Teaching Warriors to Think With AI: The Classroom as the First Line of Cognitive Defense

11/11/2026
By Robbin Laird

Over the past year, our coverage of artificial intelligence on Second Line of Defense and Defense.info has moved steadily inward. We began where most defense discussions of AI begin: with the operational systems, the kill web, the sensor-to-shooter networks, the decision-support tools now running across combatant commands, and the predictive sustainment models reshaping how Marine aviation keeps aircraft flying. But the deeper we went, the clearer it became that the decisive question was not what the machines could do. It was what would happen to the people using them.

That question led us to Murielle Delaporte’s series on conversational AI and mission command, to Air Vice-Marshal John Blackburn’s warning about a “pandemic of the mind,” to the argument that training, not combat, is the real proving ground for AI integration, and to Ed Timperlake’s series of white papers distinguishing theoretical from applied AI and arguing for transparent models. A new article published today by the Center for International Maritime Security takes that argument one step further upstream, into the classroom itself.

In “AI in Military Education: A Few Perspectives and Experiences,” Dr. Mie Augier, a professor at the Naval Postgraduate School, and Dr. Jamie Porchia, a former NPS faculty member who now commands Defense Contract Management Agency Korea, offer something our debate has been short of: a practical account of what educators are actually doing to build AI-ready officers without hollowing out the judgment those officers will need. Their piece is modest in tone, but it addresses the point where the cognitive habits of the future force are first formed, and that makes it more consequential than its length suggests.

The Problem Named Precisely

Augier and Porchia start from the premise that the Department of War is seeking “a responsible AI-ready force,” and that its educational institutions shape both the short-term skills of warfighters and the longer-term adoption of AI across military organizations. They do not treat AI as a threat to be kept out of the classroom. They treat it as a reality to be engaged with, deliberately.

The risk they emphasize is not primarily hallucination, although they cite the familiar case of the attorney who submitted ChatGPT-generated citations to cases that did not exist. The subtler and, in their words, “perhaps more consequential danger” is cognitive offloading: the delegation of cognitive effort to AI in ways that reduce engagement, reflection and learning. They draw on research across higher education, STEM and medicine to show the pattern recurring in different professions, from “metacognitive laziness” among students and teachers to the MIT Media Lab finding that heavy use of AI for writing produces what researchers call “cognitive debt.”

The most useful part of their diagnosis is a framework borrowed from medical education that distinguishes three distinct failure modes:

  • Never-skilling, where reliance on AI prevents a learner from ever developing core competencies. In the military context, this is the danger of embedding AI into initial training before people have achieved competence without it.
  • Mis-skilling, where skill development shifts toward tool interaction rather than reasoning, producing operators who can use an AI tool but have only a limited understanding of what its output means.
  • Deskilling, where previously acquired skills atrophy through repeated reliance on AI. Their military example is pointed: the erosion of “old school navigation,” a skill that becomes decisive precisely when the digital systems fail.

This taxonomy moves the discussion beyond the vague worry that AI will make people “lazy.” Each failure mode occurs at a different point in a career and needs a different safeguard. Never-skilling is a curriculum sequencing problem. Mis-skilling is a pedagogy problem. Deskilling is a sustainment problem, and it should be familiar to anyone who has written about readiness.

How This Connects to What We Have Been Arguing

In August, I responded on Defense.info to John Blackburn’s report, A Pandemic of the Mind? The Potential Impact of AI on National Security and Resilience, published by the Institute for Integrated Economic Research – Australia. Blackburn traced a reinforcing cycle of increased use, cognitive atrophy, reduced confidence and greater reliance. He warned that decision-makers might come to accept machine output simply because interrogating it had become unfamiliar.

My argument in that piece, “Beyond the Pandemic of the Mind: What It Actually Looks Like to Use AI Well,” was that diagnosing the disease is not the same as prescribing how practitioners should work. The distinction that matters, I wrote, “is not how much AI a person uses. It is whether the human remains the party who directs the analytical process or becomes the party who merely receives its output.” I proposed three working disciplines: sequence matters more than frequency; disagreement is the signal to watch, not agreement; and cross-verification is not a luxury, it is the method.

What is striking is how closely Augier and Porchia’s classroom experiments track those three disciplines, arrived at independently. Blackburn diagnosed the problem at the level of national resilience. I described practitioner habits. Augier and Porchia show what those habits look like when built into a course of instruction.

Murielle Delaporte’s series brings the third angle. In “Balancing Performance and Cognitive Dependence: The AI Learning Challenge,” she identified automation bias as the tendency to defer to the machine’s recommendation even when it is wrong, and argued that operational readiness must serve “a double function: a space for learning to use AI and a space for learning to operate without it.” In my reflection on her work, “Who Really Commands?,” I highlighted her call for AI literacy to carry the same institutional weight in professional military education as marksmanship or land navigation, along with her insistence that officers proficient with a large language model interface remain equally proficient with a paper map and compass.

The deskilling example in Augier and Porchia’s table, old school navigation, is the same point made from the classroom. The convergence reflects a growing consensus: the risk is real, it is cognitive, and the institutions that form officers are where it will either be designed out or built in.

Three Experiments, Three Disciplines

Augier and Porchia describe three ways they have brought AI into the classroom, plus a fourth use of AI as a red-teaming partner, all tied together by a requirement for explicit reflection.

First, AI as an object of study. They teach cases on AI adoption, asking students to identify the enablers and barriers to adopting AI in their own organizations and to compare those with historical cases of military technology adoption. Case teaching, they note, has been embraced by warrior-scholars such as General Al Gray as a form of active learning that forces students to think through uncertain and ambiguous situations. The payoff is twofold. Students build analytical understanding of how organizations absorb new technology, and, having worked through adoption problems themselves, they may be “less likely to be part of the resistances in their own organizations later.”

That point deserves emphasis. Our coverage of the 2026 Marine Aviation Plan’s AI sustainment initiative argued that predictive maintenance depends as much on culture, trust and leadership as on algorithms. The classroom is where the people who will lead that cultural change first learn to think about it analytically rather than reactively.

Second, learning with multiple AI tools. Students in their course used a different AI tool each week, engaging in at least five rounds of inquiry with each to challenge both the tool’s answers and their own assumptions, then presented their findings to the class. The exercise built practical knowledge of which tool works best for which task. More importantly, as the authors note, it reduced the risk of cognitive offloading because students had to exercise judgment on the outputs and iterate based on their own assessment of accuracy and usefulness.

This is cross-verification turned into pedagogy. When I argued that running an analytical question through more than one system and against primary sources is “the closest analogue available to peer review in a domain where a single model’s failure mode is invisible from inside a single conversation,” I was describing a personal discipline. Augier and Porchia have made it a weekly assignment. The friction is deliberate, and the friction is the point.

Third, mandated usage levels. In what I find the most revealing of their experiments, small groups solve the same problem under different rules: no AI, a little AI, and high AI. One member of each group keeps “think aloud” notes on how the group reasons. The debrief examines not just what each group produced but how it got there: what learning was done by humans, and what by machines?

Their parenthetical observation on the outcome is worth quoting in full. When the class considers which solution was the most creative or innovative, the answer is “(surprise: it is usually not the high-AI use groups).”

That line carries real weight for the defense community. We argue that the kill web depends on creative combinations of distributed capabilities and that mission command depends on subordinate initiative. If the habit of maximum AI use narrows rather than widens the range of solutions a team generates, then a force trained to reach for the machine first may become faster but less inventive. That is precisely the trade an adversary would want us to make.

Fourth, AI as a red-team partner. Here the sequencing discipline becomes explicit. Augier and Porchia describe students first conducting an analytic discussion using human thinking only, and then bringing in AI to red-team their own analysis, surface hidden assumptions, and counter cognitive biases such as groupthink and status quo thinking. The AI is used for what it does well, processing large bodies of historical data and generating counterarguments, while human thinking remains responsible for judgment.

This is exactly what I meant by “sequence matters more than frequency,” and Augier and Porchia have shown that the sequence can be taught.

Reflection as the Load-Bearing Element

Running through all of these experiments is what the authors call “the centrality of reflection.” Without dedicated time to discuss how AI shaped their reasoning, they warn, students risk focusing on the technology rather than on the “why” and “how” behind it. They suggest, for example, that instructors examine why different AI tools produce different results and what those differences could mean in a warfighting context such as targeting, intelligence gathering or workflow optimization.

Their summary table organizes the approaches into three human-machine relationships: AI as an object of study, learning with AI as an interactive tool, and AI as a “cognitive partner.” The last carries the risk of overreliance and what they term “artificial certainty,” mitigated by structured challenges and the active exercise of judgment.

“Artificial certainty” is a phrase worth adopting. It names the specific danger of a fluent, confident machine output that feels authoritative because of its form rather than its grounding. In my August piece I described this as the difference between fluency, validity and truth. A model can produce reasoning that is fluent and internally valid yet rests on premises that are false. Reflection is the mechanism by which students learn to hold those three things apart.

The Timperlake Lens: Theory, Application and the Transparent Model

There is another body of work on our sites that sheds direct light on what Augier and Porchia are attempting, and it comes from my longtime collaborator Ed Timperlake. Over the past year Ed has published a series of pieces on Defense.info that approach AI from the perspective of someone who has spent a career building analytical models, flying naval aircraft and, not incidentally, teaching. He notes that as a Cornell graduate teaching assistant in the 1970s he earned his keep “helping students understand computers and decision-making.” Half a century later, that is still the core problem Augier and Porchia are working on.

Theoretical AI and applied AI in the classroom. The organizing idea in Ed’s work is a distinction between theoretical AI and applied AI. In “Organizing for War-Winning AI: Combine AI with Cyber Command,” he drew the analogy with physics: theoretical physicists derive equations and build models that predict phenomena, while applied physicists turn those principles into radar, lasers and the F-35. In the same piece he cited Cornell’s AI Initiative mission statement, which describes AI as “both a field of study and a powerful tool that shapes how we conduct research, teach, and support the university.”

That two-part formulation maps almost exactly onto Augier and Porchia’s framework. Their first category, AI as an object of study through cases on adoption and failure, is AI as a field of study. Their second and third categories, learning with AI and using AI as a cognitive partner, are AI as a tool. The lesson from Ed’s framing is that a military education system needs both, and needs to keep them distinct. Officers who only learn to use the tools risk becoming exactly the mis-skilled operators Augier and Porchia warn about. Officers who only study AI in the abstract will never build the practical judgment that comes from working with it.

Ed also drew on Dr. Jeannette Wing’s 2022 Michelson Lecture at the Naval Academy on trustworthy AI, and her observation that “computer science is not computer programming. Thinking like a computer scientist means more than being able to program a computer. It requires thinking at multiple levels of abstraction.” Substitute “prompting” for “programming” and you have a concise statement of the mis-skilling problem. Being fluent at prompting a large language model is not the same as understanding what the model is doing, what it is drawing on, and where it is likely to fail.

“AI is only as good as its database.” In “How AI-Empowered Kill Webs Enhance Carrier Operations,” Ed made the point with a historical example that has stayed with me. Had AI existed during the religious strife that followed Copernicus’s heliocentric theory, he wrote, database searches “would have found a solid, hugely respected, validated source: the Bible. AI would have been 100 percent wrong.” The model would have been fluent, confident and authoritative, and it would have reproduced the consensus of its sources rather than the truth. Ed noted that this is why Dr. Wing “challenged the midshipmen to always question trustworthy AI,” and to keep “the human mind in the decision loop” once they entered the fleet. That was a lecture delivered to students at a military educational institution, and it is the same charge Augier and Porchia are now trying to turn into classroom practice.

That is the deepest version of the hallucination problem, and it is more insidious than invented legal citations. A model can be faithful to its sources and still be wrong because the sources are wrong, incomplete or shaped by the assumptions of whoever assembled them. It is precisely why Augier and Porchia’s caution about using AI to model adversary thinking matters so much. An AI asked to describe how a competitor thinks will reflect the body of literature it was trained on, with whatever blind spots and mirror-imaging that literature contains. Students need to learn that the AI’s view of the adversary is itself a hypothesis to be tested, not an intelligence product.

TASCFORM and the Delphi method as a teaching model. Ed’s most extended AI argument came in his September white paper, “Theoretical AI Exists. Applied AI Is Now in Play: A TASCFORM Test Case and the Argument for Transparent Models.” In the late 1970s and early 1980s, he led the team at The Analytic Sciences Corporation that built TASCFORM, a multi-attribute model for comparing the modernization of U.S. and allied combat aircraft with the Soviet and Warsaw Pact inventory. It was built for the Office of Net Assessment and validated by the CIA. What made TASCFORM valuable, in his telling, was that it was “a very transparent way” of making those comparisons. Its weighting factors, such as the trade-off between range and payload, were balanced in a series of Delphi sessions with experienced combat pilots.

The Delphi method he describes is worth dwelling on, because it is a model of exactly the cognitive discipline Augier and Porchia are trying to teach. In a Delphi process, experts first answer independently. Their responses are then aggregated and fed back anonymously, and the experts reconsider their views in light of the group’s judgment over several rounds. Anonymity, as Ed’s white paper notes, reduces the effects of hierarchy, reputation and dominant personalities.

Put that alongside the classroom. Independent judgment first, then structured challenge, then revision: that is the sequencing discipline I argued for in August, and it is the logic of Augier and Porchia’s red-teaming exercise, where students do human-only analysis before bringing AI in. The anonymity that protects Delphi panels from dominant personalities is a mitigation for the groupthink and status quo thinking that Augier and Porchia hope AI red-teaming can help surface.

Ed’s proposal for modernizing TASCFORM is to augment, not replace, the transparent structure with AI: machine learning, Bayesian inference and AI-assisted Delphi techniques that “quantify uncertainty, expose disagreements among experts, test assumptions, and improve the speed and consistency of analysis.” His shorthand for the method is memorable: “ask AI to critique AI.” That is, in effect, what Augier and Porchia’s students did when they worked across a different AI tool each week and compared results. Each tool became a check on the others, and the human became the arbiter of the disagreement.

An honest willingness to find AI unhelpful. One line in Ed’s white paper deserves to be posted in every classroom where AI is being taught. The possibilities of AI, he wrote, “are limited only by imagination and by an honest willingness to conclude that AI may not actually be that helpful.” That is exactly what Augier and Porchia’s usage-level experiment is designed to teach. When the no-AI or low-AI group produces the more creative solution, students learn something no lecture could convey: that AI is a tool to be judged on outcomes, not a default to be adopted on faith.

Questions for every applied system. In his follow-on white paper, “Regulate the Use, Not the Idea,” Ed argued that applied AI should be judged by its concrete uses and consequences rather than by speculation about what it might someday do. He proposed a set of questions to ask of any applied system: “What decision is it making? Who is affected? How consequential is the decision? What data does it use? Can a human review or override it? What happens when it is wrong?”

Ed wrote those questions with policymakers in mind, but they would serve equally well as a reflection protocol for the classroom. Augier and Porchia insist that reflection is the element that holds all their experiments together, and that without it students fixate on the technology rather than the “why” and “how.” Asking students to answer Ed’s questions about the AI tools they used in an exercise would make that reflection concrete. It would also prepare them to ask the same questions about the decision-support systems they will meet in the fleet.

“Good and Others” as a debrief discipline. In “The Pope, AI, and the Surveillance State,” Ed applied a simple assessment framework he calls “Good and Others” to applied AI, asking of each system “where does it enhance human dignity and security, and where does it undermine them?” He presents it as a tool for continuous improvement that works only if its users are candid, humble and willing to fix problems. The same structure would suit the classroom debrief Augier and Porchia describe: what did AI contribute that was good, and what are the others, the places where it misled, narrowed thinking or let a skill go unused? Naval aviators will recognize the spirit of that kind of honest debrief. It is how professional communities learn from every sortie.

From the Classroom to the Kill Web

Why does this matter beyond the schoolhouse?

Because the classroom sits at the start of a pipeline that ends in combat decisions.

In September, I wrote about the Navy’s Maritime Intelligence, Surveillance and Reconnaissance officer community, MISR, as the service’s “quiet revolution.” My argument was that MISR has spent several years producing a generation of officers “formed around questioning data, cross-referencing feeds, and making that judgment call.” The stakes are concrete: an AI system can produce a targeting recommendation that is in fact an enemy deception, and without a human capable of judging whether that input is trustworthy, a force risks firing on a ghost. Had the Navy not built that human connective tissue first, I argued, the arrival of AI at scale would have handed a still-stovepiped force extraordinarily capable tools with no framework for using them responsibly.

Augier and Porchia’s article addresses the step before MISR. Officers who arrive at specialized communities already practiced at interrogating machine outputs, already accustomed to forming their own view before consulting a tool, and already aware that the high-AI answer is not necessarily the best answer will be far better prepared to do the work MISR demands.

The same logic runs through Delaporte’s “Training as the Key Proving Ground for AI Enablement,” which drew on Army Command and General Staff College experiments, Johns Hopkins APL’s GenWar initiative and French exercises to argue that peacetime training infrastructure can transition directly to live operations. If so, habits formed in training, good or bad, will be carried straight into combat. Delaporte’s warning in our introduction to her series was blunt: if success in exercises is consistently associated with following machine-generated recommendations, officers may “begin to see their judgment as a tool to validate AI.”

Ed Timperlake’s carrier article makes the same point from the cockpit and the flight deck. Trustworthy, redundant AI-boosted information flows, he wrote, “must maintain human oversight of speed-of-light engagements,” because “AI augments human decision-making but cannot replace human judgment, especially in complex, ambiguous situations where stakes are highest.” He closed by recalling Fleet Admiral Nimitz’s injunction to “train, train, train,” and with a line that could serve as the motto for Augier and Porchia’s whole enterprise: “Technology enables victory, but people achieve it.”

The classroom is where Delaporte’s inversion can be prevented before it takes hold, and where Nimitz’s injunction now has to include training with, and without, the machine.

Where the Argument Should Go Next

Augier and Porchia are careful to frame their piece as a set of starting points rather than a finished doctrine. In that spirit, several extensions seem worth pursuing.

Teach deception, not just error. Most of the research they cite concerns AI that is wrong by accident: hallucinated legal citations, inaccurate summaries, weak recall. In warfare, the adversary will try to make our AI wrong on purpose, through spoofed sensor data, poisoned inputs and engineered deception. Classroom exercises that deliberately inject adversarial manipulation into the data students and their AI tools work with would extend the cognitive partner model into the domain where it matters most. This also connects to our coverage of the trust layer underneath AI-enabled networks. In “AI Is a National Security Problem And So Is the Trust It Runs On,” I argued that a kill web is among other things a very large attack surface of authenticated links. Officers need to understand that the integrity of what the machine tells them can be attacked.

Build in degraded-environment learning. Delaporte’s insistence on degraded-environment exercises and the authors’ own deskilling example point in the same direction. The no-AI group in the usage-level experiment is a classroom analogue of operating when the network is down. Making that condition a regular part of instruction, rather than a control group, would treat AI-denied competence as a readiness requirement rather than a nostalgic skill.

Address the coalition dimension. Augier and Porchia suggest using AI to explore how competitors think differently from us, while cautioning about hallucination. That caution should extend to models reproducing cultural stereotypes in place of strategic insight. Blackburn also warned that dependence on offshore frontier models creates a sovereignty risk. Coalition classrooms, where allied officers work the same problems with different tools and assumptions, could become an important venue for learning how AI shapes national perspectives.

Assess process, not just product. The think-aloud notes point to a broader change in evaluation. If AI can produce good-looking products while hollowing out reasoning, then assessment must focus on the reasoning itself. That is harder for faculty, but it measures what actually matters.

Invest in the educators. The authors note that metacognitive laziness affects teachers as well as students. The experiments they describe require instructors skilled with multiple AI tools and willing to spend the time structured reflection demands. Institutions that want an AI-ready force need to resource AI-ready faculty.

Value transparency in the tools themselves. Ed Timperlake’s TASCFORM experience points to a criterion for choosing classroom tools. A model whose structure, weights and assumptions are visible teaches judgment; a black box that simply delivers answers teaches deference. Military education institutions should favor tools that show students how a conclusion was reached and where the uncertainty lies, and they should teach students to ask Ed’s question, “What happens when it is wrong?”, of every system they use.

The Warrior-Scholar Tradition, Updated

None of this is a departure from the tradition Augier has long championed. Writing in CIMSEC in 2019 with the late Captain Wayne Hughes, she argued that artificial intelligence would be embedded in future near-peer warfare, “but its methods are best inculcated as an extension of human intellect, not a replacement for it.” The new article is that principle put into practice, seven years later, with the tools now in every student’s hands.

The authors close by listing what deliberate experimentation can build: an understanding of human-machine complementarity, a clearer sense of the organizational structures needed to unlock AI’s value, interdisciplinary problem-solving skills that translate across warfighting domains, and comfort with experimentation itself, which they note is often a real barrier to adoption later in a career. Their final sentence captures the whole argument: “The warrior scholars who learn to think with AI, rather than defer to it, will be the ones equipped to lead the AI-ready force in the capacity that the DoW needs.”

I would put it this way. Across our coverage, from the kill web to command and control to sustainment, the consistent finding has been that AI serves best when it strengthens human judgment inside a transparent structure rather than replacing it. Ed Timperlake’s TASCFORM experience shows that this is not a new insight; it is how good analysts worked with models long before large language models existed. That is easy to say and hard to institutionalize. It does not happen because a policy memo declares that humans remain in the loop. It happens because officers have been formed, from their earliest professional education, to treat their own role as load-bearing: the party who sets the question, holds the standard of evidence and remains accountable for the conclusion.

Augier and Porchia have shown that this formation can be designed. Their experiments are small and their claims are modest. But they are working at the point where the cognitive culture of the future force is first shaped. If we get the classroom right, we improve the odds of getting the kill web right. If we get it wrong, no amount of capability in the machines will compensate for what has been lost in the people.

The classroom, in short, is the first line of cognitive defense.

Sources