THE TAKEAWAY
Define the human’s intervention role as carefully as the system’s analytical capability.
My assessment: oversight is an operational capability. A confirmation button is not evidence that a person can meaningfully challenge an AI system.
- Where AI could help
- Keep research and drafting assistance separate from action-taking agents. Increasing action authority requires a new, bounded assurance argument.
- What remains uncertain
- The reviewed roadmap cannot tell us whether a particular interface, model or escalation design is adequate in a specific operation.
- The human decision
- Specify who may approve, override or stop an action—and demonstrate that workload and time allow them to do it.
- Evidence & confidence
- Human-centric framing is supported by EASA’s roadmap. Our design questions are analysis, not a certification checklist.
AI-generated interpretation · Prepared 16 September 2026 · Professionally reviewed and approved 16 September 2026. Sources are listed below; no live operational data are connected.
An approach, not a blanket approval
EASA’s 2023 AI Roadmap 2.0 sets out a human-centric programme addressing the safety and ethical dimensions of AI in aviation. A roadmap is not approval for a particular operational deployment. Source ↗
In November 2025, EASA opened NPA 2025-07 for consultation as its first aviation-AI regulatory proposal, covering Level 1 assistance and Level 2 human–AI teaming alongside technical trustworthiness guidance. This is consultation context, not a claim that the proposal is adopted law or that any particular system is approved. [aiProposal]
Our analysis: an AI assistant that sorts reports and a system that initiates an operational action need different assurance arguments. Calling both “agentic” does not make their consequences equivalent.
Make oversight possible
The reviewer needs to understand the recommendation, its evidential basis and its limits. They also need a realistic opportunity to challenge it. A request for confirmation becomes a formality if the workload, interface or time pressure makes meaningful review impossible.
For a proposed use case, identify actions the system may take, actions that require approval, and actions it must never take. Define escalation and fallback behaviour before discussing autonomy as a benefit. These are editorial design questions, not a regulatory checklist.
Test the uncomfortable cases
Use a hypothetical exercise: the assistant recommends a change based on incomplete information while the responsible person is occupied. Does the system wait, escalate or continue? Is the behaviour visible and bounded?
Also test confidently wrong output, contradictory sources, stale inputs and attempted instructions embedded in retrieved material. Retain enough evidence to reconstruct the recommendation and the human decision.
The goal is not maximum automation. It is a demonstrable improvement within a clearly defined operating boundary.
A QUESTION FOR YOUR NEXT SAFETY DISCUSSION
Could the accountable person explain, challenge and stop the proposed AI action in the time available?
Follow the evidence
Sources checked for this edition on 16 September 2026. Uncited examples and recommendations are IASMS editorial analysis, not official requirements. AI-produced editorial analysis reviewed by an experienced aviation safety professional for aviation-safety context, evidence boundaries, accountability and publication suitability. This is not regulatory, legal or independent specialist certification. How we work →