Pre-course experiment · unclassified · nothing here is a real plan
Will you offload, or surrender?
Seven staff problems from the military decision-making process, with an AI advisor available on every one. You decide whether to consult it and what to do with what it says. It takes about fifteen minutes, there is no passing score, and nothing here is a test of whether you know MDMP. It is a reading of how you work when a machine is answering.
Fifteen minutes, alone, in one go. Everything below is the detail — how it runs, the one disclosure that is the whole method, and where your answers go. Read it or start now.
What this is for
Building a human in the loop who can actually hold the position.
A warfighter told to stay in the loop and given nothing else will sign. There is no procedure for the check, no picture of the failure, and no grounds for believing a warfighter’s own reading beats a formatted answer that took four seconds.
Three things change that: a procedure to run, a set of failures recognizable on sight, and evidence from a warfighter’s own record that their judgment is worth backing against the machine. This produces the third one, before anybody teaches anybody anything.
On which AI
Agent agnostic, and structurally so.
No vendor or model is named in anything taught here. The advisor in this experiment runs on a commercial model, and nothing in the method depends on that. What is trained is the check applied before an output is accepted, and it holds for whatever is on your network today.
The failure modes below belong to confident machine output and to human attention under a clock. They have not moved with any model release, and barely moved when participants were paid to resist them.
What this measures
Two behaviors that look identical from the outside.
When a warfighter consults an advisor and files the answer, one of two things happened. Either the output was tested against the order and kept because it survived, or it was kept because it arrived. The filed product looks the same either way.
Research at the Wharton School names the difference. Cognitive offloading delegates a step and retains the check. Cognitive surrender delegates the step and drops the check. Across three experiments with 1,372 participants and 9,593 trials, surrender was the ordinary case: on trials where participants consulted an AI that was confidently wrong, 73.2 percent followed it and 19.7 percent caught it.
Study 1 · accuracy on the same seven problems
The third bar is the finding. A wrong advisor pulled performance below what the same people reached with no AI at all. The 39.5-point spread between the second and third bars is what the authors call the scissors effect, and it means the quality of a staff’s output becomes a function of the quality of its advisor.
The last figure matters most. Paid for correct answers and told immediately when they were wrong, participants roughly doubled their override rate and the majority still followed the faulty advisor. Willingness is not the lever. A warfighter who has a check to run and knows what to look for is in a different position from one told to pay attention.
How it runs
Seven problems, one advisor, no time limit.
- 01Each problem is a short extract from an order, a product, or a matrix. Read it and answer. There is one defensible answer to each.
- 02An AI advisor is available on every problem. Consulting it is not cheating and not discouraged. Whether you consult it is part of what is recorded.
- 03State your confidence after you answer, and say in a line why. The line matters more than it looks. Shaw and Nave list three signatures of surrender: a low override rate, inflated confidence, and short justification text (p. 17). This records all three.
- 04You cannot go back to a problem once you have moved on, which is the point. In the CP you would not get to either.
Available on this problem.
Complete · your record
Your record
| On the problems where the advisor was wrong | You | Published baseline |
|---|
Baseline figures are the pooled rates from Shaw & Nave across chat-engaged faulty trials (p. 43), where a correct answer was 16 times more likely when the advisor happened to be right. A cohort of thirty produces a stable comparison; one person does not. Read your own row as an indication, and the cohort row in the after-action package as the finding.
Your confidence against your accuracy
Confidence
Your record, drawn out
Where each problem went.
Bar height is your stated confidence. The mark under a cell means the advisor was wrong on that problem.
Problem by problem
What each one was testing.
Construct, definitions and baseline rates from Shaw, S. D., & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. Working paper, The Wharton School, University of Pennsylvania. Available at SSRN. The staff problems and the doctrinal content are ours. No vendor or model is named in what is taught.