The Diagnosis Is Backwards.

A response to Norman MacLeod’s “Treating Pilots Like Cattle” (LinkedIn, 18 August 2026)

https://www.linkedin.com/pulse/treating-pilots-like-cattle-norman-macleod-1z1de

Cedric Paillard — Chief Executive Officer, Amris Aviation

Norman MacLeod’s piece posted 18 August is the most useful thing written about EBT in a while, and I want to start by agreeing with the part most people will skip past.

His prediction is that EBT becomes EBTA — Evidence Based Training Audit. Regulators who don’t have the capacity or the domain knowledge to oversee competency-based frameworks will do what they did with SMS, FRMS, IOSA and LOSA: push the burden of demonstration onto the operator. The accountable manager then discovers, at the worst possible moment, that they are accountable for something nobody ever built the means to evidence.

That is exactly right, and it is coming faster than the industry thinks.

Where I part company with Mr. MacLeod is on why we got here, and therefore on what to do about it.

Start with the Cows

He opens with the Economist piece on GPS-tracked cattle — pasture nutritionally mapped, each animal’s intake measured, tailored supplements closing the gap so the whole herd reaches market in the same condition. He offers it as a slightly uncomfortable joke about where pilot training might go.

But look at what actually makes that system work. It is not standardisation. It is not treating the herd as a herd. It is the opposite: continuous, cheap, per-individual measurement, which is the only thing that makes individual treatment affordable at scale. The farmer stopped managing an average and started managing each animal, because a sensor made the individual visible.

Which means the metaphor points the other way. The genuinely cattle-like practice in commercial aviation is the one we run today: put every pilot through the same recurrent event, on the same six-month cycle, against the same script, regardless of what that specific pilot’s line flying has or hasn’t exposed them to. That is herd management. Individualisation isn’t the dehumanising thing. Unmeasured uniformity is.

So I would reframe the question. It isn’t “should we treat pilots like cattle?” It’s “why does a dairy farmer have better per-individual data on a cow than a Head of Training has on a Captain?”

The Dark Secret is a Measurement Problem, Not a Concept Problem

Mr. MacLeod’s central charge is that CBTA and EBT have been rolled out at breakneck speed with no evidence they actually work. The facts of that are hard to dispute. The inference I would resist is that this tells us something about EBT.

Consider what we ask an instructor to do. In a two-crew sim detail they are expected to observe, hold, classify and subsequently recall behaviour against nine competencies and several dozen observable behaviours — while simultaneously operating the box, injecting the malfunction, managing the scenario, monitoring safety, and thinking about the debrief.

Here is where the cognitive science becomes unavoidable, so it is worth being precise about it.

Nelson Cowan is a cognitive psychologist at the University of Missouri. In 2001 he published The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity — a deliberate revisiting of George Miller’s famous “seven plus or minus two” from 1956. Cowan’s argument was that when you strip out the tricks people use to cheat the limit — rehearsal, chunking practised material into larger units, long-term memory support — the raw capacity of working memory is closer to four items held simultaneously, not seven.

A “chunk” is the unit that matters. It is not a fact; it is a bundle of information the mind can treat as one thing. “Airspeed, attitude, thrust and configuration” is four separate chunks to a novice and can be one chunk — “energy state” — to an expert. Expertise doesn’t raise the ceiling. It changes what fits inside a slot. That distinction is the whole argument, and it is why “experienced instructors manage fine” is not the rebuttal it first appears to be: an experienced TRE chunks more efficiently, but they are still working with four slots, and they still have to spend some of them on running the exercise.

The number itself is contested, and honestly so. Cowan’s paper was published as a target article with open peer commentary, so the objections are on the record alongside it. A recent review by Doherty and Logie (2024) concludes that whether you land on four, five or seven depends heavily on the experimental paradigm. That debate does not rescue us. In our own analysis of a standard CBTA framework, the observable-behaviour load in the take-off, approach and landing phases runs to around nineteen concurrent chunks. At Cowan’s four, that is roughly five times capacity. At Miller’s seven — the most generous number anyone credible defends — it is still about 2.7 times over. The conclusion survives the entire range of the disagreement, which is the only kind of conclusion worth building on.

So the grades that come out of that process are not fabricated. They are real professional judgments — applied to whatever sample of the session human attention under load was able to retain. Different instructor, different fragment retained, different grade. The variance we all quietly acknowledge in inter-rater reliability isn’t an instructor quality problem. It is the predictable output of using unaided human attention as a primary sensor in a high-rate data environment.

Now aggregate that up. Roll unreliable individual observations into fleet-level competency analytics, feed it into your EBT programme design, and you have a picture whose confidence interval nobody has ever computed — because nobody could.

That is why there is no evidence that EBT works. Not because the framework is wrong. Because we have never possessed an instrument capable of testing it. You cannot validate a training methodology using a measurement system whose own reliability is unknown and probably poor. The absence of evidence here is a fact about the sensor, not about the concept.

Why Design Discipline Alone Won’t Close It

Mr. MacLeod’s prescription is that training design, delivery and management is a discipline like any other — foundational concepts, processes, deliverables, auditable outputs — and that this discipline is missing in commercial aviation.

Necessary, and I would sign it. But not sufficient, and there is a trap inside it worth naming.

Instructional systems rigour produces auditable process. It documents that you analysed, designed, developed, implemented and evaluated. It does not improve the reliability of the observation on which the whole edifice rests. You can run a textbook-perfect ADDIE cycle on top of a measurement layer with poor inter-rater reliability and produce a beautifully documented programme built on noise.

And the auditor he correctly predicts will not, in 2030, ask whether you had a design process. They will ask three questions: (1) what is your evidence that this pilot is competent, (2) how was it captured, and (3) what is the demonstrated reliability of that capture? Process documentation answers the first. It does not touch the other two — and those are the two that create exposure for the accountable manager.

Which Makes the Technology Argument Second-order

Mr. MacLeod is right that distributed simulation, VR and AI crew members are technically achievable now, and that regulation will be the sea anchor. I would add a caution and a correction.

The caution: adding delivery modalities to a system that cannot reliably measure outcomes doesn’t shrink the evidence gap, it widens it. More training environments, more configurations, more variance — all assessed by the same overloaded observer. You end up with a larger and more expensive set of things you can’t evidence.

The correction: the first job of AI in flight training isn’t to be a synthetic first officer. It is to be a working-memory prosthetic for the human instructor — capturing the behavioural stream continuously and completely, indexing it to phase of flight and exercise, and presenting it back so that professional judgment is applied to a full record rather than a remembered fragment. The instructor remains the assessor. That is not negotiable and should not be. The machine takes the part humans are demonstrably bad at — sustained high-rate capture and recall under load — and leaves the part humans are uniquely good at, which is interpretation.

And note where the sea anchor actually sits. The regulatory drag Mr. MacLeod describes is on the exotic delivery technology — networked sims across jurisdictions, AI avatars as crew members. It is not on the evidence layer. An operator that wants to instrument its own observation process, capture behaviour systematically and measure the consistency of its own assessors can begin that work inside the frameworks as they stand today. Which is a good argument for building the evidence layer first: it is the part you are already permitted to do, and it is the part that determines whether any of the rest can ever be validated.

Back to the Herd

Mr. MacLeod closes by saying that before we start treating pilots like cattle, we might reflect on whether the systems we have in place will get them to market in tip-top condition. I agree completely — and I would only sharpen what is missing.

It isn’t a design methodology. It’s an instrument.

Commercial aviation currently measures the competence of the person at the front of a transport-category aircraft using a sensor — unaided human attention, saturated, retrospectively recalled — that no other safety-critical industry would accept as primary. Fix that, and the evidence problem, the validation problem, the audit problem and the accountable manager’s exposure all become tractable at the same time, because they are all the same problem wearing different clothes.

Until then, the uncomfortable truth isn’t that we’re treating pilots like cattle.

It’s that we’re making individual, career-affecting, safety-critical judgments about them on thinner data than a dairy farmer has about a cow.

References

Cowan, N. (2001). The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity. Behavioral and Brain Sciences, 24, 87–185.

Cowan, N. (2010). The Magical Mystery Four: How Is Working Memory Capacity Limited, and Why? Current Directions in Psychological Science, 19(1), 51–57.

Doherty, J. M., & Logie, R. H. (2024). Modelling Working Memory Capacity: Is the Magical Number Four, Seven, or Does it Depend on What You Are Counting? Journal of Cognition.

Miller, G. A. (1956). The Magical Number Seven, Plus or Minus Two. Psychological Review, 63(2), 81–97.

The chunk-load analysis referenced above is set out in full, with the statistical treatment of sampling and confidence intervals, in The Cognitive Limits of CBTA (Amris Aviation, 2026). https://aviationtraining.ai/the-cognitive-limits-of-cbta/