EASA's AI concept paper Issue 03: what it means for LLMs in aerospace

What the document is

The concept paper — Guidance for safety-related artificial intelligence applications, proposed Issue 03 — is the next step of EASA's AI Roadmap 2.0 after the March 2024 Issue 02, which covered Level 1 and Level 2 machine-learning applications. Issue 03 sets out objectives from Level 0 ("low automation") through Level 3 ("advanced automation") and extends them to the whole taxonomy of AI techniques: supervised, unsupervised and reinforcement learning, logic- and knowledge-based systems, and hybrids of these.

It is not yet regulation. EASA describes it as a set of actionable objectives, not definitive guidance, that informs rulemaking task RMT.0742 — the task whose first Notice of Proposed Amendment (NPA 2025-07) set out generic AI-trustworthiness specifications aligned with the EU AI Act. Until those deliverables apply, applicants can use the concept paper as a reference for an approval basis, and EASA may point to it through a Certification Review Item.

Where LLMs land

The paper treats generative-AI large language models under hybrid AI: in its description, LLMs are pre-trained and fine-tuned with supervised learning, then aligned with reinforcement learning from human feedback, with rule-based elements possibly layered on top. It names the challenge it considers unique to them — emergent behaviour and unexpected outputs, including what is commonly called hallucination.

The practical consequence sits in the table that limits assurance levels by AI technology. For "large OTS models" — general-purpose LLMs, whose billions or trillions of parameters and training data make control of the design process intractable — the proposed ceiling is the lowest level in every column:

Source: EASA concept paper, proposed Issue 03, Table 2 (extract). The paper also states that an AI constituent involved in catastrophic failure conditions cannot be allocated these lowest levels.

What this means for engineering teams

Most aerospace teams are not putting an LLM on board an aircraft. They are using one on the ground — to draft requirements, write code, size a structure or summarise a standard. Read in that light, the proposal draws a clear line:

• An LLM's output cannot carry safety-significant trust on its own. A TQL-5 ceiling is the level of a verification tool whose failure could only fail to detect an error — not a development tool whose output goes unchecked into the product.
• Verification has to come from somewhere else. Every artefact an LLM produces still needs the review, analysis and testing the existing standards require — DO-178C for software, the requirements and safety processes of ARP4754B and ARP4761A for systems.
• Architecture beats model choice. A newer, larger model doesn't change which row of the table it sits in. What changes the risk is whether the model's contribution is independently checkable.

How AETHER is built for this

AETHER treats the language model as the least-trusted component in the loop. Numbers are executed as code, not generated; multi-step calculations need two independent derivations to agree; THEMIS and TELOS ground regulatory claims in live eCFR and NASA sources and flag anything they can't support; and the engineer keeps design authority and sign-off. That is the posture the proposed Issue 03 describes: use the model for what it is good at, and verify everything it touches.

Our reading, not EASA's position — the concept paper is a proposal and will change through consultation and rulemaking. Read the source document at easa.europa.eu/ai.

Related notes

• The White House superintelligence accord: what human control means in engineering
• AI-generated flight code under DO-178C
• Verifiable AI: no hallucinated math
• How to check an AI's engineering calculation

Published · 3 August 2026 · All technical notes