EU AI Act classification

v1.0 · effective 2026-04-27

Mill is an AI-powered course-authoring tool that uses large language models for instructional design + content writing, and image / TTS models for media. This page documents how Mill classifies under Regulation (EU) 2024/1689 (the AI Act).

Summary

Mill is a limited-risk AI system under the AI Act (Article 50). It is not a high-risk system because it does not autonomously make decisions that affect access to education, employment, or essential services for learners. Mill assists authors; the author retains full editorial control over what ships.

Classification reasoning

• Annex III §3 (Education) covers AI systems that determine learner access, score exams, or evaluate learning outcomes for a binding decision. Mill does none of these — it produces course content that authors review and publish; pass/fail decisions are made by the customer LMS.
• Article 50 transparency obligations apply: we disclose every AI-generated section to learners (visible attribution in course metadata + xAPI extension), and we inform authors about AI involvement at every step (visible throughout the wizard + editor).
• Article 5 prohibited practices: Mill does not perform social scoring, biometric inference, or subliminal manipulation. Affect-recognition is not used.

Author obligations Mill helps fulfill

• Disclosure: every Mill-generated course carries an “AI-assisted” metadata flag (xAPI extension generation.aiAssisted = true).
• Human review: the editor's review-and-publish flow is mandatory; courses cannot be auto-published from the generation pipeline.
• Source content tracking: when a course is generated from a document upload, the source-document hash is logged on the Course row for provenance audits.
• Right to rectification: authors can edit any section after generation and republish (the SCORM zip regenerates).

High-risk content guardrails

When a course's topic or audience matches Mill's compliance keyword scan (regulatory training, safety, legal), Mill automatically:

• Disables engagement gamification (points, narrative wrappers, streaks, timed challenges).
• Forces a complianceContext flag through to the LRS export.
• Refuses certain deprecated patterns at the AI architect level (e.g. story-wrapping decoy details).

See the architecture page for technical details on how this is enforced.

Model providers

Mill uses third-party general-purpose AI models as sub-processors. Each model provider carries its own AI Act obligations and is responsible for meeting them; Mill selects providers that publish GPAI-tier documentation under Articles 53+ and records the model used for every generation in the audit trail.

• Anthropic Claude (Sonnet, Haiku) via the Anthropic EU endpoint.
• Google Gemini for image generation, through the Gemini Developer API. The request carries no region and is not EU-pinned.
• Google Cloud Text-to-Speech for narration, on its global endpoint. The request carries no region and is not EU-pinned.
• Where the EU AI region is selected — a separate choice for outline and content, for section images, and for narration: Mistral AI (Paris) for the outline and section writing, FLUX by Black Forest Labs (Freiburg) for images, pinned to that provider’s EU API host, and the operator-configured speech endpoint for narration.

Where each provider processes, per role and per setting, is in the table on the architecture page.

Classification review cadence

This classification is reviewed every six months and after any change to: (a) Mill's AI feature set, (b) the AI Act guidance from the European AI Office, or (c) the customer deployment context (e.g. integration into an LMS that performs binding learner decisions).

AI Act compliance contact: compliance@colport.eu