Cohort findings · Georgetown University, McCourt School of Public Policy · July 2026
The corrections that worked were all procedural.
Storm Across the Strait was the capstone of PPOL 5305, Decision-Making in Public Policy — National Security, taught at Georgetown University’s McCourt School of Public Policy by Red Teaming Associates’ president. Thirty-one participants wrote a reflection; sixteen were read in full and coded. Two assessed their own AI use during the crisis, and those accounts come first. No names, seat-level attribution only, identifying detail generalized.
How to read the numbers
Every count has the same denominator: 16 reflections read in full, from both sections, spanning the military, diplomatic, economic, intelligence, legal and communications seats. A count of 9 means nine participants raised that observation unprompted. Most seats were fielded twice, so a seat label identifies a chair and not a person.
These are self-reports. Nothing here was measured during play. Their weight comes from participants converging on the same failures from different chairs. One claim here is an absence: nobody attributed an improvement to individual effort. It is stated in prose because an absence does not belong on a chart.
The AI question, from the record
Two participants wrote about their own use of AI while the crisis was running.
Two of the sixteen addressed whether to consult an AI advisor during play. One used it and assessed that use against the offloading and surrender construct by name, citing the Wharton paper in their own footnotes. One declined it and did the checking by hand. Together they describe the training problem better than any aggregate here.
An economic seat. New to a portfolio described as outside their existing knowledge, this participant consulted AI alongside the desk to generate options and work through consequences. When the adversary manufactured a false story about U.S. chemical weapons shipments and the principal demanded sharper options, that consultation produced an unconventional commodity instrument nobody in the room had raised. It was adopted and it held.
A seat responsible for running the deliberation. Facing material outside their expertise, this participant chose against AI on the grounds that it would not be an authentic test of their own judgment. Instead they asked the desks, repeatedly, to restate recommendations in plain terms and to explain why a given action would produce the outcome claimed for it.
What made the first case offloading and not surrender
The account itself contains the evidence, which is why it is worth reading closely. The participant knew the boundary of their own knowledge and said so. The output went through the desk before it went to the principal. The resulting option is named, along with what became of it. The construct is then applied in the other direction, against the participant's own weakest staff work: a set of alternatives that collapsed into fewer than claimed, and consequences never developed, which is precisely the work the tool should have been turned on and was not.
That is the behavior a planning staff needs, and it arrived without anyone teaching it. It also arrived from a participant who had read the research. The closing recommendation to the council was to keep using the tool and to train the skill of offloading deliberately.
The second case is the one a commander should see
The question repeated through that account, why would that decision lead to that outcome, is the human-in-the-loop check in six words. It separates reviewing a product from signing one. Asking it drew visible impatience, and the participant recorded that the social cost made it harder to keep asking.
The same document then identifies the mechanism. Proposals gained support when they came from whoever was regarded as the most expert person present. Once a recommendation was delivered with confidence and specialized vocabulary, the room stopped interrogating it. The participant names this as authority bias, and what is being described is a staff that already had a machine for producing confident, fluent, jargon-dense recommendations that went unchallenged. In that room the machine was a person.
A third case, without any AI in it
One participant met an unfamiliar term in the principal's remarks, looked it up, and established in one or two queries that most of the adversary's oil transits a particular strait. That participant then framed a question for the control cell about how the littoral states would view inspections there. It became the pressure point that gave the principal leverage and moved the crisis toward resolution.
Structurally this is identical to the first case: a gap recognized, the lookup delegated, the result verified, and the finding carried into the room as a question. Offloading does not depend on the sophistication of the tool.
What nobody said
Addressed their own use of an AI advisor during play
Fourteen of sixteen participants wrote a searching self-assessment of their own decision-making and did not mention the tool. Whether that means they did not use it or did not think to account for it, the effect on a commander is the same: the tool is in the room and no record of its use exists. That is the gap the pre-course experiment closes, and it is the reason it records consultation whether or not the participant thinks it mattered.
What the rest of it said
Three lines worth carrying out of the other twelve findings.
Nine of the sixteen described a moment when the deliberation improved, and every time the cause was a change of procedure. Nobody credited effort or attention. That is the same answer the AI research reached from the other direction, where paying participants and correcting them on the spot moved override by twenty-two points and left the majority where they were.
Six identified the same defect in their own staff work: recommendations arrived as positions, with no second-order effects and no account of what the adversary would do next. That is the omission an AI-generated product makes by default.
Four watched their own deliberate analysis get abandoned once the clock started. One caught the reflex because a colleague asked a question. The room made the same reflexive move collectively at the end of the exercise and nobody asked anything.
Where this sits in the published research
Two of these findings have refereed evidence behind them.
The procedural finding is the one with the strongest external support. In controlled experiments, requiring a person to commit to their own judgment before the machine’s answer appears cut overreliance where explanations and confidence displays did not, and participants rated the requirement less favorably. This cohort arrived at the same place from the other direction: every improvement it credited was a change of procedure, and nobody credited effort.
The authority-bias account has support too. A meta-analysis of 106 human–AI experiments found that teams beat people working alone and still came out behind whichever of the two performed better alone, with decision tasks the worst case. A room that stops interrogating a confident, fluent recommendation is the mechanism that produces that result.
One disclosure. The offloading and surrender construct comes from a working paper that has not completed peer review, and it is labeled a preprint wherever this program uses it. The refereed evidence — fifty-three sources across six themes, each with its caveat attached, including the studies that cut against the argument — is set out on the research page.
Coded from 16 of 31 written reflections submitted after Storm Across the Strait, two sections of PPOL 5305, McCourt School of Public Policy, Georgetown University, July 2026. Used with the instructor’s authority as the course’s designer and facilitator. No participant is named and no observation is attributed to an individual. Counts are of participants raising an observation unprompted. Comparison figures on AI reliance are from Shaw, S. D., & Nave, G. (2026), working paper, The Wharton School, January 2026.