AI Regulation Tracker / Guidance and consultation
EASA Proposes Extending Aviation AI Certification Guidance to Level 3 Advanced Automation
On June 3, 2026, the European Union Aviation Safety Agency released the proposed Issue 03 of its Artificial Intelligence Concept Paper, the final deliverable under the EASA AI Roadmap 2.0. It extends certification guidance to Level 3 advanced automation, adds reinforcement learning and symbolic AI, and anticipates a design assurance ceiling on how far AI can be certified into safety-critical avionics. This is a consultation draft, not a rule, and comments are open until August 12, 2026.
On June 3, 2026, EASA released the proposed Issue 03 of its Artificial Intelligence Concept Paper for public comment. This is the document EASA has been building toward. Issue 01 set out first guidance and a trustworthy-AI framework. Issue 02 gave detailed guidance for Level 1 and Level 2 machine-learning applications, the assistance and cooperation cases where a human stays in the loop. Issue 03 is described as the final deliverable foreseen in the EASA AI Roadmap 2.0, and it completes the technical scope by reaching into the harder territory.
What Level 3 actually means
The headline extension is Level 3. In EASA's own words, the paper explores "Level 3 AI applications, corresponding to 'advanced automation'." These are applications that, again quoting EASA, "open the way to novel types of operations in which the human end user may be either remotely present, or not present during the operation."
That last clause is the whole point. Level 1 and Level 2 assume a person is there, assisted by or cooperating with the system. Level 3 is where the system carries operations the human is not continuously supervising in the cockpit sense. Think advanced automation in the flight deck, remotely supervised operations, and the building blocks that sit under more autonomous concepts. Extending certification guidance to that tier is the difference between guidance for AI that helps a pilot and guidance for AI that does things a pilot used to do.
New techniques in scope: reinforcement learning and symbolic AI
Issue 03 also widens the toolbox it will certify. EASA says it "further broadens the framework of technical guidance by addressing additional AI techniques, including reinforcement learning and symbolic AI." Earlier issues centered on supervised machine learning. Reinforcement learning, where a system learns from interaction rather than a fixed labeled dataset, and symbolic AI, the older logic-and-rules tradition, are both now inside the frame, as are hybrid architectures that combine learned and symbolic components.
The practical machinery behind this is EASA's learning assurance process, an adaptation of the classical V-shape development process that the agency has extended into what it describes as a W-shape to reinforce data management and operational design domain validation. Issue 03 extends that thinking to reinforcement-learning and logic-based constituents. For engineers, that is the part to read closely, because it is where the certification evidence you will have to produce actually lives.
The design assurance ceiling
The part that will shape strategy most is the limit. EASA is candid that the guidance does not open the door to putting AI everywhere. With the current state of knowledge of AI and machine-learning technology, the agency anticipates a limitation on the validity of applications when the AI or machine-learning constituents would carry the highest design assurance levels, the ones reserved for the most safety-critical functions where a failure has the most severe consequences.
Read plainly, that is a ceiling. AI and machine-learning constituents can be certified for a broad range of functions, but not, for now, as the highest-integrity items in the most safety-critical avionics. That is not a permanent bar. It is EASA describing what today's evidence and today's techniques can support, and it is exactly the kind of position that a consultation is meant to test. If you are planning a product architecture around AI doing the most critical job on the aircraft, this draft tells you the regulator is not there yet.
What this is, and what it is not
I want to be precise, because a regulator's name on a certification document invites over-reading. This is a concept paper. EASA released it for comment. It is not a regulation, it is not a Certification Specification, and it is not an Acceptable Means of Compliance that you can point to as the approved path today. It is EASA setting out proposed guidance and inviting the industry to push back before anything hardens.
What it does is tell you where European certification is heading and give you the structure early. When the safety regulator for EU aviation publishes its proposed final tier of AI guidance, that is a strong signal about the objectives, the assurance process, and the limits that future binding material is likely to reflect. It is a map of the road, not the road itself. Treat it as planning intelligence, not as a compliance obligation that took effect on June 3.
Why US aerospace and avionics teams should care
The FAA is your certifying authority, and this EASA paper does not bind a US program. But two things make it matter on this side of the Atlantic. First, if you build aircraft, engines, or avionics that fly in EU airspace, EASA validation is part of your reality, and this is the framework your AI-enabled systems will be measured against there. Building to it early is cheaper than retrofitting evidence later. Second, FAA and EASA have a long history of harmonizing certification approaches through bilateral arrangements, and EASA has been the more public mover on AI certification. The concepts here, the learning assurance process, the automation levels, and the design assurance ceiling, are the vocabulary the harmonized approach is likely to borrow from. For certification engineers and aviation counsel, reading Issue 03 now is how you anticipate where FAA guidance may land.
What to do now
Read the actual paper, not the summary, if AI touches your certification basis. Map your AI-enabled functions against EASA's automation levels and against the anticipated design assurance ceiling, because a function you assumed could be fully AI-driven may sit above what the current guidance supports. If reinforcement learning or symbolic or hybrid components are in your architecture, study the learning-assurance sections that now cover them. And if the ceiling or any objective would hurt or help your program, this is the window to say so: comments go to EASA by August 12, 2026. After that the draft moves on without your input.
Questions professionals are asking
Is EASA Issue 03 a binding rule that certifies AI for aviation?
No. Issue 03 is a concept paper released for public consultation on June 3, 2026. It is not a regulation, a Certification Specification, or an Acceptable Means of Compliance. It sets out proposed guidance and objectives, and comments are open until August 12, 2026.
What is Level 3 advanced automation?
EASA describes Level 3 AI as advanced automation, applications that open the way to novel operations in which the human end user may be either remotely present or not present during the operation. It goes beyond the Level 1 and Level 2 assistance and cooperation cases covered by Issue 02, where a human stays in the loop.
What is the design assurance ceiling?
EASA says that with the current state of AI and machine-learning knowledge, it anticipates a limitation on applications where AI or machine-learning constituents would carry the highest design assurance levels. In practice that caps how deep AI can be certified into the most safety-critical avionics for now. It is a current limit tied to the state of the evidence, not a permanent ban.
Does this affect US aerospace and avionics suppliers?
Not as US law. The FAA is the US certifying authority and this EASA paper does not bind a US program. It matters when you certify AI-enabled products for EU airspace, and because FAA and EASA tend to harmonize certification approaches, the concepts here are a strong preview of the shared vocabulary. Certification engineers and aviation counsel should read it as planning intelligence.
How do I comment, and by when?
EASA invites stakeholders to use the dedicated comment-response document and send feedback to ai [at] easa.europa.eu no later than August 12, 2026. If any objective or the design assurance ceiling affects your program, that window is the time to weigh in.
RELATED BRIEFINGS
Browse the full AI Regulation News tracker
Informational analysis for working professionals, not legal or certification advice. Confirm how any guidance or requirement applies to your program with qualified professionals in the relevant jurisdiction.