Evidence before authority for aviation AI
An international research ecosystem measuring benefits, risks, and safeguards of specialized LLMs on official aeronautical information.
What technical and human evidence is required to identify the benefits and risks of a large language model specialized in official aeronautical information, and to guide its responsible development and adoption?
More traffic, more information, less margin to search
ICAO’s Global Air Navigation Plan warns that air traffic could double within fifteen years. Every new flow, procedure, and requirement adds information that professionals must locate, interpret, and verify under high responsibility.
Searching faster must never mean verifying less. Capacity of information and people must grow with traffic.
Document friction signals
Exploratory consultation with ~100 aviation professionals across Americas, Europe, and Asia, mainly ATC.
Location described as slow or complex
Difficulty recalling where information is
| Metric | Value |
|---|---|
| Location described as slow or complex | 52% |
| Difficulty recalling where information is | 36% |
Sources consulted most often (exploratory)
Official sources consulted most often
| Source | Consulted % |
|---|---|
| National AIP | 87% |
| Local regulations | 82% |
| METAR / TAF | 79% |
| Letters of agreement | 72% |
Exploratory data, not generalizable scientific evidence. Novelty claims are updated via literature review before each protocol or publication.
Observed access time (exploratory exercises)
| Method | Range |
|---|---|
| Traditional methods | 34-223s |
| Specialized tool | 12-16s |
| Avg. observed savings | ~61s |
Six exploratory exercises with ATC. Not a generalizable scientific protocol; used only to justify rigorous research.
Technology is moving faster than evidence
Generic generative tools are already used informally by aviation students and professionals, even though they were not designed for document validity, regulatory hierarchy, or the cost of an incorrect answer.
Technical risks
Hallucinations, incomplete retrieval, outdated documentation, context loss, inconsistent answers.
Human risks
Overtrust, automation bias, reduced vigilance, incorrect interpretation, weak human-system handoff.
Institutional risks
Data governance, accountability, cybersecurity, training, decision logging, and approval criteria.
First understand, then expand. Higher automation must advance in stages with stop criteria.
Early evaluations run in academic, training, simulation, or controlled research settings, not live ATC or aircraft operations.
An international applied-research ecosystem
Evaluate how an aviation-specialized LLM can support access, understanding, and verification of official aeronautical information, identifying benefits, limits, risks, and responsible design requirements before higher-authority uses.
ATC, piloting, dispatch, maintenance, management, education, aeronautical information, and human factors.
Scientific gap
No open primary study was identified that integrates, in a single platform, traceable consultation of AIP, NOTAM, METAR/TAF, and official documentation across multiple professional roles.
Aeronautical NLP and LLMs
Progress exists on NOTAMs and specific corpora, but computational metrics still dominate without multi-role evaluation on realistic documentary tasks.
Pilot and ATC assistants
Projects measure workload and trust, yet their main focus is not integrated, source-traceable document consultation.
Data and governance
FAA, EASA, EUROCONTROL, ICAO, and EUROCAE advance assurance and standards; those principles still lack multi-role evidence on realistic tasks.
The program aims to evaluate model, document, interface, and user as one system. It does not seek to prove that AI is always faster or better.
Scientific evidence
Search time, accuracy, traceability, workload, trust, usability, acceptance, over-reliance.
Better technology
Findings converted into requirements, controls, interface changes, and abstention criteria.
Better governance
Practical input for institutions, education, staged implementation, and future certification paths.
International capacity
Replicable studies across the U.S., Dominican Republic, Europe, and beyond.
SARA as the experimental platform
SARA (System for Aviation with Real-Time AI) turns dispersed official documentation into a verifiable interaction (text or voice) with source traceability so studies can measure model, document, interface, and user together.
- AIP, NOTAM, METAR, TAF, regulations, procedures, and authorized docs
- Dynamic route and airspace visualization with contextual info
- Private document upload under access controls
- Form assistance (e.g. flight plans) with human review
- Semantic search with documentary references
- Instrumentation for time, sources, interaction patterns, and failure events in authorized studies
Not only “is the answer correct?”, but why users trust, how they verify, and what happens when evidence is insufficient.
Institutional traction
Southern Illinois University (R1)
SIU Carbondale School of Aviation issued a formal letter of support and prepared a preliminary academic estimate. SIU authorizes presenting the project as a topic of research interest, subject to funding and institutional approvals.
ENAIRE Open Innovation
First place in phase one of ENAIRE’s 3rd Business Ideas Contest (Spain), among 70+ international projects. Recognition of problem relevance, not a substitute for scientific validation.
ICAO NACC / AI Community
Participated as focal point in the ICAO NACC AI Governance Survey and was invited to the AI in Aviation Community via iSTARS. Membership does not imply ICAO approval or endorsement of the platform.
Multinational coverage
Documentary and functional coverage across six countries (U.S., Dominican Republic, Spain, Argentina, Haiti, Cuba), with 200+ registered users and ongoing expansion.
Exploratory beta
Closed beta signals (Dominican Republic)
Volume signals
Acceptance outcomes
Exploratory beta feedback, not generalizable scientific evidence. Used to show need and early acceptance.
Who can join
Select a role to see how you can contribute, and prefill your expression of interest.
Sponsorship & early access
Sponsors choose whether their contribution supports research, SARA development, or both. Levels are flexible references; scope and recognition are defined together.
- Recognition in approved press, web, and presentations
- Acknowledgement in papers when editorial policies allow
- Briefings on public results (no control over methods or conclusions)
- Early access for individuals, teams, or institutions by agreement
- In-kind welcome: cloud credits, simulators, engineering hours, translation, authorized access
Research sponsorship is delivered by the sponsor directly to the responsible university. Technology funding is formalized in a separate agreement with the technology party.
Dollar amounts and access packages are not listed publicly; they are tailored in conversation. Co-authorship depends solely on substantive intellectual contribution.
How collaboration works
- 1
Initial conversation
Identify the problem, professional profile, site, and what each collaborator can contribute.
- 2
Scientific lead
Define the university or lead researcher and choose scope: pilot, small study, area line, full study, or replica.
- 3
Protocol design
Metrics, instruments, sample, responsibilities, budget, data governance, and ethics review.
- 4
Configure SARA
Sources, permissions, version, logging, controls, and technical support for the study.
- 5
Execute & publish
Independent analysis, publication, and translation of findings into product improvements or new questions.
Scientific leadership always sits with the university or principal investigator. Platform infrastructure remains with SARA / Nolim Studios; preexisting IP stays separate from scientific results.
Expression of interest
Tell us how you want to participate. Submitting this form creates no automatic commitment for either party.
Olga Isabel Pérez Encarnación
Creator and co-director of the research program
SARA · Dominican Republic / United States
Email: saraapp.dev@gmail.com