EASA AI Concept Paper Issue 03: What the DAL C Ceiling Means for Your Certification Strategy

EASA released Proposed Issue 03 of its Concept Paper on Artificial Intelligence on 3 June 2026 [2]. The document is open for consultation until 12 August 2026 [3]. Stakeholders are invited to submit comments using the dedicated comment-response document, addressed to ai@easa.europa.eu. As of 19 June, approximately eight weeks of that ten-week window remain. If you build safety-critical inference engines for aviation, here is what the document actually says, why the ceiling exists, and how to architect around it now.

What the Ceiling Says

Table 2 of the document "highlights the limitations that are currently observed based on the available means of compliance and the state of the art of technology" [1]. The substance of the limitation has been consistent across all three issues of the Concept Paper [4]. With the current state of knowledge of AI and ML technology, EASA anticipates a limitation on the validity of applications when AI/ML constituents include IDAL A or B, SWAL 1 or 2, or AL 1, 2 or 3 items. No assurance level reduction should be performed for items within AI/ML constituents. The limitation will be revisited when experience with AI/ML techniques has been gained [1].

In plain terms: if your AI constituent touches a failure condition classified at DAL A or B under airworthiness rules, or their ATM/ANS equivalents, EASA does not currently have means of compliance to certify it. DAL C is the ceiling. The language in Issue 03 is not aspirational, and it is not a starting point for case-by-case negotiation. It is a documented limitation tied explicitly to the current state of compliance tooling, not a permanent constitutional judgment about learned systems.

This is the final Concept Paper deliverable foreseen under the EASA Artificial Intelligence Roadmap 2.0 [2]. The document provides actionable objectives but does not constitute definitive or detailed guidance at this stage. It serves as a reference for Roadmap 2.0 Phase II, when formal regulatory development comes into force via RMT.0742 [1].

Why the Ceiling Exists

The limitation is an engineering diagnosis, not a policy preference. AI assurance differs from traditional development assurance because the system's intended function is driven by data, scenarios or knowledge, depending on the technique used, making full requirements-design-implementation traceability infeasible [1]. Classical software achieves DAL A and B compliance by letting a certification authority walk the logic tree from requirement to code line and bound every behavioural path. A trained model learned a mapping from data. You can test it, bound its behaviour over a defined operational design domain, and apply formal robustness methods, but you cannot audit what the model learned the way you audit a line of C.

A correct and as complete as possible definition of the ODD is a prerequisite to an adequate level of quality of the data sets, scenarios or knowledge bases involved in the AI assurance process. However, an exhaustive exploration of all possible operating conditions is acknowledged to be intractable for high-dimensional input spaces [1]. For a catastrophic-function application, that intractability is not an acceptable residual risk. A 99.999% correct rate on historical data is cold comfort if the failure mode clusters in exactly the meteorological or traffic geometry conditions that matter most for the safety case.

The Strongest Version of the Counterargument

Let me state the critics' case fairly, because it is serious.

Empirical reliability data for AI systems is increasingly compelling. Waymo has driven over 200 million autonomous miles on public roads as of February 2026, and according to Waymo's own safety reporting, the Waymo Driver was involved in 13 times fewer serious-injury-or-worse crashes than human drivers in the same cities [7]. In medical imaging, seminal work by Rajpurkar et al. demonstrated radiologist-level pneumonia detection, establishing transfer learning as a standard practice for CNN-based classifiers on defined tasks [8]. More broadly, convolutional neural network-based algorithms now automate many radiology tasks such as lesion detection, classification, and quantification with performance often comparable to radiologists [9].

The argument from this data is that empirical reliability, demonstrated over statistically representative conditions, should count for more in a risk-based regulatory framework. If a system demonstrably achieves a well-bounded failure-rate estimate, why should the certification methodology be the gating factor?

There is also a competitive concern. EASA is moving faster and more comprehensively than any other aviation regulator, and because the EU AI Act is already law, EASA's framework is likely to become a global reference model for AI assurance in aviation [10]. The joint SAE G-34 and EUROCAE WG-114 standards committee, working with global industry and regulators on AI certification means of compliance, is on track to publish its first recommended guidance, ARP-6983, which will detail assurance methods for building and integrating trustworthy AI into aerospace systems up to a DAL C safety standard [11]. Their work is aligned with the EASA AI Roadmap and is designed to also support FAA guidance [10]. The DAL C ceiling is not uniquely EASA's. It is converging toward a trans-Atlantic standard. If you are betting that the FAA will certify your DAL A AI constituent while EASA holds at DAL C, that bet is looking increasingly narrow.

That said, competitive asymmetry is a real near-term concern for European ventures. A product that cannot pursue DAL A or B functions in the European market today faces a constrained addressable space relative to what the technology may eventually permit.

Why the Ceiling Is Defensible, for Now

The unfashionable view is that EASA is largely right, and the autonomous vehicle data above actually illustrates why.

The Waymo figure is impressive precisely because it accumulated in a well-defined operational domain: urban speeds, extensive redundancy, remote monitoring infrastructure. Generalising that reliability argument to a collision avoidance function operating at cruise altitude, in degraded weather, with no remote operator, at the tail of a distribution that directly causes catastrophic outcomes, requires exactly the generalisation evidence that EASA says does not yet exist. A recent systematic review found that when a CNN trained on one patient cohort was tested on a new trauma cohort of older patients, specificity dropped from 94% to 70%, even with stable sensitivity [8]. Distributional shift is not a theoretical concern. It is the mechanism by which catastrophic-function failures accumulate.

The ceiling is not permanent. The consistent language across all three Concept Paper issues is that the limitation "will be revisited when experience with AI/ML techniques has been gained" [1]. That is the stated mechanism of change, and Issue 03 opens more architecture space than it closes.

Building on Issue 02, which explored Level 1 AI and Level 2 AI, Proposed Issue 03 completes the technical scope foreseen in the EASA AI Roadmap 2.0 [4]. It further broadens the framework by addressing reinforcement learning and symbolic AI, and explores Level 3 AI applications corresponding to advanced automation [1] [2]. The document explicitly covers hybrid architectures combining symbolic and learned components. The W-shape process adapts the classical V-shape development assurance process to reinforce data management and ODD validation, and Issue 03 extends this to reinforcement learning and logic-based constituents, giving architects a richer toolkit than a simple supervised-learning model [1] [17].

This also matters for the regulatory pipeline ahead. On 10 November 2025, EASA opened its first regulatory proposal on AI in aviation for public consultation: NPA 2025-07, setting out detailed specifications plus AMC/GM that operationalise the EU AI Act's high-risk system requirements for aviation [12]. The publication is the first step of Rulemaking Task RMT.0742, which will be followed by a second NPA in 2026 to deploy the generic framework to the relevant aviation domain regulations [12]. The Concept Paper is the technical foundation. The rulemaking will translate it into binding acceptable means of compliance. The ceiling can rise when those AMC include methods that close the generalisation gap.

What This Means for Founders Building Safety-Critical Inference Engines

DAL C is not a death sentence for your product. It is a scope definition. A vision-based detect-and-avoid system certified at DAL C with a well-specified ODD can generate real commercial value, particularly in the uncrewed aviation market. The important question is not "how do I get to DAL A today?" It is "what is the highest-value product I can build within the current ceiling?"

The ACAS X architecture is the template. RTCA DO-385 sets forth minimum operational performance standards for ACAS X equipment, including both active surveillance variants [5]. The paired standard, DO-385/ED-256, covers ACAS Xa and ACAS Xo [6]. The neural-network-compressed implementations of the ACAS X decision tables have been the subject of extensive academic verification work [16]. The ACAS Xu variant for unmanned aircraft uses a large lookup table in the design, with a neural network compression of the policy proposed to reduce the memory constraint of on-board avionics hardware [14]. Critically, ACAS Xu remains in standards-development testing: Reliable Robotics recently completed an FAA contract demonstrating an ACAS Xu-based detect-and-avoid system on a Cessna Caravan during simulated collision avoidance scenarios, with data collected to guide development of industry standards for DAA systems [15]. That is still standards development, not a certified product. The point for your architecture is the pattern itself: a learned or compressed component feeding a deterministic certified decision layer. The AI constituent handles perception or policy compression; the certifiable logic layer makes the final resolution advisory. The criticality allocation works because the deterministic layer is auditable at higher DAL.

Architectural decisions made today are certification decisions. Every week you accumulate training data without a formally documented ODD definition is a week of certification evidence you cannot retroactively reconstruct cheaply [1]. The first V of the W-shape process is your data management and ODD validation cycle [17]. Instrument your data pipeline against an ODD specification from day one.

The consultation window is your policy lever. Following the consolidation phase, EASA will continue to work with stakeholders in support of the AI strategy for aviation [2]. If your team has empirical evidence that specific means of compliance, such as formal verification of ODD coverage, conformal prediction intervals, or runtime monitoring architectures, are capable of supporting a higher assurance level than current guidance assumes, submit it before 12 August. The stated mechanism by which the ceiling moves is industry-supplied experience with AI/ML techniques [1]. That is not rhetoric.

Do not build a catastrophic-function product on pure supervised learning today and assume the ceiling lifts before your certification date. It might not. Build your system so the learned AI constituent sits at DAL C and the safety-critical decision layer above it is deterministic and certifiable at a higher level. That is not a compromise. It is sound engineering for the regulatory environment that actually exists.

The consultation opened 3 June 2026 and closes 12 August 2026. Read Table 2 of the document. Identify which failure conditions your AI constituent touches. Architect to keep the learned component below DAL C. Send substantive technical comments to ai@easa.europa.eu. That is the mechanism by which industry shapes what comes out of RMT.0742.