EASA AI Concept Paper Issue 2 explained for suppliers

The levels walkthrough is EASA AI Concept Paper levels explained. This page is the supplier brief: what Issue 2 actually asks you to produce, and what it is not.

What Issue 2 is

EASA published Artificial Intelligence Concept Paper Issue 2 as guidance for Level 1 & 2 machine-learning applications. The PDF does not constitute final detailed guidance. EASA’s news item says Issue 2 introduces human-AI teaming for Level 2.

It is preliminary guidance, not an acceptable means of compliance and not an EASA approval of your product. See the EASA Issue 2 explainer.

What a supplier should be ready to show

Operational design domain. EASA says a correct and complete ODD definition is a prerequisite to an adequate level of quality of the data sets. If you cannot name the ODD, you cannot yet talk about data quality.

Independent data sets. EASA bases learning assurance on three separate and independent data sets for training, validation and testing. Objective DM-01 is to identify data sources and collect data in accordance with the defined ODD while ensuring data quality requirements are satisfied. The data-assurance article walks that W-shaped path: Data assurance for ML avionics.

Learning assurance as an extension, not a replacement. EASA proposes learning assurance as an extension of traditional development assurance so trained models generalize on unseen operational data and are robust in foreseeable conditions. It also refers to the limitations of traditional development assurance for data-driven approaches. Issue 2 does not say learning assurance replaces DO-178C or DO-330.

Criticality cap. With supervised learning, EASA will initially accept only applications where AI/ML constituents do not include IDAL A or B / SWAL 1 or 2 / AL 1, 2 or 3 items, because some anticipated means of compliance for the highest criticality are not yet available.

Human-AI teaming (Level 2). The Issue 2 news item names human-AI teaming. A supplier of a Level 2 application should expect to explain the human role, not only the model metrics.

What vendors publish (not an EASA approval)

LDRA names EASA guidance for level 1 and 2 machine learning applications on its AI/ML page. That is a vendor statement that the product is aimed at those levels, not an EASA letter. MathWorks AI Verification points to a worked airborne deep-learning example that mentions prospective EASA guidelines; that is a vendor example. The Commission’s AI Act Service Desk is the official scope walkthrough for the Act, not an EASA AMC.

How this sits next to the EU AI Act

Article 108 amends Regulation (EU) 2018/1139 so that implementing and delegated acts on AI systems that are safety components take into account the requirements set out in Chapter III, Section 2 of the AI Act. Extra-territorial reach for a non-EU supplier is in EU AI Act obligations for aerospace suppliers outside the EU. Issue 2 is still not the Act’s conformity assessment.

Research, not a means of compliance

EASA’s MLEAP project aims to identify concrete means of compliance with the learning-assurance objectives, covering generalisation guarantees, data completeness and representativeness, and algorithm and model robustness. That is research. It is not a published AMC you can put in a certification plan as if it were DO-178C.

Questions

Does Issue 2 let me ship a Level A ML constituent?

Issue 2 says EASA will initially accept only applications where AI/ML constituents do not include IDAL A or B / SWAL 1 or 2 / AL 1, 2 or 3 items. A Level A claim needs a project-specific conversation with the authority, not this concept paper.

Is a completed MLEAP paper an AMC?

No source we cite says that. MLEAP aims to identify concrete means of compliance. Identification is not issuance.