We’re researching new ways AI can improve safety and risk management.
Aviation should bring in AI the way it brings in anything that affects safety, and test it before trusting it. That’s why we research as well as build.
What we’re working on, and what we’d like to.
Current project · Benchmarks
A benchmark for aviation regulation questions
An open benchmark of how well AI answers these questions. Practitioners write them with help from a model, part of the set is held back, the marking is blind, and we’ll publish the method with the results.
Read about the project
Published · Just culture
Same actions, different endings
Do large language models judge flight crews by how a flight ended? We gave three of them 48 real occurrences, each with a safe and a bad ending, and asked them to apply Reason’s culpability decision tree. All three were harsher when it ended badly.
Read about the project
Published · Safety & compliance
Off-the-shelf AI in aviation safety and compliance
General-purpose AI tools are easy to trust on things they get wrong. Seven ways that happens in safety and compliance work, with an example of each, and what to do about it.
Read about the projectFuture research opportunities
Questions we’d like to work on with partners. None has started yet.
Safety dataOpen to partners
How consistently is an occurrence coded?
Two trained people can code the same occurrence differently. Measuring how often, and where, sets a limit on what any AI trained on those codes can claim.
People and AIOpen to partners
Does a suggested code change the coder?
If an analyst sees an AI’s proposed classification before making their own, do they follow it, even when it’s wrong? The answer decides how a suggestion should be shown.
CompetenceOpen to partners
What does it take to supervise AI?
What someone overseeing an AI tool needs to be able to do, and how that fits the competencies aviation already trains and assesses.
BenchmarksOpen to partners
A benchmark for occurrence classification
A shared, versioned set of occurrence reports with reference classifications, so everyone can check claims about AI classification the same way.
New uses for AI, and the problems it brings.
We want to do this with companies and organisations across the industry, and with universities.
Finding new uses
New ways AI could help keep aviation safe. We look for them with the people who do the work.
Solving the problems it brings
How good are AI’s answers, how far should people rely on them, and how do you show it can be trusted? Our first project starts with how good they are.
Who we’d like to work with.
We need people who do the work, and people who know how to test AI.
Across the industry
Operators, maintenance organisations, airports, air navigation service providers and the organisations around them. That’s where the questions come from, and where an answer has to hold up.
In academia
Universities and research groups who want to look further into what we find.
The research and the product feed each other.
Our product is Chuck, the AI workspace for aviation’s safety and compliance teams.
Everything it gives you is made to be checked. The source is a click away, some checks don’t use AI at all, and everything it writes has a line for someone to sign off.
Research → Chuck
What the research finds shapes what we build
Chuck → Research
How people use it shapes what the research looks at
Talk to us about research
Got an idea for a project, a question you want answered, or something to ask about the benchmark? We reply within one working day.