EU AI Act Annex III §3 disclosure · Koydo policy

EU AI Act Annex III §3 disclosure.

The EU AI Act classifies AI systems used in education as 'high-risk' under Annex III §3. Koydo's Cortex AI assist features fall under this classification when used by EU schools. The specific disclosures + risk-management measures Koydo implements.

Status: Active for EU AI deployments · Last reviewed: 2026-05-20

Why Cortex is Annex III §3 high-risk

Annex III §3 covers AI systems intended to be used for: (a) determining access to educational institutions, (b) evaluating learning outcomes, (c) assessing appropriate level of education, (d) monitoring + detecting prohibited behavior during exams. Koydo's Cortex assist features (homework grading drafts, lesson recommendations, engagement summaries) fall under (b).

Risk management system (Article 9)

Koydo maintains a documented risk management system covering: hazard identification (what could go wrong with AI-drafted homework comments), risk estimation (how often is the comment wrong + how bad is wrong), risk control measures (human-in-loop required, confidence thresholds, fallback to manual mode), post-market monitoring (accept/edit/regret rates surfaced via /admin/ai-quality).

Data + data governance (Article 10)

Training data for Cortex's underlying models (OpenAI and Google models): we don't control this; we rely on OpenAI + Google's own training-data governance. For Koydo-side fine-tuning + prompt data: pseudonymized before transmission to provider (student names → 'student', etc.); not retained beyond the inference call.

Technical documentation (Article 11)

Koydo maintains full technical documentation per Annex IV: AI system intended purpose, AI techniques used (LLM fallback chain), hardware/software descriptions, training/validation/testing datasets (provider-owned), human oversight measures, accuracy + robustness specifications, cybersecurity measures. Available to EU controllers + supervisory authorities on request.

Human oversight (Article 14)

Koydo's Cortex assist is always human-in-loop. Drafts are NEVER auto-sent — they're surfaced for tutor approval (in Suggest mode) or tutor override (in Autopilot mode, which still requires explicit opt-in per tutor). The system can be paused or rolled back via /educator/ai-assist settings.

Transparency to users (Article 13 + 52)

Users (tutors + parents + students) are informed when AI is in use. Drafts are clearly labeled as 'engine-drafted' in the UI. Parents can opt their child out of AI assist features (see /parent/notifications). Tutors can switch between Manual / Suggest / Autopilot modes per their preference.

Conformity assessment (Article 43)

High-risk AI systems require conformity assessment before deployment in the EU. Koydo's Cortex is preparing for conformity assessment via the internal-control procedure (Annex VI) — appropriate for AI systems with extensive human oversight that don't fall under stricter Annex VII third-party assessment.