Most ATO instructors first heard about NPA 2025-07 in a group chat, not from their competent authority. The document runs to over two hundred pages, references six AI authority levels, and connects to an EU AI Act enforcement deadline that is already running. For a head of training managing a roster, a simulator booking sheet, and a Part-ORA audit, that is a lot to absorb before the next student walks in.
EASA NPA 2025-07 Is the First Regulation That Treats AI as an Aviation-Wide Problem
EASA describes NPA 2025-07 as 'the first regulatory proposal of its kind in global aviation.' No equivalent exists at the FAA, ICAO, or Transport Canada. That claim is worth sitting with, because it explains why the document matters far beyond the training community.
What makes it structurally different
Previous AI-adjacent guidance in aviation was either domain-specific or advisory. NPA 2025-07 does neither.
Rather than amending Part-FCL or Part-ORA, it proposes an entirely new standalone instrument: a set of Detailed Specifications (DS.AI) with associated Acceptable Means of Compliance and Guidance Material. The EASA Explanatory Note states explicitly that no existing regulation is being amended. This is a new layer that sits above all domain rules and will connect to them in a second NPA planned for 2026.
The scope is genuinely horizontal. Affected parties include:
- Aeroplane operators
- Helicopter operators
- Training organisations (Part-ORA)
- Maintenance organisations (Part-145)
- Design organisations (Part-21)
- Flight crew licensing authorities (Part-FCL)
- Air navigation service providers
That breadth is deliberate. A single framework avoids the inconsistency that would result from each domain developing its own AI approval criteria independently.
The authority classification and what it limits
DS.AI.110 defines six sub-levels across three tiers, from Level 1A (AI assistance to humans) through Level 3B (AI decisions that cannot be overridden). The current NPA covers only Levels 1 and 2.
Andrew Mitchell, Head of Training at FTE Jerez, confirmed that EASA does not expect Level 3A approval before 2035, with Level 3B not anticipated until 2050. For training organisations, that timeline sets a concrete ceiling. Any AI tool deployed in a certified training environment for the foreseeable future must sit within Level 1 or Level 2, and the DS.AI trustworthiness framework applies to all of them.
Understanding that regulatory boundary is only the first step. The harder question for most ATOs is whether their internal organisation is ready to operate within it.
The Barrier to AI Adoption in ATOs Is Not the Regulation. It Is What Is Inside the Organisation
The compliance window looks wider than it is. An ATO can, in principle, deploy an AI-assisted scheduling tool or adaptive ground school platform under existing EASA frameworks. The regulatory permission exists. What most small ATOs do not have is anything close to the internal infrastructure that makes that permission usable.
The governance gap no one is measuring
Research published in Frontiers in Education (Gorbunova et al., 2026, N=105 instructors) found that external institutional barriers to AI adoption scored significantly higher than personal skill deficits (mean 3.43 vs. 2.81, p<0.001). The lowest-scoring barrier of all was personal digital competence. Instructors are not the bottleneck. The absence of institutional infrastructure is.
The same study found 84.76% of instructors would integrate AI tools if appropriate technical and methodological support were in place. That support does not exist in most ATOs.
The ITU's AI Readiness Assessment Framework identifies six foundational factors for deployment readiness: data availability and governance, digital infrastructure, skills, research and innovation capacity, standards, and sandbox environments. Small ATOs typically lack structured capability in at least four of those. They hold student performance records, simulator logs, and instructor assessment data in formats that no AI tool can reliably ingest, and they have no formal policy governing any of it.
The Article 4 clock is running
EU AI Act Article 4 requires all deployers of AI systems to make sure staff have sufficient AI literacy, with national enforcement beginning 3 August 2026. Compliance does not require formal certification, but it does require documented training efforts, a records burden that lean ATO administrations are poorly positioned to carry.
The 70% of AI projects that never reach production share one root cause: no readiness framework before deployment. Organisations that assess readiness first spend 2.3 times less and reach production 40% faster. For an ATO considering AI tools under competitive pressure, that figure is not abstract. The evidence on what those tools can actually deliver in a training context is worth examining before any procurement decision is made.
What the Evidence Actually Shows About AI Simulator Effectiveness in Pilot Pre-Training
The strongest published evidence comes from a 2025 study by Ryan Guthridge at the University of North Dakota, published in the Journal of Aviation/Aerospace Education & Research. With 37 student pilots enrolled, Guthridge used Roscoe's Transfer Effectiveness Ratio to measure whether AI simulator pre-training produced real gains in the pre-solo block of Private Pilot training.
The results were specific and positive across all three outcome variables measured:
- Total flight hours to solo: 62.4 hours (treatment) vs. 64.8 hours (control), Δ = −2.4 hours
- Instructor-logged deficiency count: 7.2 (treatment) vs. 11.4 (control), Δ = −4.2 errors per student
- Checkride pass rate on first attempt: 89% (treatment) vs. 76% (control), p=0.031
Guthridge's own framing is the clearest way to read that number: for every 10 hours of AI simulator pre-training, students saved 3.4 hours of aircraft time on average. That is not a 1:1 substitution, but it is a measurable, positive transfer to real flight training.
The learning science behind why this works points to two factors. Immediate corrective feedback after each attempt, and repeated practice distributed over time, both drive procedural skill acquisition. A 2015 randomised controlled study by Bosse et al. found that high-frequency feedback after each repetition produced significantly better procedural performance (p=0.004) compared to low-frequency feedback. Repetition alone also mattered: both feedback groups improved substantially from baseline (p<0.001), suggesting that volume of deliberate practice is doing real work independent of feedback quality.
Where the evidence runs out is worth stating plainly. No peer-reviewed study has specifically measured ATC radio communication training outcomes from AI-based tools. The Guthridge study measures pre-solo flight efficiency, not phraseology accuracy or readback error rates. Adacel's ICE tool reports that 82% of users found it useful for practice, but that is satisfaction data, not a controlled outcome study. The case for AI radio training tools rests on solid adjacent evidence, not direct proof. That distinction matters when an ATO must document the rationale for introducing such a tool under Part-ORA.
What Level 1 AI Compliance Looks Like in an ATO's Radio Communication Training Programme
Before an AI radio practice tool reaches a single student, the head of training has regulatory work to complete. What follows is the sequence that matters.
Before deployment
Under ORA.GEN.130, introducing an AI-based training aid that modifies your training programme triggers a prior approval requirement with your competent authority. Submit the application at least 30 days before the intended change.
Simultaneously, Article 4 of the EU AI Act requires documented AI literacy training for any personnel deploying the tool. That means the CFI completes training before students use the system, not alongside them. The training manual must then be updated under ORA.ATO.130 to describe the tool's intended use, its limitations, and the instructor's oversight role. A student handout does not satisfy this.
The ATO's annual organisational review under ORA.GEN.200(c) must capture the introduction as a significant change, with hazard identification documented before deployment. The risk assessment must address one specific scenario: what happens if the AI gives incorrect phraseology feedback, and how does the instructor catch it before that error reaches a live frequency.
During and after each student session
EASA's AMC guidance on CBT tools sets the expected standard: instructor briefing before each AI practice session, debrief after. For a student using the tool between lessons, this means the CFI reviews session logs at the next ground briefing. Almost 80% of pilot radio transmissions contain at least one error; if a student's logs show a persistent readback pattern problem, that item goes on the next briefing agenda.
Two further obligations apply throughout:
- Session logs may constitute training records under ORA.ATO.120, requiring three-year retention with a backup updated within 24 hours.
- Any software version change to the AI tool must be notified to the CFI, who re-validates phraseology accuracy against current ICAO standards before continued use.
- Where AI feedback conflicts with local ATC procedures, the instructor must be able to explain the discrepancy in plain language. If the tool's logic cannot be explained, it should not be used.
The documentation habits an ATO builds during Level 1 deployment will determine how well-positioned it is when the regulatory framework tightens.
The Level 1 Window Will Not Stay Open Indefinitely: What ATOs Should Build Before 2035
EASA does not expect any Level 2 or Level 3A AI system to receive regulatory approval before 2035. That sounds like a long runway. It is not, once you account for what needs to be in place before the compliance bar rises.
NPA 2025-07 is explicitly described as the first stage of a two-stage rulemaking effort. The second NPA will link EASA's AI trustworthiness framework directly to existing aviation regulations, which means the documentation and assurance requirements ATOs face today are a floor, not a ceiling. ATOs that treat current Level 1 deployment as a procurement decision rather than a safety management activity are accumulating compliance debt that will be expensive to clear under Level 2 pressure.
Three things ATOs can build now, while the requirements are still manageable:
- Data infrastructure that meets EASA's stated expectation of measurable performance indicators, traceable records, and documented deviation detection, because retrofitting this after Level 2 rules are finalised will cost significantly more than building it once.
- AI literacy programmes that satisfy Article 4 of the EU AI Act, stratified by role: instructors, examiners, safety managers, and administrators each interact with AI systems differently and need different levels of understanding.
- Explainability documentation for any AI tools already deployed, given the European Cockpit Association's clear position that opaque systems are unsuitable for aviation regardless of their current classification.
Two things remain genuinely unresolved. EASA has not published specific ATO adoption benchmarks or outcome metrics for Level 1 deployments, so there is currently no regulatory yardstick for what good looks like. And it is not yet clear how evidence gathered during this window will feed into the second NPA. What is clear is that the FAA's parallel roadmap was shaped primarily by operational data from early deployments. The quality of ATO documentation between now and 2035 will likely influence how Level 2 rules are written, not just whether a given ATO can comply with them.
Source: EASA, EUROCONTROL, FAA, Studies refferenced in text