AI System Technical Summary
EU AI Act (Regulation 2024/1689) — Article 13 Transparency
HIGH-RISK AI Classification
Cognaium's AI-powered assessment scoring, interview conductor, and candidate ranking systems are classified as HIGH-RISK under the EU AI Act (Regulation 2024/1689, Annex III, Section 4 — Employment, workers management, and access to self-employment). We are implementing compliance measures ahead of the August 2, 2026 applicability deadline.
Purpose of AI Systems
| AI System | Purpose | Inputs | Outputs |
|---|---|---|---|
| Assessment Scoring | Evaluate candidate responses to skill-based questions | Text responses, reference answers, scoring rubrics | Numerical scores (0-100), per-question feedback |
| Interview Conductor | Conduct adaptive text-based interviews on job-relevant skills | Job description, required skills, candidate text responses | Follow-up questions, skill evaluations, summary report |
| CCI Ranking | Rank candidates by composite competency indicators | Assessment scores, interview evaluations | Ranked candidate list with confidence indicators |
Training Data Description
Our AI systems use third-party foundation models (OpenAI, Anthropic, Google). Cognaium does not train models on candidate data.
Evaluation uses structured rubrics, golden test datasets, and scoring calibration sets. No candidate personally identifiable information (PII) is used to train any model.
Prompt engineering and evaluation are performed using synthetic and anonymized datasets only.
Accuracy and Performance Metrics
| Metric | Target | Description |
|---|---|---|
| Scoring Accuracy vs. Reference | ≥ 85% | Agreement with expert human scoring on calibration sets |
| Inter-rater Reliability | ≥ 0.80 (Cohen's Kappa) | Consistency between AI scores and human reviewer scores |
| Scoring Consistency | ≤ 5% variance | Score variation on identical input across multiple runs |
| Hallucination Rate | < 2% | Rate of factually unsupported claims in AI-generated feedback |
Detailed metrics are available upon request for deployers. Contact [email protected].
Known Limitations and Failure Modes
- Language Bias Potential: Non-native English speakers may receive lower scores on communication-heavy questions. Mitigated by LanguageBiasGuard module that adjusts for ESL patterns.
- Hallucination Risk: Generative scoring may occasionally produce feedback not directly supported by the candidate's response. Mitigated by HallucinationDetector and structured rubric constraints.
- Third-Party API Dependency: System availability depends on external LLM provider uptime. Mitigated by multi-provider failover configuration.
- Scoring Inconsistency Across Model Versions: Provider model updates may introduce scoring drift. Mitigated by version pinning, regression testing on calibration sets, and model evaluation pipeline.
- Context Window Limitations: Very long candidate responses may be truncated, potentially missing relevant content.
Human Oversight Mechanism
All HIGH-risk AI outputs are subject to mandatory human oversight:
- All AI-generated scores are labeled “AI Advisory — Human Review Required”
- Adverse hiring decisions (rejections) are blocked until human review is confirmed
- Every rejection requires a
reviewed_byandreviewed_attimestamp before processing - HR managers can override any AI score with documented justification
- Candidates can appeal any AI-driven outcome through our appeals process
Fundamental Rights Impact Assessment (FRIA)
Our assessment identified potential impacts on the following fundamental rights (EU Charter):
- Right to Non-Discrimination (Art. 21): AI scoring could disadvantage certain groups. Mitigated by bias audits, fairness monitoring, and the four-fifths rule.
- Right to Fair Working Conditions (Art. 31): AI outcomes influence employment access. Mitigated by mandatory human review before adverse decisions.
- Right to Data Protection (Art. 8): Personal data processed for scoring. Mitigated by data minimization, purpose limitation, retention policies, and DSAR processes.
Full FRIA documentation is available on request. Contact [email protected].
Contestation Mechanism
Candidates can contest any AI-driven outcome through the following process:
- Submit an appeal through the candidate portal (available after receiving results)
- A human reviewer (separate from the original hiring manager) will re-evaluate the assessment
- The candidate will be notified of the appeal outcome within 30 days
- If the appeal is upheld, the AI score will be overridden with the human-reviewed score
For questions about the appeals process, contact the hiring organization's HR team or email [email protected].
EU Representative
EU Authorised Representative: [To be appointed]
For EU-related inquiries regarding our AI systems, contact [email protected].
Compliance Timeline
Cognaium is implementing EU AI Act compliance ahead of the August 2, 2026 applicability deadline for high-risk AI systems. Our compliance programme includes: risk management system implementation, technical documentation, transparency obligations, human oversight mechanisms, accuracy and robustness testing, and conformity assessment preparation.