AI Autonomy's Blind Spot Is the Signal It Trusts

The Signal Under the Autonomy

Most AI assurance conversations in aerospace center on decision authority: how much oversight a human retains as systems move from assistance to teaming to advanced automation. That is the right question for flight control logic. It is the wrong question for the data those systems are built on.

Positioning, navigation and timing data is the substrate underneath most AI-enabled autonomy, from flight management to autonomous approach systems. And PNT is no longer a stable substrate. Industry analysis heading into 2026 points to a threat environment defined by jamming, spoofing and deliberate NAVWAR activity, alongside a market response of diversifying into MEO and LEO PNT sources and standing up sovereign and regional navigation systems as hedges against single-point GNSS dependence Inside GNSS. That is a resilience story for the navigation community. It is an assurance problem for anyone certifying AI that consumes that data.

Learned Models Don’t Know What They Weren’t Shown

The distinction matters because of how these systems are built. Machine learning models fit behavior from data rather than encoding every rule explicitly, which is precisely what makes them useful for perception, reasoning and decision support in autonomous functions Aerospace America. A model trained predominantly on nominal PNT conditions has no principled basis for behaving safely when that input degrades or is spoofed. It will produce an output. Whether that output is safe is exactly the kind of question a certification body cannot answer after the fact.

EASA’s own roadmap treats human oversight as the central variable across its AI levels Aviation Business ME. That framing assumes the AI’s inputs are trustworthy and the open question is what a human does with the AI’s output. When the input itself is the attack surface, oversight of the human-AI loop doesn’t close the gap. The safety case has to extend further upstream, into whether the system was ever tested against degraded, denied or deliberately corrupted navigation data as a first-class condition rather than an edge case.

What This Changes for Program Leads

For programs building AI-enabled autonomy, this is a scoping decision, not a research question. Multi-source PNT architectures that blend GNSS with LEO constellations and regional systems are being adopted for operational resilience Inside GNSS. Those same architectures should be treated as certification evidence generators, producing the degraded-input test conditions an AI safety case needs to demonstrate graceful failure rather than confident error.

The programs that fold adversarial PNT scenarios into their AI verification plan now will walk into certification with evidence in hand. The ones that treat navigation integrity as someone else’s resilience initiative will find out during flight test that their model has an opinion about a jammed signal, and that opinion was never reviewed.

Autonomy doesn’t fail quietly. It fails confidently, on bad data it was never told to distrust.

---

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. The central argument—that AI certification frameworks focus on human oversight of outputs while neglecting adversarial input conditions—is logically coherent and the causal chain from PNT vulnerabilit
Source & Claim Verification	Qwen · local	cleared. All factual claims are supported by citations, but the relevance of some sources could be improved.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately identifies a critical gap in AI assurance for aerospace but does not explicitly map its arguments to ISO 42001, EU AI Act, FDA, or MDR/IVDR requirements.
Technical Accuracy	Llama	cleared. The article accurately highlights the critical issue of AI autonomy’s reliance on potentially compromised PNT data and the need for rigorous testing against degraded or spoofed inputs.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies a critical counterargument to common AI assurance approaches by highlighting the vulnerability of underlying PNT data, and largely avoids vendor hype by focusing on
Novelty & Non-Duplication	Grok	held. The PNT-threat-to-AI-certification link is a competent synthesis of widely wired 2025-26 GNSS/NAVWAR and EASA-oversight coverage rather than a genuinely new claim or angle absent from the existing cat
Validation	DeepSeek	cleared. The central claim that AI models trained on nominal PNT data lack a basis for safe behavior under degraded or spoofed conditions is validated by established machine learning principles and the documen

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