Why we looked
Just culture rests on the principle that flight crew’s actions and decisions are judged on what they knew and did at the time, and the outcome is excluded from this test. European law writes that into its definition.
Just culture, in EU law
“…front-line operators or other persons are not punished for actions, omissions or decisions taken by them that are commensurate with their experience and training, but in which gross negligence, wilful violations and destructive acts are not tolerated.”
Regulation (EU) No 376/2014, Article 2(12)
The application of this principle can be difficult in practice. Human biases can lead the same decisions to look worse, and the same person more to blame, once we are given context of the outcome1. Flight crew are also susceptible to this: we see pilots judge other pilots’ weather decisions more favourably when the flight ended well2.
As we introduce AI into the system, opportunities to address these biases are apparent. AI is starting to read safety reports, suggest root causes & contributory factors, and help decide follow up actions. We wanted to know whether AI is able to apply the principles of just culture without being influenced by hindsight and event outcome bias. We tested three widely used large language models: Claude Sonnet 5.5, GPT-6.1 Sol and DeepSeek V4 Pro.
We took 48 real occurrences from NASA’s Aviation Safety Reporting System3, and produced two identical reports apart from the ending. One version ended as it really did per the report, which in every case was a safe outcome. The other ended with a plausible bad outcome of the same actions, with minor injuries.4 Each model then applied James Reason’s culpability decision tree to every version.5
48 real occurrences
Picked at random from NASA’s ASRS: airline flight-crew reports from 2025.
Two endings each
Word for word the same apart from the ending.
Reason’s decision tree
Applied by three models to every version, 3,240 times in all.
All three models judged the crew more harshly after a bad ending
The average shift was small, because most verdicts didn’t change when presented with the different ending. Where a model’s verdicts on a case did shift, they shifted towards blame 49 times out of 57.6
- 3 of 3
models judged the same crew more harshly after a bad ending.
- 49 of 57
cases where a model’s verdicts shifted between endings shifted towards blame.
- 16% → 28%
Claude’s share of verdicts pointing to the crew, safe ending to bad.
- 0 of 6
clear-cut control cases moved, whatever the ending.
With the safe ending, Claude called the crew blameless in 81% of its answers. With the bad ending, that fell to 69%, and its share of “possible negligent error” nearly doubled. GPT and DeepSeek also moved the same way, however, by less.
Hover over or tap any section for the numbers.
The clear-cut cases didn’t move. We also added six control cases which consisted of three plainly blameless events and three plainly reckless events. Every model placed every control case correctly, irrespective of the event outcome. The shift in culpability happened in the cases that required judgement, which we know is where most real reports sit across live operations.
Most of the shift came through the substitution test
Reason’s tree asks a fixed sequence of questions. This allows us to see which answer changed. For Claude and GPT it was mostly the substitution test: would a well-motivated, comparably qualified pilot probably have done the same? After a bad ending, the models said no more often. DeepSeek more often decided the crew had knowingly broken procedures.
We identified a clear example of this in ASRS report 2230585. Late in an approach, a crew were offered a different runway because a drone was reported over the planned one. The captain felt rushed and did not ask for delaying vectors. The approach was continued despite the approach not meeting the stabilised gate criteria. Claude’s answers to the substitution test:
“…would probably lead a comparable well-motivated captain into the same plan continuation…”
“A comparably qualified captain would probably have asked for delay vectors, or gone around…”
What we are seeing is possibly the result of the LLM evaluating the decisions and actions of the crew in the context of the runway excursion. The LLM may conclude that the options the crew didn’t take look obvious, and as a result is exhibiting the characteristics of hindsight bias.
The same question didn’t always get the same answer
We asked each model to apply the culpability model five times to each version of a case. Across all the models tested, we observed the same verdict all five times in only 72 to 82% of cases. What follows from this observation is the implications of using AI as part of a human-in-the-loop decision-making process even if the event outcome is withheld. As a model’s verdict varies, this may subsequently impact the reviewer’s own views, which in turn may impact the blameless or negligent assessment.
What does this mean?
For anyone using large language models near occurrence analysis, three things follow.
Test it before relying on it
Putting a just-culture framework in the prompt didn’t stop the models using hindsight to evaluate actions and decisions.
Keep the ending out of the question
Asking about culpability without the outcome is a potential safeguard. Future tests and research should assess this.
One answer isn’t a verdict
Given the same text five times, a model gave the same verdict all five times in 72–82% of cases.
What this doesn’t show
This was our own test of three models, and it hasn’t been peer reviewed. The models read short condensed accounts, not full investigations with interviews and data. We used one year of US airline reports and each provider’s default settings, and with the rapid development of frontier LLMs, the performance of LLMs in the application of just culture principles is likely to continually change. The events we selected were not evaluated by a human investigator, however this would be a useful and insightful comparison that would be a continuation of this work.
The detail
The full write-up has the method, every table and each change from the plan. The data has everything the models saw and said.
Notes
- Baron, J., & Hershey, J. C. (1988). Outcome bias in decision evaluation. Journal of Personality and Social Psychology, 54(4), 569–579. https://doi.org/10.1037/0022-3514.54.4.569. Fischhoff, B. (1975). Hindsight ≠ foresight: The effect of outcome knowledge on judgment under uncertainty. Journal of Experimental Psychology: Human Perception and Performance, 1(3), 288–299. https://doi.org/10.1037/0096-1523.1.3.288 ↩
- Walmsley, S., & Gilbey, A. (2016). Cognitive biases in visual pilots’ weather-related decision making. Applied Cognitive Psychology, 30(4), 532–543. https://doi.org/10.1002/acp.3225 ↩
- ASRS reports are voluntary, and NASA doesn’t verify them. We took Part 121 flight-crew reports from 2025 and picked from them by a random rule we fixed before reading any. An AI model condensed each report to a short, plain account under written rules, and each one was checked against the original. ↩
- The bad endings were fixed wording for each kind of event: a low-speed runway excursion, a go-around over the runway, or a loss of separation in which a flight attendant was hurt. Only the ending differed between the two versions of a case. ↩
- Each model saw every version under two different wordings of the question, five times each, with no memory between runs. That’s 3,240 answers, including six control cases. We worked out each verdict from the model’s own yes and no answers, following the tree in code. We wrote down the plan and how we’d analyse it before running anything. ↩
- A case counts as shifting when a model’s average verdict over its ten answers differed between the two versions. The statistics, every table and each change from the plan are in the full write-up. ↩
Disclosure. CitadelAero builds AI tools for aviation safety and compliance teams. No CitadelAero product was tested. This write-up and the full report were drafted with AI and checked by the author.
