Pilots Are Flying With AI Already. Nobody Has Agreed How to Check It.

The gap is methodological, not jurisdictional

Most aerospace AI coverage this year has focused on where autonomy gets certified. The more consequential question is whether anyone knows how to certify AI at all, even inside aircraft that still have a human at the controls. The Royal Aeronautical Society’s review of AI in the cockpit lands on a blunt conclusion: the FAA has reached the same place as much of the industry, which is that there is still no settled method for AI safety assurance, even as manufacturers keep building AI into training tools and decision support systems aerosociety.com. This is not a debate about whether AI belongs in aviation. It is already there. It is a debate about whether regulators can verify it before it scales further.

That distinction matters for anyone budgeting a certification program right now. DO-178C and its peers were built for deterministic software, where the same input reliably produces the same output and every path can be traced and tested. Machine learning systems used for pilot training feedback or in-flight decision support do not behave that way, and neither the FAA nor EASA has published an equivalent framework for probabilistic, learning-based systems that airlines and OEMs can build against with confidence aerosociety.com.

Why the electric aviation precedent does not transfer

It is tempting to assume this resolves the way novel airframes have. The Velis Electro, a battery powered trainer aircraft, secured EASA type certification and was then validated by the UK Civil Aviation Authority and Transport Canada, while still operating under exemptions from the FAA nomadlawyer.org. That is a real model for cross-border reciprocity, but it worked because the underlying engineering problem, a liquid-cooled battery system replacing a combustion engine, was still deterministic and testable against existing airworthiness logic. AI safety assurance does not have that shortcut. There is no existing rulebook to extend.

Where the assurance function is being built instead

Some jurisdictions are treating this gap as an opening rather than a stall. The Lowy Institute argues that countries with deep experience regulating aviation, mining, and workplace safety are well positioned to build specialized capability in testing, certification, and human-machine interaction assurance for physical AI systems generally, aviation included lowyinstitute.org. Meanwhile, insurers and risk advisors are not waiting for regulatory consensus. Firms tracking liability exposure across autonomous systems and advanced aviation technology are already pricing risk into AI-enabled aircraft programs, which means underwriting standards may arrive before certification standards do us.bbrown.com.

The decision in front of compliance leads

Airlines and OEMs deploying AI-assisted training or cockpit decision tools should not wait for the FAA or EASA to publish a finished assurance method before building their own evidence trail. That means documenting model behavior, human factors testing, and override logic now, aligned to the risk management structure ISO 42001 already expects of AI management systems, so the paper trail exists before an insurer or regulator asks for it. The airframe can be certified in isolation. The AI running inside it cannot.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. Core argument is coherent and well-structured—the distinction between jurisdictional debates and methodological gaps is genuinely useful—but the claim that ‘underwriting standards may arrive before ce
Source & Claim Verification	Qwen · local	cleared. All factual claims are supported by citations, but the article could benefit from more specific references to the cited sources for clarity.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing correctly identifies ISO 42001’s risk management expectations but does not detail specific clauses or controls from the standard, nor does it explicitly address EU AI Act tiers or FDA/MDR
Technical Accuracy	Llama	cleared. The article accurately highlights the challenge of certifying AI in aviation due to its non-deterministic nature and the lack of a settled method for AI safety assurance.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively anticipates and counters potential arguments, particularly regarding certification precedents, and avoids vendor hype by focusing on the challenges rather than solutions.
Novelty & Non-Duplication	Grok	held. The core thesis—that aviation already uses cockpit/training AI without a settled non-deterministic assurance method—is long-running trade-press consensus (EASA/FAA/EUROCAE/DO-178C limits), not a new p
Validation	DeepSeek	cleared. The central claim that there is no settled regulatory method for AI safety assurance in aviation is validated by the cited industry review and the absence of an equivalent to DO-178C for probabilistic

Sources cited: 7. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.