By: Charlton Evans, Autonomy Global Ambassador – Certification
Something significant has happened inside the FAA’s waiver and exemption review process, and most of the aerospace industry hasn’t fully registered it yet. The agency is deploying artificial intelligence (AI) to assist in the evaluation of regulatory applications. It’s doing so across a spectrum, from using structured institutional tools to inspector-level prompting against the FAA’s own regulatory corpus. Meanwhile, the companies filing those applications also increasingly leverage AI to draft them. As AI now works on both sides of the equation, it has profound implications for processing speed, consistency, feedback quality and for the risks that come with embedding machine intelligence into a regulatory process. Here’s my take.
There Is A Dual Mandate
The “Unleashing American Drone Dominance” Executive Order 14307, signed by President Trump on June 6, 2025 directs the FAA to initiate the deployment of AI tools to assist in and expedite the review of UAS waiver applications under 14 C.F.R. Part 107. It gave the agency a 120-day deadline from the order’s signing date (i.e., by October 4, 2025) to start that process.
The FAA’s own internal interim AI policy under Notice N 1370.52 (effective March 2025), which governs how FAA personnel may use generative AI (GAI) tools across internal workflows say that FAA organizations, managers and employees must not use AI to: “Be cited as direct evidence or authority for a determination/decision” and use must be backstopped by human review.
The fact of the matter is that the FAA was evaluating waiver applications under Part 107 with GAI. The notice simply made explicit what was already developing behind the scenes already inside the agency.
This isn’t a speculative future. It reflects the current state of the FAA’s operating reality. Here’s how they use GAI, in at least two distinct ways. The first is the employment of institutional and structured purpose-built GAI tools with defined criteria against which the agency evaluates applications. The FAA has already evaluated UAS operational data from Part 107 waivers to identify trends and characterize the sufficiency of information provided in waivers that resulted in approvals, as well as the insufficiency of information in those that did not receive approval. That trend analysis provides precisely the kind of structured, labeled dataset that trains a model to recognize if the elements of a safety case are present or if they are not.
The second form is more flexible, and perhaps more consequential, in the near term. It involves individual inspectors using FAA-sanctioned GAI tools to prompt analysis of a given application against the regulatory text, legal interpretations, orders and advisory circulars that govern the operation in question. Notice 1370.52 establishes the agency’s interim policy on this use of generative GAI for all FAA employees and contractors and formalizes it as a capability that enables case-by-case AI-assisted review. What does this mean? The inspector isn’t just applying his or her own interpretive judgment anymore. They’re getting a second, or possibly even first, opinion from a machine trained on a regulatory library and historic approvals.
There Is Hope
This use of GAI could prove attractive for both the FAA and to industry. The Part 107 waiver backlog has been a persistent pain point for operators trying to reach revenue operations. Applications sometimes languish. Feedback, when it comes, is often generic. The gap between what an applicant intended and what an evaluator needed has historically been bridged only through multiple rounds of correspondence or often, simply not bridged at all. This resulted in waiver denials. GAI-assisted review, done well, could change that calculus substantially.
The first benefit could be consistency. Every application would be evaluated against the same criteria, without the variability that comes from reviewer experience, workload or interpretation. The FAA’s own data shows that many waiver applications are brief two sentences or less and do not provide adequate information to effectively assess the risk of the desired operation. A well-configured AI tool can flag that deficiency at intake, rather than weeks into review. This would enable faster feedback and better applications.
Speed also follows from consistency, as long as we trust the outcomes. If structured criteria can be applied automatically to screen and score applications, reviewers can focus their analytical effort on the genuinely novel questions, the ones that require human judgment about the application of rules to unprecedented circumstances. That seems like a better use of expert capacity.
For applicants, this also creates an opportunity. If the FAA uses GAI to evaluate applications, sophisticated operators can use AI to pressure-test their submissions against the same criteria before filing. In theory, this might result in a tighter feedback loop and higher-quality applications reaching the agency. That would benefit everyone.
There Is Risk
On the other hand, the concerns in using GAI to evaluate waiver applications are real and worth naming directly.
Bias Inheritance
GAI systems learn from historical data. The FAA’s approval and denial history reflects the agency’s past interpretive posture. That history has not always been consistent or favorable to innovation, BVLOS operations or novel concepts without established precedent. A model trained on that history will tend to favor applications that look like previously approved applications. That might be the exactly wrong filter to evaluate genuinely new operations that don’t fit the historical mold. The first commercial BVLOS pipeline patrol authorization didn’t look like anything that had been approved before. It shouldn’t have been rejected because an AI couldn’t find a precedent.
Prompt Variability And Inspector Discretion
When individual inspector prompting drives an AI review, the quality of the analysis depends on the quality of the prompt. Two inspectors evaluating similar applications may ask their AI tools fundamentally different questions. This would reintroduce the inconsistency that structured AI was supposed to eliminate. This is not a hypothetical risk. It is the natural consequence of deploying general-purpose generative tools without standardized query frameworks.
Hallucination And Regulatory Misapplication
FAA Notice 1370.52 explicitly requires that all generative AI output undergo human review for validity, accuracy and completeness. That requirement exists because the agency knows its own tools can be wrong. GAI trained on regulatory text can produce confident-sounding analysis that misapplies an order, miscites an Advisory Circular or draws an incorrect inference from a legal interpretation. In an administrative process with no real-time correction mechanism, a confident but incorrect AI analysis that shapes a reviewer’s conclusion could present a serious problem.
The Opacity Problem
When the FAA denies or conditions an application with extensive limitations, the applicant deserves to understand why. If the reasoning behind that outcome is partially or substantially generated by an AI tool, applicants have a right to know that, and to know the basis on which the AI was prompted. The administrative record must be transparent enough to enable meaningful response. A black-box denial is not an acceptable outcome in a safety-critical regulatory process.
AI tools need governance built in by the owners, rules that prevent them from speculating or “guessing” for lack of facts and data. I had an AI tool craft me a letter recently and it added a name to the signature block that wasn’t actually my name. When I asked the AI why, it apologized and said that my name had “fallen out of memory” despite my robust master prompt. The next prompt I created had governance parameters to establish a “no guessing, flag for verification of all unconfirmed information.” But I still read the signature lines.
Both Parties Are Learning
Here’s the deal. The FAA is learning to use AI just as the industry is learning to use AI. Neither party has this fully figured out. That is not a criticism. It is the reality of deploying a genuinely new class of tools against problems that were previously solved by human expertise alone.
What it means practically is that the current moment calls for dialogue, rather than adversarial positioning. Industry practitioners who understand both the regulatory requirements and the failure modes of AI tools have a responsibility to engage constructively with the FAA on how AI-assisted review should be structured, validated and governed. The FAA, for its part, has an obligation to be transparent about how AI is being used in its review processes. The integrity of the regulatory process depends on it.
The human backstop is not optional. The OMB guidance with which the FAA must comply requires that agencies ensure there is sufficient training and periodic retraining so that operators of AI systems can appropriately interpret and act on AI outputs, including the ability to recognize and manage situations where the AI generates erroneous or inappropriate outputs. The same obligation applies on the industry side. When you use AI to draft a waiver application, you are responsible for what it says. When the FAA uses AI to evaluate that application, a human reviewer is accountable for the outcome.
The Takeaway For Applicants And Operators
As a member of industry, you need to know this is happening. Structure your applications accordingly with explicit, unambiguous safety case logic that maps clearly to the FAA’s stated evaluation criteria. Don’t make an AI reviewer work to find your argument. Put it in the document, in plain language, in the right sequence.
Understand that consistency is now a feature of the system. That cuts both ways. Strong applications will benefit from consistent, criteria-based evaluation. Weak applications that previously slipped through on reviewer discretion will not. If your case is novel, you probably should say that up front.
Outside of your own application, engage in the larger governance discussion. How AI is used in the FAA’s review process is a policy question, not just a technical one. The companies that will shape how this plays out are the ones participating in the public conversation through comment, collaboration and demonstrated expertise.
Clearly, AI sits on both sides of the table now. The question isn’t whether we all need to adapt to that reality. We do. The real question is whether we all are going to do it deliberately.