQ, and the GPAI guidelines.

System context

See the compliance overview.

Scope

Is Eigin an AI system?

Yes. The App takes natural-language input, infers text and tool calls from a machine-learned model, and acts on the user's environment. That meets the Article 3(1) definition.

What is Eigin's role?

RoleAppliesWhy
Provider of an AI system (Art. 3(3))YesEigin ApS places the App on the EU market under its own name
Provider of a GPAI model (Art. 3(3), 3(63))NoEigin does not train or fine-tune a model today. Gemma 4 E2B is mirrored unmodified, and the Relay forwards to models hosted by others. Per the GPAI guidelines, serving or distributing another's model does not make you its provider
Deployer (Art. 3(4))InternallyEigin ApS staff use the App in their work, which makes Eigin a deployer of its own system. For a non-high-risk system the deployer duties are AI literacy (Art. 4) and Article 50(4) labelling, both covered below. Users deploying the App for personal use are out of scope (Art. 2(10))
Open-source exemption (Art. 2(12))PartlyThe App is AGPL-3.0, so the Act does not apply to it, except that Article 5 and Article 50 still do. Both are assessed below

Risk classification

TierVerdictWhy
Prohibited (Art. 5)NoSee Article 5
High-risk (Art. 6, Annex III)NoEigin is a general-purpose personal assistant. It has no intended purpose in any Annex III area: no biometrics, no access decisions for education, employment, credit, or essential services, no law enforcement or migration use, no safety components. The Aula integration reads a parent's school messages; it does not evaluate students or determine access to education
Transparency (Art. 50)YesThe agent interacts directly with a person, and it generates text. Article 50(1) and 50(2) apply
MinimalOtherwiseNo further obligations

Article 50 applies from 2 August 2026, and Eigin complies with all of it from that date without relying on the Omnibus grace period for marking.

Obligations

Article 5: prohibited practices

PracticeVerdictEvidence
5(1)(a), (b) Manipulation, exploiting vulnerabilitiesNot presentThe agent acts on explicit user requests. Proactive behaviour is a user setting with a "reactive" off level (initiative). No subliminal, dark-pattern, or engagement-optimising techniques exist; there is no tracking or ad incentive (philosophy)
5(1)(c) Social scoringNot presentNo scoring of persons. The knowledge graph is private to the device and serves the user only
5(1)(d) to (h) Predictive policing, facial scraping, emotion recognition, biometric categorisation, remote biometric IDNot presentNo biometric processing. Speech recognition transcribes words on-device and identifies nobody (voice). No emotion or sentiment inference exists in the codebase
5(1)(i) Non-consensual intimate imagery, CSAMNot presentThe App generates no images, audio clones, or video

Article 50(1): interaction disclosure

The user must be informed they are interacting with an AI system, unless obvious.

AspectVerdictEvidence
Disclosure at first interactionMetThe App is marketed and described as a personal AI agent. The first conversation is an onboarding chat whose prompt frames the agent as "a new personal AI agent being set up inside Eigin" (chat.jinja)
Never claims to be humanMetEvery system prompt opens with "an AI agent in the Eigin app" (preamble.jinja). The onboarding prompt instructs: never pretend to be human or to have a body
VoiceMetLive call uses the native call UI with the agent's name and symbol, initiated by the user in the App (live call). The same prompts apply
Third partiesMetContent the agent produces for others leaves the device only through tools the user approves per call, from the user's own accounts, after reviewing what is sent (permissions). It carries the machine-readable mark below

Article 50(2): marking synthetic content

Outputs must be marked in a machine-readable format and detectable as artificially generated, as far as technically feasible.

ModalityVerdictEvidence
TextMetAn invisible zero-width mark is appended when text leaves the App: full-message copy and agent-sent email (watermarking). A detector ships in the CLI (eigin watermark detect). The mark is not robust to paraphrase or stripping; the Act's feasibility standard weighs cost and the state of the art, and generation-time watermarking is unavailable to an app that does not run the model
Text exclusionsN/ASource code and short strings are excluded by the guidelines. Text leaves the App only through the marked paths above
AudioN/AText-to-speech renders to the device speaker in real time. No audio file is produced, stored, or shareable, so there is no artefact to mark. The listener is the user who chose the voice
Image, videoN/ANot generated

Article 50(4): deepfakes and public-interest text

These are deployer obligations. They bind Eigin ApS for its internal use, and users only in professional use.

AspectVerdictEvidence
DeepfakesN/ANo image or video generation. TTS voices are generic synthetic voices from PocketTTS and Plapre Pico, or OS voices. None imitates an existing person
Public-interest textMetEigin ApS publishes no agent-written text without human editorial review, which the guidelines exempt. A user publishing such text professionally carries the duty themselves; the watermark helps them and others detect it

Article 4: AI literacy

Providers must ensure sufficient AI literacy of their staff.

AspectVerdictEvidence
Staff literacyMetEigin's team builds the agent harness and works with model behaviour, prompt injection, and permission design daily (privacy and security). This document and the CASA assessment are part of that practice

Value chain: GPAI models

Eigin integrates models it does not train. The GPAI provider obligations (Art. 53) fall on the model providers.

ModelProviderHow Eigin uses it
Gemma 4 E2BGoogleUnmodified weights, mirrored and checksum-verified (local model)
Hosted models via the RelayMistral, TensorXForwarded unread (relay)
User's own providerChosen by the userAny OpenAI-compatible API, under the user's key and terms

Eigin relies on the providers' published model documentation to understand capabilities and limits. Model routing and the catalog record per-model capabilities (providers). Because Eigin does no fine-tuning, the downstream-modification threshold in the GPAI guidelines is not approached.