FAA Roadmap for Artificial Intelligence Safety Assurance

What it is

The FAA’s Roadmap for Artificial Intelligence Safety Assurance, Version I says its primary purpose is to provide a path to assuring the safety of AI in aircraft and related systems for inflight operations. The scope expanded to also introduce using AI for safety. An FAA advisory committee presentation dates the first official release to August 2024.

The roadmap states that the treatment of the ethical use of AI is outside its scope, and that it is a living document the FAA plans to update periodically.

Who it applies to

The primary purpose is a path to assuring the safety of AI in aircraft and related systems for inflight operations. The scope also introduces using AI for safety. It is a roadmap, not a standard, advisory circular, or means of compliance. Applicants should work with the FAA directly on project-specific criteria.

Guiding principles

The executive summary lists these principles:

1. Work within the aviation ecosystem - use existing aviation safety requirements.
2. Focus on safety assurance and safety enhancements - the safety of AI, and using AI for safety.
3. Avoid personification - treat AI as a tool, not a human, with clear assignment of responsibility.
4. Differentiate between learned and learning AI - distinct assurance methods for each.
5. Take an incremental approach - learn and adapt assurance methods from real-world experience.
6. Leverage the safety continuum - start with lower-risk applications to gain experience.
7. Leverage industry consensus standards - adopt them as appropriate to promote global harmonization.

Learned versus learning AI

The roadmap defines a learned AI implementation as static in the operating environment, typically ML trained offline and subjected to design-time safety assurance before use. A learning AI implementation is dynamic in the operating environment, so it is not possible to qualify each version and the assurance strategy has to address the learning itself. The roadmap treats systems that continue to learn in operation as ongoing research topics.

Five areas of next steps

The roadmap identifies next steps in five areas: collaboration; FAA workforce readiness; assuring the safety of AI; using AI for safety; and aviation safety research.

How it connects to standards

For learned AI the FAA describes a strategy built from experience, consensus standards, guidance and research. On experience, it encourages companies to start with lower-criticality AI systems, and notes that project-specific issue papers and special conditions can document tailored criteria. On standards, the FAA is supporting application and possible extension of existing standards including SAE ARP4754, RTCA DO-254, DO-178C/ED-12, DO-330/ED-215 and DO-331/ED-218, and the development of new AI standards in SAE G-34. See DO-178C and DO-330 and SAE G-34.

What it is not

The roadmap sets principles and priorities. It is not a standard, an advisory circular, or a means of compliance. Applicants should work with the FAA directly on project-specific criteria. Our plain-language explainer walks through the document. Official pages are also collected on the FAA Roadmap filter. For how that sits next to tool qualification, see How to map an AI model to DO-178C / DO-330 tool qualification.

Primary source

The roadmap is published by the FAA at faa.gov; the FAA’s AI/ML technical discipline page links to it.