Static vs. Dynamic: How AI Architecture Determines Certification
Principle 4: Differentiate Between Learned AI and Learning AI
How do you certify a system that’s changing?
Traditional certification is straightforward. You test a system exhaustively. You prove it safe. You certify it. The system doesn’t change. You monitor it in service, but the core behavior is fixed.
But what if the AI system keeps learning after deployment? What if it evolves based on new data it encounters in the real world? How do you certify something that’s not done changing?
The FAA’s answer to that is that you don’t. Not yet at least. Maybe not for years.
This post is part of the FAA Roadmap For AI Safety Assurance series. Over 8 weeks, I’m breaking down the seven guiding principles that will define how AI gets integrated into aviation safely.
The Fourth Principle
Distinguish between the safety assurance methodology for learned AI implementations and learning AI implementations
Not all AI is created equal. The architecture of the AI system fundamentally determines whether aviation can certify it now or has to research it for years.
Most people think “AI is AI.” The roadmap is saying that is not true. The difference between learned and learning AI changes everything.
Learned AI: The Static Model That Can Be Certified
Learned AI is what most people think of when they think of modern AI.
It’s AI systems trained offline on historical data. The model is fixed at a specific point in time. Once deployed, it doesn’t learn or change from new operational data. Same input always produces the same output.
Here’s a concrete example: A predictive maintenance AI trained on 10,000 hours of engine data. The system learns to recognize degradation patterns in vibration signatures, temperature trends, fuel consumption. The model is frozen. It’s tested, validated, and certified. Deployed on aircraft.
In service, it analyzes new engine data and flags anomalies. But the algorithm doesn’t change. A year from now, it will behave exactly the way it does today. Ten years from now, same behavior. The system is static.
Why learned AI can be certified:
You can test it exhaustively before deployment. You create test scenarios that cover the operating environment. You validate it against historical data. You understand its behavior completely because it doesn’t change.
Once deployed and passed assurance, it’s accepted. In-service monitoring becomes part of continuous operational safety (COS). You’re watching for anomalies, not for the system to evolve.
The FAA is explicit:
“The safety assurance for a learned AI implementation can be performed as part of the system design and validation. Once completed, the AI implementation is accepted, and the in-service monitoring of the AI implementation is part of the continuous operational safety (COS) programme for the aircraft.”
You certify it once. You monitor it continuously. But you’re not constantly re-certifying it.
How updates work with learned AI:
The operator collects in-service data. The developer uses this data to retrain the model. A new version is created. That new version goes through safety assurance again. Only when the new version is certified is it deployed.
This process can take weeks or months. It’s slower than continuous learning. But it’s manageable. It’s safe.
Learning AI: The Dynamic System That Aviation Isn’t Ready For
Learning AI is the scary one. And the FAA is explicitly saying they’re not ready to certify it.
Learning AI systems continue to learn in the operational environment. The model evolves based on new data encountered during actual flights. You can’t test every possible state because the system keeps changing. Same input might produce different outputs as the system learns.
This creates a fundamental problem: How do you assure safety of something that’s constantly changing?
With learned AI, you certify a snapshot. With learning AI, you have to certify a process. A process that evolves.
The FAA captures this challenge in two critical quotes:
“A system that continues to learn in the operating environment must build its safety assurance into the operating environment or include safety assurance as part of the process of learning. Learning systems may necessitate new regulations to assure the continued safety of the evolving system, as for active monitoring of performance or recurrent certification.”
Translation: We don’t have a framework for this yet. It might require continuous recertification. We’re not ready.
And then:
“Learning AI implementations may adapt in a manner that degrades performance, ultimately weakening their original safety profile. Cases in which a system learns anomalous, ungeneralizable, or inaccurate information will require new mitigation strategies.”
This is the core fear: The system could learn its way to being unsafe.
Think about this scenario. An AI system trained to optimize fuel efficiency learns a pattern that saves fuel by reducing engine strain in a particular flight regime. It’s a legitimate efficiency gain. It deploys this “optimization.”
Over time, something changes. Weather patterns shift. Traffic patterns change. The optimization that worked in the training environment starts producing different results. The system adapts, learning new patterns. But these new patterns were never validated. They’ve never been tested against the full operational envelope.
Gradually, the system’s behavior diverges from what was originally assured. It’s learning. But it’s learning in a direction that degrades safety. How do you catch this? How do you stop it before it causes an accident?
Aviation doesn’t have answers to these questions yet. That’s why the FAA is deprioritizing learning AI.
The roadmap is explicit: Safety assurance for learning AI is a research problem, not a near-term deployment problem. The FAA expects learning AI safety assurance to remain in the discovery phase for more than three years. Focused development activities won’t begin until safety assurance methods are proven.
Why This Distinction Changes Everything
The difference between learned and learning AI isn’t just technical. It’s the difference between “we can certify this now” and “we need years of research.”
Here’s what this means for safety assurance methodology:
With learned AI, you need:
Design assurance level (DAL) based on criticality
Exhaustive testing against expected scenarios
Validation that the model performs across the operational envelope
In-service monitoring for anomalies
Controlled updates with recertification
This is hard. But it’s solvable with existing frameworks adapted for machine learning.
With learning AI, you need all of the above, PLUS:
Runtime assurance (continuous validation during operation)
Anomaly detection (catching when learning goes wrong)
Graceful degradation (the system safely reduces capability if learning degrades performance)
Possibly continuous recertification
New regulatory framework (doesn’t exist yet)
This is not solvable with current methods.
The strategic implication is clear: The FAA is not saying learning AI is impossible. They’re saying it requires research and new methods. Until then, learned AI is the gateway technology.
You gain experience with learned AI. You document your processes. You build institutional knowledge about certifying and deploying AI systems. When learning AI research matures, you already understand the domain.
This is how you sequence innovation safely in safety-critical systems.
What This Means Beyond Aviation
Most industries are deploying AI without thinking about whether it’s static or dynamic. Aviation is forcing clarity. And other domains desperately need this distinction.
In healthcare: A learned AI diagnostic system is trained on 100,000 patient records and frozen. It’s deployed. Hospitals can assure it works consistently. A learning AI diagnostic system adapts based on each new patient case. It evolves. How do you assure it? Hospitals are deploying both without this distinction.
In finance: A learned AI fraud detection system identifies patterns in historical fraud. It’s static. A learning AI fraud detection system adapts to new fraud techniques in real-time. It evolves. Banks want learning AI because fraud changes. But they can’t assure learning AI works safely, so they deploy it anyway.
In autonomous vehicles: Current systems are mostly learned AI. They’re trained on driving data and frozen. They’re being deployed with limited success. Future systems might be learning AI, adapting to new road conditions. But we don’t have safety assurance methods for learning AI in vehicles. Aviation is saying: research this first, deploy later.
The principle applies everywhere: Be honest about whether your AI is static or dynamic. Have different assurance frameworks for each. Don’t deploy learning AI without understanding the safety implications. Use learned AI to gain experience before moving to learning AI.
This is the lesson aviation is teaching the world: sequence adoption carefully. Distinguish between what you can assure now and what requires research. Build experience with the former before attempting the latter.
The Pattern Emerging
Four principles in, you’re seeing how this works:
Principle 1: Work within the existing ecosystem. Don’t reinvent governance.
Principle 2: Address safety assurance AND enhancement. Ask both questions.
Principle 3: Be ruthlessly clear about responsibility. Personification erodes accountability.
Principle 4: Understand the AI distinction. Static and dynamic require different approaches.
The through-line is clarity and sequencing. Clarity about frameworks. Clarity about goals. Clarity about responsibility. And now, clarity about the type of AI you’re deploying.
Each principle builds on the previous. And Principle 4 explains why the incremental approach (Principle 5) actually works.
Next Week
Post 6 drops next Friday: Principle 5 — “Take an Incremental Approach”
Now that you understand the distinction between learned and learning AI, you understand why incrementalism matters. You can’t jump straight to learning AI. But you can learn with learned AI, build experience, and then apply those lessons when learning AI research matures.
This principle explains the actual deployment strategy: how to introduce AI into aviation safely, step by step, with clear milestones and decision points.