Can machine learning be certified for flight-critical applications?

EASA's AI Concept Paper provides a pathway for certifying ML in flight-critical applications through learning assurance, which requires rigorous data quality management, model verification with demonstrated behavioral bounds, and operational validation. The requirements increase with criticality level, and DAL A applications face the most stringent evidence requirements.

What is EASA's W-shaped development model for AI?

The W-shaped model extends the traditional V-model used in aviation software development by adding parallel data and learning assurance activities. One arm covers system and software development as in traditional DO-178C, while the parallel arm covers data management, model training, and learning verification.

Do drone AI detect-and-avoid systems need certification?

Yes. AI-based detect-and-avoid systems for BVLOS drone operations require demonstration of reliability and performance levels as part of the SORA risk assessment process. The specific requirements depend on the operational risk class (SAIL level) and are assessed by the national aviation authority.

How does the EU AI Act interact with aviation AI certification?

Aviation AI systems certified under EASA regulations qualify as high-risk under EU AI Act Article 6(1). The AI Act defers to sector-specific conformity assessment where it exists, meaning EASA certification procedures take precedence, but the AI Act's general requirements for risk management and transparency still apply.

What standards apply to AI in air traffic control systems?

ATM AI systems must comply with EUROCAE ED-109A for software assurance, ED-78A/DO-264 for safety assessment, and Commission Implementing Regulation (EU) 2017/373 for ATM/ANS common requirements. EASA's AI Concept Paper requirements apply in addition to these sector-specific standards.

Quick answer

Aviation AI must be certified under EASA's Artificial Intelligence Concept Paper framework, which extends traditional DO-178C/DO-330 software assurance levels with AI-specific requirements for learning assurance, data quality, and explainability, varying by the criticality level (DAL A through E) of the AI function.

Updated June 2026 · MmowW AI Compliance

AI Compliance in Aviation: Air Traffic, Drone AI, and Safety Certification

Aviation AI Certification Framework

Aviation has the most mature AI certification framework of any industry, built on decades of software assurance practice. The European Union Aviation Safety Agency (EASA) published its AI Concept Paper in 2024, establishing a structured approach to certifying AI in aviation applications. The framework builds on existing Design Assurance Levels (DAL A through E) from DO-178C and extends them with AI-specific requirements.

EASA identifies three levels of AI: Level 1 (human-assisted AI providing information), Level 2 (human-AI collaboration with shared decision authority), and Level 3 (advanced AI with high autonomy). Each level triggers increasing certification requirements for learning assurance, data quality, robustness, and explainability.

Certification Requirements by AI Application

Aviation AI Application	Criticality (DAL)	EASA AI Level	Key Certification Standard
Flight control automation	DAL A (catastrophic)	Level 2-3	DO-178C + EASA AI Concept Paper Level 2/3
Air traffic management decision support	DAL B-C	Level 1-2	EUROCAE ED-109A + EASA AI Concept Paper
Predictive maintenance	DAL C-D	Level 1	DO-178C + airline safety management system
Drone detect-and-avoid	DAL B-C	Level 2-3	EUROCAE ED-318 (forthcoming) + EASA SORA
Cabin crew decision support	DAL E	Level 1	Airline operational approval
ATC conflict detection	DAL A-B	Level 1-2	EUROCAE ED-109A + national ANS regulations

Learning Assurance

EASA's framework introduces the concept of learning assurance, which replaces traditional deterministic verification for AI/ML components. Learning assurance covers the entire ML lifecycle: data management (collection, labeling, validation), model training (architecture selection, hyperparameter tuning, training process documentation), model verification (performance metrics, out-of-distribution detection, adversarial robustness), and model validation (operational environment testing, edge case coverage).

For DAL A and B applications, EASA requires independent verification of training data quality, formal analysis of model behavior bounds, and demonstrated coverage of the operational design domain. The W-shaped development model extends the traditional V-model with parallel data and learning assurance activities.

Drone AI Regulations

AI in unmanned aircraft systems (UAS) falls under EASA's regulatory framework established by Commission Implementing Regulation (EU) 2019/947 and Delegated Regulation (EU) 2019/945. AI-specific requirements arise in several areas:

Detect-and-avoid systems for Beyond Visual Line of Sight (BVLOS) operations require SORA (Specific Operations Risk Assessment) authorization with demonstrated AI reliability levels
Autonomous flight planning AI must comply with U-space Regulation (EU) 2021/664 for operations in U-space airspace
AI-based geo-awareness systems must access authoritative airspace data and enforce geographic restrictions
Swarm operations using AI coordination are addressed in emerging EASA guidance on multi-UAS operations

The JARUS (Joint Authorities for Rulemaking on Unmanned Systems) guidelines on AI and automation for UAS, published in 2023, provide additional guidance on AI-specific risk assessment for drone operations.

Air Traffic Management AI

AI in ATM falls under the Single European Sky framework and EASA oversight. The SESAR programme (Single European Sky ATM Research) has identified AI applications including trajectory prediction, conflict detection and resolution, arrival and departure management, and weather impact assessment. Each requires certification under the ATM/ANS common requirements (Commission Implementing Regulation (EU) 2017/373).

ATM AI systems are subject to safety assessment per EUROCAE ED-78A/RTCA DO-264 and software assurance per EUROCAE ED-109A. The human factors implications of AI-assisted ATC are addressed in EUROCONTROL guidance on human-machine teaming.

International Standards Alignment

The FAA in the United States has not yet published equivalent AI certification guidance, though it participates in RTCA Special Committee 240 on AI/ML. ICAO Annex 8 (Airworthiness) and Annex 10 (Aeronautical Telecommunications) are being updated to address AI. RTCA DO-178C remains the primary software assurance standard, with the supplement DO-330 for tool qualification. SAE ARP4754B for system development and ARP4761A for safety assessment are being revised to incorporate AI/ML considerations.

Compliance Pathway

Classify the AI application by criticality level (DAL) and EASA AI Level (1, 2, or 3)
Apply DO-178C/DO-330 software assurance as the baseline, supplemented by EASA AI Concept Paper requirements
Implement learning assurance processes covering data management, model training, verification, and validation
For drone AI, complete SORA risk assessment and obtain operational authorization from the national aviation authority
For ATM AI, follow the SESAR deployment pathway and obtain ANS provider approval
Maintain configuration management and post-deployment monitoring throughout the AI system lifecycle

Check your AI compliance readiness — free.

Take the Readiness Check 3 minutes · 10 questions · no signup required

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.