The Predictive Soldier: How the Brain’s Internal Models Shape Threat Detection—and Why That Matters for Cognitive Security
Here’s something that might surprise you: two soldiers can walk down the same street, in the same city, under the same conditions, and literally see different things. Not because one is more alert or better trained, but because their brains are running different prediction engines. And those differences — where they come from, what they mean, and how they can be exploited — should concern anyone working in information security.
I want to walk you through a line of research that started with people drawing kitchens and ended up raising some uncomfortable questions about cognitive vulnerability in operational environments.
Your Brain Doesn’t Just See — It Predicts
Most of us think of vision as a camera. Light comes in, the brain processes it, and we see what’s out there. That model is wrong. Or at least, it’s badly incomplete.
What the brain actually does is run a continuous prediction loop. Before your eyes even land on something, your brain has already generated an expectation about what should be there, based on a lifetime of accumulated experience. These expectations are what neuroscientists call internal models — your brain’s best guess about how the world is structured. When the incoming sensory data matches the prediction, processing is efficient and fast. When it doesn’t match, you get a prediction error — a neural signal that says “something here needs attention.”
This framework, broadly called predictive processing, has been gaining traction in cognitive neuroscience for over a decade. But a recent study brought it into sharper focus by asking a deceptively simple question: do individual differences in internal models explain why people look at the same scene differently?
Drawing Kitchens, Tracking Eyes
A research team led by Micha Engeser at Justus Liebig University Giessen tackled this question with an elegant experimental design (Engeser, Babaei, & Kaiser, 2026). They asked participants to draw their most typical version of a kitchen and a bathroom — not from a photograph, but from memory. These drawings served as a window into each person’s internal model. What does your brain think a kitchen looks like?
The drawings varied considerably across participants. Some people put the stove front and center. Others emphasized counter space or a window over the sink. These aren’t random differences — they reflect years of living in specific environments, cooking in specific kitchens, building up a personal template of “kitchen-ness.”
The researchers then converted these drawings into photorealistic images using AI tools and fed them through a deep neural network to extract high-level feature representations. This gave them a quantitative way to measure how similar any two participants’ internal models were.
Next came the eye tracking. The same participants viewed 300 photographs of real kitchens and bathrooms while the researchers recorded exactly where they looked, how often they moved their eyes, and in what order they inspected different objects.
The critical question: do people with more similar internal models explore scenes in more similar ways?
When Predictions Matter Most
The answer turned out to depend on the conditions. In a standard free-viewing experiment — just look at the pictures however you want — internal models didn’t predict gaze behavior. People looked at scenes idiosyncratically, but those idiosyncrasies weren’t explained by their drawings.
But the researchers ran a second experiment with a twist. They introduced gaze-contingent viewing, where only a small window around the current fixation point was shown in detail. Everything in the periphery was blurred and desaturated. Participants were also told to memorize the scenes for a later test.
Under these constrained conditions, internal models suddenly mattered. Participants with more similar drawings explored scenes with similar fixation patterns — similar numbers of fixations per image and similar ordering of which objects they looked at first.
This finding is important, and here’s why. The gaze-contingent condition forced participants to actively sample the environment rather than passively absorbing it. They couldn’t just let bottom-up salience guide their eyes. They had to make decisions about where to look next based on limited information. And when that happened, their brain’s internal predictions took the wheel.
Think about what that means in operational terms. A soldier clearing a building isn’t casually browsing a well-lit scene. They’re moving through environments with restricted sightlines, degraded lighting, time pressure, and incomplete information. They are, in effect, living in a permanent gaze-contingent condition. Their internal models aren’t a background process — they’re the primary navigation system.
The Tactical Internal Model
Now let’s extend this logic. If people develop idiosyncratic internal models for kitchens based on their personal experience with kitchens, what kind of internal models do soldiers develop for operational environments?
A soldier who has done multiple deployments in urban settings has built up a detailed predictive model of what those environments look like and how threats are distributed within them. Windows and rooftops carry a different predictive weight for someone who has experienced sniper fire than for someone who has not. Roadside debris gets coded differently after IED exposure. Crowd density patterns get parsed through a lens shaped by specific operational histories.
These internal models are a feature, not a bug. They make experienced soldiers faster and more efficient at detecting threats because their brains are generating better predictions about where danger is likely to be. The prediction errors — the mismatches between expectation and reality — become more diagnostic. An experienced operator doesn’t just notice something unusual; their prediction error is calibrated to flag tactically unusual things.
This is part of what we mean when we talk about expertise. It’s not just procedural knowledge or reaction time. It’s a finely tuned generative model of the operational world, built through experience and refined through feedback.
The Vulnerability No One Is Measuring
And here’s where this gets uncomfortable for information security professionals.
If internal models are the primary driver of threat detection under operational conditions — and the Giessen research suggests they are — then corrupting those models is one of the most effective ways to degrade a soldier’s performance. And no one is systematically measuring or protecting these models.
Consider what an adversary could do. Doctored intelligence imagery that subtly shifts a soldier’s expectations about where threats appear in a particular environment. Manipulated drone feeds that train analysts to associate threat signatures with the wrong contextual cues. Physical staging of operational environments to create systematic prediction errors — planting decoy indicators in locations where soldiers have learned to expect threats, while placing real threats in locations their internal models treat as low-priority.
None of these attacks target equipment. None of them exploit network vulnerabilities. They target the prediction engine running inside the soldier’s skull. And because internal models are built gradually through experience, the corruption can be subtle, cumulative, and extremely difficult to detect through conventional assessment.
A soldier whose internal model has been gradually shifted might perform normally on standard evaluations. Their reaction times might be fine. Their procedural knowledge might be intact. But their predictions about where to look and what to prioritize would be miscalibrated in ways that only become apparent under operational stress — exactly the conditions where those predictions matter most.
Measuring What We Can’t See
This brings us to what I think is the most actionable takeaway from this line of research: we need methods to assess and monitor the integrity of soldiers’ internal models.
The Giessen team’s drawing-based approach is a starting point. Asking a soldier to draw their most typical version of an operational environment gives you a crude but useful snapshot of their internal model. Comparing those drawings across experience levels, deployment histories, and over time could reveal systematic shifts that flag potential model corruption or maladaptive recalibration.
But drawings are coarse. The real potential lies in combining behavioral measures with neuroimaging — specifically, EEG. Event-related potentials time-locked to fixations during scene exploration can reveal the magnitude and timing of prediction error signals. If a soldier’s neural response to threat-relevant objects changes over time — smaller prediction errors where there should be large ones, or misplaced prediction errors in the wrong spatial locations — that’s a measurable signature of internal model drift.
Imagine a periodic assessment protocol: a soldier views a set of operationally relevant scenes under gaze-contingent conditions while EEG records their neural responses. Their fixation patterns and prediction error signatures are compared against their own baseline and against the patterns of experienced operators in similar roles. Deviations from baseline could flag individuals whose internal models have shifted in potentially dangerous ways — whether through adversarial manipulation, trauma-related recalibration, or simply degraded training.
The PTSD Connection
This framework also offers a new lens on post-traumatic stress. PTSD can be understood, at least partly, as a failure of internal model contextualization. The soldier’s prediction engine, tuned for operational environments, doesn’t switch off when they return to civilian settings. Everyday scenes get parsed through threat-calibrated internal models. The rustle of a plastic bag generates a prediction error calibrated for IED detection, not grocery shopping.
This isn’t a metaphor. If the Giessen framework holds, you should be able to measure this directly. A soldier with PTSD should show operationally-tuned fixation patterns and prediction error signatures when viewing civilian scenes — their internal model is leaking from one context into another. The magnitude of that leakage could serve as a quantitative marker of severity, and tracking it over time could provide an objective measure of treatment progress.
Current PTSD assessment relies heavily on self-report and clinical observation. A neurobiological measure rooted in predictive processing could complement these approaches, particularly for individuals who underreport symptoms or whose presentation doesn’t fit standard diagnostic categories.
Where This Fits in the Cognitive Security Landscape
I’ve written previously about frameworks for detecting manipulation in information environments — approaches that treat cognitive systems as dynamical systems and look for signatures of destabilization before full-blown disruption occurs. The internal models framework connects to that broader program, but from a different entry point.
Where information-environment analysis focuses on the inputs to cognition — what signals people are exposed to and how those signals propagate — the internal models approach focuses on the architecture that processes those signals. Both matter. An adversary can attack the information stream, or they can attack the prediction engine that interprets the stream. The most sophisticated attacks will do both.
For information professionals, the practical implication is that cognitive security can’t stop at controlling the information environment. We also need to understand and monitor the predictive models that people bring to that environment. Two analysts receiving identical intelligence briefings will process that information differently based on their internal models — and if one analyst’s models have been subtly corrupted through a long-term influence campaign, the consequences won’t show up until a critical decision point.
What Comes Next
This is still early-stage thinking, and I want to be transparent about that. The Giessen study demonstrated the basic science — internal models predict exploration behavior under conditions of uncertainty. The extension to military and cognitive security contexts is a reasoned inference, not established fact. The experimental work applying this framework to operational populations hasn’t been done yet.
But the theoretical foundation is solid, the measurement tools exist, and the stakes are high enough to warrant serious attention. Here’s what I think the research agenda should look like:
First, characterize how internal models differ across military experience levels and specializations using the drawing-based and neuroimaging approaches described above. Second, test whether those differences predict performance on operationally relevant tasks under degraded sensory conditions. Third, develop protocols for monitoring internal model integrity over time. And fourth — the piece that matters most for this audience — investigate how adversarial information operations could target and corrupt these models, and build detection mechanisms before that threat fully materializes.
We spend enormous resources hardening networks and securing communications channels. The prediction engine between a soldier’s ears deserves the same attention.
References
Engeser, M., Babaei, N., & Kaiser, D. (2026). Can individual internal models predict idiosyncratic scene exploration? bioRxiv. https://doi.org/10.64898/2026.04.01.715777
Friston, K., Adams, R., Perrinet, L., & Breakspear, M. (2012). Perceptions as hypotheses: Saccades as experiments. Frontiers in Psychology, 3, 151.
Henderson, J. M. (2017). Gaze control as prediction. Trends in Cognitive Sciences, 21(1), 15–23.
de Lange, F. P., Heilbron, M., & Kok, P. (2018). How do expectations shape perception? Trends in Cognitive Sciences, 22(9), 764–779.
DISCLAIMER: The following article is an original work published by the Information Professionals Association. Opinions expressed by authors are their own, and do not necessarily reflect the views of or endorsement by the Information Professionals Association, the United States Marine Corps, or US Government.