NASHUA, N.H. – Artificial intelligence (AI) is already supporting many commercial aviation operations, but expanding those systems into safety-critical roles presents a different engineering challenge. Before AI can move beyond decision support, manufacturers must demonstrate that it meets the rigorous certification standards required for every flight-critical system.

That process is not simply a matter of proving that an AI model performs well. Developers must demonstrate that AI behaves predictably, responds appropriately to abnormal conditions, and satisfies certification requirements originally developed for conventional software. Those expectations have guided commercial aviation for decades, but machine learning introduces technical questions that existing certification methods were not designed to answer.

Part one: How AI is already changing commercial aviation—and why it still isn't flying the airplane

Aviation certifies software differently from most industries

Commercial aviation treats software as a critical part of aircraft safety. Before regulators approve software for use on commercial aircraft, engineers must document system requirements and validate that the software performs as intended throughout its operating environment.

Certification also looks at how software behaves when something goes wrong. Engineers evaluate equipment failures, invalid inputs, degraded sensor data, and other abnormal conditions to confirm that a system continues operating safely.

Just as importantly, every function must be traceable. Certification demands that developers connect software behavior to documented requirements and demonstrate through testing that each requirement has been satisfied. That level of documentation and verification is the foundation of aviation software certification.

Machine learning changes the development process

Traditional software follows logic written directly by programmers. Engineers define how the software should respond to specific inputs, then verify that the resulting code performs exactly as intended.

Machine-learning systems are developed differently. Instead of programming every decision explicitly, engineers train models using large collections of operational data. During training, those models identify statistical relationships that allow them to recognize patterns, classify information, or generate predictions.

That difference also changes how regulators evaluate software. They must assess the software itself, the data used to train the model, and the methods used to validate it. Developers must also understand the operating conditions that could affect the model's performance.

Accuracy alone is not enough

High accuracy does not automatically satisfy aviation certification. Engineers also need to understand how a model performs when it encounters conditions that differ from its training data. Rare weather events, unusual equipment failures, incomplete sensor information, or operating scenarios that happen infrequently may all influence model behavior.

Certification therefore extends beyond measuring overall performance. Engineers must establish the limits of a model, identify situations where confidence decreases, and demonstrate that those limitations don’t introduce unacceptable safety risks.

The required evidence depends on the role AI will perform. Software that supports operational planning presents a different certification challenge than software that could directly influence aircraft control.

Explainability becomes part of the engineering process

Traditional software allows engineers to trace a specific output through documented program logic. Machine-learning models make that process more difficult because their recommendations often result from relationships learned during training rather than explicitly programmed decision paths.

As a result, engineers need additional tools to understand why a model reached a particular conclusion. That capability, often called explainability, helps developers investigate unexpected results and determine whether the model's recommendation is supported by the available data.

Explainability also supports certification. That understanding helps regulators determine whether an AI system can be trusted in higher-consequence applications.

Updating AI introduces another certification challenge

Unlike many conventional software systems, engineers may retrain machine-learning models as new operational data becomes available. Retraining can improve performance, but it also creates new certification questions.

Researchers are developing methods to answer those questions as regulators refine guidance for AI-enabled systems.

Regulators are adapting existing certification principles

The Federal Aviation Administration (FAA) and the European Union Aviation Safety Agency (EASA) are developing guidance for AI. Instead of replacing existing certification standards, regulators are examining how established safety principles apply to software that learns from data.

Current work focuses on model validation, data quality, transparency, human oversight, lifecycle management, and continued monitoring after deployment. The objective is to apply aviation's longstanding safety expectations to a technology that behaves differently from conventional software.

Aircraft manufacturers, software developers, and aerospace organizations continue developing practical methods to certify AI-enabled systems alongside regulators.

AI certification will likely advance gradually

Commercial aviation rarely introduces new technology all at once. Manufacturers typically gather operational experience and expand applications as engineering knowledge and regulatory confidence increase.

AI may follow that same pattern. Manufacturers will likely pursue certification first for narrowly defined applications that engineers can thoroughly verify, then expand to more complex systems. More advanced applications will depend on breakthroughs in verification methods, certification guidance, and engineering confidence.

The next challenge is demonstrating that AI can meet the aviation industry's longstanding standards for safety, reliability, and certification.