AI crash reconstructions expose gaps in aviation safety regulation

The old accident record was built to be read slowly. Flight data, cockpit voice recordings, maintenance histories, weather information, photographs, interviews, metallurgical reports, docket exhibits: these materials were meant to support reconstruction by investigators and, later, careful review by regulators, parties, courts, insurers, and families. They were not built for a world in which a public file can be scraped, modeled, voiced, animated, and returned to the internet as a seemingly immediate version of what happened.

That is the pressure point now forcing aviation safety regulators to rethink what happens after crash discovery. The important development is not simply that AI tools can make crash materials more vivid. It is that synthetic reconstructions can sit uneasily beside official records while appearing, to non-specialist audiences, more complete than the records themselves.

Reports that AI systems have reconstructed cockpit audio from public NTSB investigation files have been circulating through aviation and technology outlets. The underlying primary account is not fully accessible from the available materials, so the technical claims should be treated cautiously. The firmer institutional fact is the response attributed to the NTSB: public access to more than 40 investigation cases has reportedly been restricted while the agency develops AI governance under its OMB M-25-21 compliance plan. That administrative move matters more than any single viral reconstruction because it shows the regulator already sees a problem in the reuse of investigation files, even before the law has settled what to call the output.

The record is public, but the reconstruction is something else

Aviation investigation depends on controlled categories. A cockpit voice recorder is one thing. A transcript is another. A factual report, a party submission, a safety recommendation, and a probable-cause determination each carry different institutional weight. The discipline of the system is that these categories do not collapse into one another just because they describe the same event.

AI reconstruction disturbs that discipline. A model-generated cockpit-audio sequence may be trained or prompted from public documents, but it is not the cockpit voice recorder. An animation may be consistent with selected flight data, but it is not the data. A voice synthesis may feel evidentiary because it is emotionally legible; that is precisely why it needs sharper handling, not looser handling.

The NTSB’s ordinary safety-recommendation process is aimed at identifying safety problems and urging corrective action; the agency tracks recommendations and their status as part of that public safety function.[1] That process assumes a record capable of being disclosed, interpreted, challenged, and separated from advocacy. AI reconstructions introduce a secondary object: not the underlying evidence, but an interpretation generated from evidence-like material and packaged in a form that may influence public and legal judgment before any official finding is complete.

That is why the access restriction reported around more than 40 cases should not be dismissed as a records-management footnote. Closing or limiting access to investigative materials changes who can examine the record, who can independently test official conclusions, and who can prepare for litigation. It also signals that open accident dockets, long understood as tools of transparency, can become inputs for synthetic outputs that the agency does not control.

A regulator can move before it has a vocabulary

Regulators often reveal pressure first through procedure. Before there is a final rule, a guidance document, or a settled evidentiary category, there may be a narrower release policy, a revised disclosure practice, an internal review, or a compliance plan. That is what makes the NTSB response significant. It is not a complete AI evidence framework. It is an early defensive adjustment by the institution that owns the investigative record.

For legal professionals, the practical question is not whether the NTSB has solved the problem. It plainly has not. The question is what follows when crash-discovery materials are no longer merely documents to be read, but datasets to be converted into persuasive synthetic media. Public access, party access, protective orders, discovery requests, evidentiary objections, and expert disclosures all become harder when the contested item is neither a raw record nor a conventional expert opinion.

A reconstruction might be created by a journalist, a litigant, an insurer, a consultant, an expert witness, a family member, or a vendor selling analytic tools. Each origin changes the legal analysis. A regulator’s internal model output may raise administrative-record and due-process questions. A plaintiff’s demonstrative simulation may raise authentication and prejudice issues. A defense expert’s AI-assisted timeline may raise discovery and methodology questions. A viral video made outside litigation may still affect jury pools, settlement posture, and reputational pressure.

The aviation system has seen regulatory tightening after major crashes before. The Boeing 737 MAX crashes led to changes in how U.S. regulators approached aircraft design changes and required safety information disclosures by manufacturers.[2][3] But those reforms concerned certification, disclosure, and design oversight. AI crash reconstruction presents a different problem: the accident has already happened, the record already exists, and outsiders can create new artifacts from it.

The existing investigation frameworks do not map cleanly onto AI-generated evidence

FAA Order 8020.11D and ICAO Annex 13 matter because they represent the procedural architecture around accident and incident investigation. They define responsibilities, coordination, notification, evidence handling, and the safety purpose of investigation. They were not written around a world in which machine-generated reconstructions derived from accident records would circulate as quasi-evidence.

ICAO has continued to strengthen the global framework for accident investigations, including measures addressing conflicts of interest.[4] That kind of reform goes to institutional independence and credibility. It does not, from the available materials, answer whether an AI-generated cockpit-audio reconstruction is a record, a demonstrative exhibit, expert work product, investigative assistance, synthetic media, or something else.

That category problem is not academic. The label determines who must disclose the item, whether the underlying model and prompts must be produced, whether the output can be authenticated, whether a party can meaningfully cross-examine it, and whether a court treats it as probative analysis or prejudicial theater. Aviation safety investigation was designed to reconstruct causal sequences from physical, recorded, and testimonial evidence. AI-generated reconstructions add another layer: a machine-mediated representation of the sequence.

The gap is especially visible because aviation is not ignoring AI in general. The FAA issued an AI roadmap in August 2024, with principles that include avoiding personification of AI, differentiating learned AI from learning AI, and taking an incremental approach that starts with low-risk applications.[5] Those are sensible principles. They are also aimed mainly at AI adoption in aviation systems and agency activity, not at the evidentiary status of AI outputs after an accident.

AI in the aircraft is not the same as AI after the crash

A great deal of aviation AI governance concerns operational systems: avionics, flight assistance, maintenance analytics, air traffic management, autonomy, and certification. Those debates are difficult, but at least the regulatory target is visible. The question is whether a system can be safely designed, certified, monitored, updated, and used in an aviation environment.

Post-accident AI is different. It may never fly an aircraft. It may never be installed on a certified product. It may be built by an outside vendor using public materials. It may operate after the event, outside the certification perimeter, but still affect liability, regulatory conduct, and public understanding. A rulebook for AI in the cockpit does not automatically govern AI applied to the wreckage file.

EASA’s developing AI framework shows how much aviation AI governance already exists on the operational side. Its approach uses levels of AI authority, moving from Level 1 human assistance toward Level 3 advanced automation. Reported planning horizons indicate that Level 2 and Level 3A approvals are not expected before 2035, and Level 3B not before 2050; those are planning horizons described in industry analysis, not statutory deadlines.[6] EASA’s NPA 2025-07 also prohibits continuous or online learning in certified environments.[6]

That prohibition is revealing. Certified aviation systems are being approached with caution because regulators need bounded behavior, version control, validation, and auditability. But a post-accident reconstruction tool can be outside that certified environment while producing material that looks authoritative. The model may change. The training data may be opaque. The prompt may be unavailable. The output may be reproducible only if someone preserved the exact system state at the time of generation.

The EU AI Act adds another layer by classifying AI systems in air traffic management and AI used as safety components in aviation products as high-risk, with obligations around conformity assessment, risk management, documentation, and human oversight.[7] That overlay is important, but it still points mainly to AI systems embedded in aviation operations or products. The synthetic reconstruction made after the crash remains harder to place.

The evidentiary questions arrive before the regulations

A court or agency does not need a comprehensive aviation AI rulebook before the first dispute appears. It needs to decide what to do with a particular output in a particular proceeding. That is where legal professionals should expect the hard questions to surface first.

The most immediate disputes are likely to involve provenance. Who created the reconstruction? What source materials were used? Were any public docket files omitted? Did the tool infer missing data? Was the cockpit voice generated from actual audio, transcript text, flight parameters, or a narrative account? Was the output edited for clarity, duration, or dramatic effect? If the model hallucinated a sound, phrase, warning, or sequence, how would anyone know?

Those questions are familiar in digital evidence practice, but aviation adds a special sensitivity. Accident investigation is not supposed to be a free-for-all search for blame. It is a safety process with legal consequences orbiting around it. Once a synthetic reconstruction begins shaping liability theories, witness memories, public campaigns, or settlement positions, the distinction between safety investigation and adversarial use becomes harder to maintain.

This is why analogies to AI video evidence are useful but incomplete. A disputed synthetic video in an ordinary civil or criminal matter may turn on authenticity, prejudice, and chain of custody. An AI crash reconstruction adds the institutional status of the accident record itself. The question is not only whether the output is reliable; it is whether the output is being mistaken for a safety finding.

Auditability is the hinge

The strongest version of AI-assisted investigation is not theatrical. It is analytic. A tool might help align time stamps, compare flight parameters, flag anomalies, identify maintenance-pattern correlations, or test whether a proposed sequence is consistent with known data. Used that way, AI can make hidden sequences easier to inspect. The problem begins when the output becomes persuasive faster than it becomes auditable.

A 2026 systematic literature review on AI for aviation safety reflects growing research interest in using AI to enhance safety, but adoption and demonstrated legal reliability are not the same thing.[8] The aviation industry has long moved cautiously with safety-critical AI; a Berkeley CLTC report observed that the lack of AI-based standards and regulations has constrained adoption, while also identifying collaborative standard-setting tendencies in the sector.[9] That caution is often frustrating to technologists. In accident investigation, it is also what prevents a plausible sequence from becoming an official sequence merely because software rendered it cleanly.

Auditability requires more than saying that a tool used AI. It requires enough information for affected parties to understand the source inputs, the model behavior, the assumptions, the uncertainty, and the human choices embedded in the output. If a reconstruction is offered in litigation, a party may need to know whether it is deterministic, whether repeated runs produce the same result, whether the model used external training data, and whether the vendor can separate factual inputs from generated inferences.

The FAA roadmap’s instruction to avoid personifying AI is more than a communications preference.[5] In evidence disputes, personification becomes a reliability trap. A model does not “hear” the cockpit, “know” the crew’s intent, or “remember” the sequence. It generates or analyzes outputs according to system design, training, inputs, and constraints. Treating the tool as an observer smuggles authority into the record without cross-examination.

Other safety-critical fields are already wrestling with the same boundary

Aviation is not alone in this evidentiary uncertainty. Medical AI diagnostics can influence treatment decisions while raising questions about validation, explainability, and responsibility. Autonomous vehicle crash analysis can involve sensor data, software logs, simulation, and post-event reconstruction. In both settings, the hard problem is not simply whether AI is useful. It is how to distinguish operational data, model inference, expert interpretation, and legally accountable decision-making.

The comparison has limits. Aviation accident investigation has its own international architecture, safety-reporting culture, and restrictions on using certain materials. But the common lesson is plain enough: when AI outputs are introduced into a safety-critical evidentiary chain, the law needs provenance, validation, and contestability before it needs fascination.

What legal teams should treat as contested from the start

No lawyer needs to wait for a final FAA, NTSB, EASA, or ICAO rule before treating AI crash reconstructions as contested evidence objects. The first discipline is separation. Keep the official record separate from synthetic interpretation. Keep the raw input separate from the model output. Keep the expert’s opinion separate from the tool’s generation. Keep public media artifacts separate from materials created for litigation or investigation.

• Identify the creator of the reconstruction and the purpose for which it was made.
• Preserve the model name, version, settings, prompts, source files, and output history where available.
• Ask which portions are directly sourced from official records and which portions are inferred or generated.
• Test whether the output is reproducible and whether alternate inputs produce materially different sequences.
• Avoid describing synthetic audio, animation, or timeline material as an official finding unless the investigating authority has actually adopted it.

This is not a rejection of AI-assisted reconstruction. Some tools may become valuable to investigators and experts, especially when they make complex sequences easier to test. But the legal system should resist the shortcut in which visual or audio immediacy substitutes for evidentiary status. A polished reconstruction can be useful, misleading, or both.

The current regulatory picture is therefore uneven rather than empty. The FAA has an AI roadmap. EASA has a developing framework for AI authority and certification. The EU AI Act creates high-risk obligations for certain aviation AI systems. ICAO continues to refine accident-investigation governance. The NTSB has reportedly restricted access to more than 40 cases while developing AI governance under a federal compliance plan. Regulators have started to respond.

But most of that activity still governs AI in aviation systems, not AI applied to aviation evidence after the accident. That is the blind spot. Until it closes, AI crash reconstructions should be handled as synthetic, challengeable, source-dependent artifacts: potentially useful, never magical, and not to be confused with the black box itself.

References

1. NTSB Safety Recommendations – Statistics and Process, NTSB, Updated July 10, 2026.
2. FAA adopts strict new policy on aircraft design changes in wake of deadly Boeing 737 MAX crashes, GlobalAir, Nov. 2023.
3. After Boeing Max crashes, U.S. regulators detail safety information that aircraft makers must disclose, PBS, July 2023.
4. ICAO strengthens global framework for accident investigations, addresses conflicts of interest, ICAO.
5. FAA issues roadmap for AI in aviation, GlobalAir, Aug. 2024.
6. Navigating the Runway: How EASA's AI Framework Will Reshape Aviation Safety, Halldale Group, April 2026.
7. AI: Innovation in Aviation, Oracle Law Global.
8. Enhancing aviation safety with artificial intelligence: A systematic literature review, ScienceDirect, 2026.
9. The Flight to Safety-Critical AI: Lessons in AI Safety from the Aviation Industry, CLTC Berkeley, 2020.