The FAA AI Safety Assurance Roadmap, explained

The FAA’s roadmap is short on rules and long on principles. That is deliberate, and it shapes how an applicant should read it.

Status as of 14 October 2026: the FAA AI/ML technical discipline page still links Roadmap for Artificial Intelligence Safety Assurance, Version I as the listed resource. We found no Version II on that page.

What problem it addresses

The roadmap’s executive summary says AI must demonstrate its safety before use in aviation, and that it presents a new challenge for systems that achieve performance by learning rather than design, because of the absence of engineering principles that guide the traditional engineering design process.

Its primary purpose is a path to assuring the safety of AI in aircraft and related systems for inflight operations. The FAA also expanded the scope to introduce using AI for safety. Ethical use of AI is outside its scope.

The principles that matter most in practice

Work within the existing aviation ecosystem. The first principle is to use existing aviation safety requirements. The FAA’s presentation to its research advisory committee lists “Use as much existing regulation as possible” among the guiding principles.

Learned versus learning AI. The roadmap distinguishes static (learned) from dynamic (learning) implementations. For learned AI, each new version is subjected to safety assurance and in-service monitoring becomes part of the continuous operational safety program. For learning AI, the roadmap says safety assurance must be built into the operating environment or the learning process, and treats it as an ongoing research topic.

Incremental approach and the safety continuum. The roadmap encourages companies to start with lower-criticality AI systems so the industry and the FAA can gain experience, scaling assurance to the risk involved.

Avoid personification. AI is treated as a tool, not a human, with the designer and maintainer accountable unless responsibility is allocated elsewhere by law.

Lean on consensus standards. The FAA is supporting the development of voluntary consensus standards, including application and possible extension of DO-178C, DO-330 and related documents, and new AI standards in SAE G-34.

The five areas of work

Collaboration, FAA workforce readiness, assuring the safety of AI, using AI for safety, and aviation safety research are the five areas for next steps.

What it does not give you

The document is not an advisory circular and does not define a compliance checklist. The route it describes for learned AI goes through experience on specific projects: project-specific issue papers provide a flexible means to develop tailored criteria, and final special conditions document the resultant criteria. The FAA also encourages companies to work with it early in product development.

Practical reading for applicants

Analysis, not sourced fact: the roadmap tells you the agency will meet you project by project, will expect existing development assurance practice to be applied and extended, and will wait for SAE G-34’s ML standard to mature. Plan for early engagement and for evidence that maps to DO-178C-style objectives plus the data-centric activities EASA describes. See DO-178C and ML: where the gaps are, How to map an AI model to DO-178C / DO-330 tool qualification, and the FAA roadmap reference page. Official documents are also listed on the FAA Roadmap filter.

The FAA’s AI/ML technical discipline page still links to Version I and describes the agency’s research focus on measuring functionality and performance of AI systems. For whether that amounts to a certification pathway today, see Does the FAA certify AI today?.