The Worst PowerPoint Slide in the World

An analysis of how the 2009 Afghanistan COIN map is a perfect entry point for a discussion on the limits of human cognition versus the promise of algorithmic synthesis.

DatassFigure 1: PA Consulting Group COIN presentation to ISAF in 2009

The slide

In 2009, the UK firm PA Consulting Group gave a presentation on COunter INsurgency (COIN) strategies to be considered for Afghanistan to General Stanley McChrystal, who was then the Commander of US and ISAF forces in the country. When this slide popped up on the screen, he interrupted the presentation and quipped that “once we understand this slide, we will have won the war.”

Since then, this slide has been regarded as one of the best examples of how NOT to illustrate a point in a presentation.

But here’s my point: I like it. There is no better way to summarise in one diagram the extent of the arguably deluded belief that Afghanistan could be “pacified” under a Western set of ideals. In my mind, it also serves another great purpose: to show how trains of thought are connected, and, in the context of AI, show where the issues emerge in the search for better ways to dissect a problem to find a path to its solution.

This is a loose follow-up to a previous article I wrote about knowledge foundations. Although my primary job is in the aerospace field, my company also works in cybersecurity and surveillance, and I am professionally interested in AI analysis of unstructured data. So, I see this slide, and wonder: would AI have been able to follow all these paths and weigh an optimal “solution” to the problem of suppressing the never-ending insurgency problems that the ISAF was attempting to solve?

One company that is in the news thought about this well before me: Palantir. Karp and his staff looked at this “unstructured complexity,” and it became the catalyst for early iterations of Palantir. The goal wasn’t just to see the lines, but to make the nodes searchable and the connections quantifiable.

Before AI was a thing, Palantir thought about offering a way to structure database complexity across different government databases. Nothing technically wrong with that, especially because it was—in part—database compartmentalisation between government departments (FBI, CIA, FAA, etc.) that caused the dots leading up to 9/11 not to be connected.

Palantir’s early success didn’t actually solve the “Spaghetti Slide” problem—it bypassed it. Palantir realised that trying to map an entire society’s stability mathematically is a fool’s errand. Instead of attempting to calculate systemic social balance, they focused on *unstructured entity extraction*—turning messy data silos into concrete actionable links (e.g., matching a phone number found in Iraq to a bank account in Dubai).

But let’s leave Palantir aside for now and return to our slide.

I uploaded the slide to Gemini and asked it to count the nodes and the connections.

Figure 2: Gemini gives up

It gave up, which is something that I had done as well, trying it by tracing my mouse over this mess of Gordian knots. Most likely because the resolution of the images I found was poor. I am sure that with better image recognition, this will be possible, so let us assume it is. Can we discover anything useful by knowing this? Not really, except for a confirmation of the huge number of actors and stakeholders in the field. Even if we then move on to counting the links between nodes, we gain little insight. This is because of something that is missing in this already complex diagram: the weight (or importance) of those connections (I should note that there is one qualifier for them, which is “delay”; can you spot it in the image?)

So, I am wondering, what might such a structure look like with weights added? Or rather—given the immediate answer would be an even bigger mess — would adding weights and then having the whole thing analysed by an AI agent give any better insight into a solution?

So, this is another article on how to gain knowledge from an unstructured data set, and what might be needed to do so.

The physics of the arrow: Bayesian Inference and Sensitivity Analysis

Assuming that you have uploaded a high-resolution version of this image to an AI agent, how would the agent treat it? I asked Gemini, and here is the answer:

Figure 2: Gemini analysis of one section of the slide.

A lot of words, so let’s see if they make sense. First, what is a Bayesian Network?

To explain this, we need to understand what a Directed Acyclic Graph (DAG) is. A DAG is a structural mathematical framework used to map complex relationships without creating infinite logical loops. In data science, it is composed of “nodes” (which represent variables, events, or data points) and “directed arrows” (which signify the directional flow of cause, influence, or time from one node to another). The defining characteristic of a DAG is that it is “acyclic,” meaning it is physically impossible to start at any single node, follow the sequence of arrows, and loop back to that same starting point. By enforcing this one-way flow, a DAG prevents the system from being trapped in feedback loops, allowing algorithms to cleanly sequence and untangle massive webs of unstructured dependencies into a clear, chronological pipeline of cause and effect.

Figure 4: Representation of a Directed Acyclic Graph (DAG), showing no loopbacks

When you inject probability and statistical uncertainty into this structural skeleton, a DAG evolves into a Bayesian Network. In real-world environments, connections between variables are rarely absolute; instead, they are probabilistic. A Bayesian Network calculates these relationships using Conditional Probability Tables (CPTs) for each node, computing the exact likelihood of an event occurring based on the shifting states of its immediate “parent” nodes. For instance, rather than assuming an action always yields a fixed result, the network continuously updates its beliefs as new field data arrives—calculating how the probability of an outcome changes when multiple background variables fluctuate simultaneously.

Figure 5: Bayesian Networks assign probabilities to the connections between nodes in a DAG

By combining the structural rigidity of the DAG with the fluid mathematics of Bayesian probability, this network transforms a static visual map into a dynamic predictive software engine. It allows data analysts to conduct Sensitivity Analysis, mathematically flexing individual weights within the network to observe how small variations in a single obscure node ripple through the entire system. This makes the Bayesian Network an incredibly powerful tool for diagnosing complex, high-stakes environments where human intuition fails, exposing hidden critical paths where seemingly minor variables hold catastrophic systemic leverage over the outcome. And this is one of the fundamental omissions from the COIN slide.

Figure 6: Final sequence showing how a Bayesian analysis on a DAG can determine the likelihood of events (sensitivity analysis).

How could this be applied to “the slide” in today’s AI capabilities context?

The sensitivity analysis is one of the keys that were missing in the original PA Consulting slide. So, my next question is: if the database of nodes contained enough “relevant” information, could an AI agent assign a sensitivity value to the connectors between them and forecast an outcome?

To answer that question, we first must assume that the slide is a DAG. It appears to be, as there are no recursive loops (although, given its confusion, that is hard to state with certainty, but let’s assume yes). In that case, we move on to a Bayesian analysis and weigh the probabilities of the links (beyond the original “significant delays” qualifier).

The original drafters could have done this, although creating a mathematical model of weights and probabilities was probably beyond the scope of the initial analysis. Even so, with the tools available in 2009, I am not sure that would have even been possible. But could AI do it today? Could AI analyse nodes and create weights and dependencies around them in such a way as to predict outcomes and assign probabilities to them?

In 2009, constructing a dynamic network required painstaking manual entry by human teams to map variables into system-dynamics tools like *Vensim or [Stella](https://www.stella.systems/). Today, Large Language Models (LLMs) combined with [Graph Neural Networks](https://distill.pub/2021/gnn-intro/)* (GNNs) can ingest thousands of unstructured operational reports, historical intelligence logs, and socioeconomic data points simultaneously. The AI performs structured entity extraction—much like the database architecture principles pioneered by companies like Palantir—to translate conversational text into concrete nodes. It then scans these data sets for semantic correlation to construct a functional DAG automatically (the blue box—DAG structure—in Figure 5), capturing the dependencies human planners tried to hand-draw.

Once the graph structure is assembled, the AI shifts from parsing text to calculating probabilities—the key to predicting possible outcomes. It reads the historical data to build the Conditional Probability Tables (CPTs) required by a Bayesian Network (the green box—Bayesian network—in Figure 5). Rather than guessing a static weight for a connector based on subjective expert opinions, the algorithm evaluates how often changes in ‘parent’ nodes historically caused shifts in ‘child’ nodes—a process known in computer science as algorithmic parameter learning from empirical data (Koller & Friedman, 2009). For instance, by processing years of local economic records and regional security metrics, a machine learning agent can compute the exact conditional probability of civilian cooperation fluctuating when external aid drops but localised security increases.

The next step is to determine sensitivity. Modern compute allows the AI to run millions of simultaneous simulations to conduct deep Sensitivity Analysis (the purple and orange boxes in Figure 5). This means the software can mathematically fluctuate individual dependencies across the network to identify which obscure paths exert the most system-wide leverage. Furthermore, an AI agent today can execute Inverse Modelling (back-propagation)—ingesting the known, definitive outcome of the conflict (the 2021 state collapse) and computing backwards through the node architecture to discover the exact path weights that made that outcome mathematically inevitable. That wouldn’t have helped McChrystal at the time, but Inverse Modelling can be used today to value path sensitivities in historically similar scenarios as the one being analysed.

So, to answer the question above: could an AI agent assign a sensitivity value to the connectors between them and forecast an outcome? Yes, modern AI can do this today, but with a critical architectural caveat. If you feed a high-resolution version of this visual map and its underlying intelligence databases into a modern AI agent, it will not see an unreadable “spaghetti” mess. Instead, it uses natural language processing to extract information from the node descriptors and interprets the lines as probabilistic dependencies. By training on historical field logs, the AI can automatically generate Conditional Probability Tables for every single intersection, effectively transforming a static, confusing slide into a live, predictive software model. Well… in theory at least. To do that effectively would require massive, clean, labelled historical data (which Afghanistan notably lacks). But, for what I write, let this slide, please!

However, this reveals the true limitation of the tool. While the AI can mathematically calculate the weights of these connections through Sensitivity Analysis, the machine is still entirely bound by the structural limits of the map it was given. If the original human planners left out invisible variables—such as psychological resilience, cultural pride, or deep-seated historical trauma—the AI will attempt to optimise a flawed network. The algorithm can tell you with high mathematical probability exactly how the system will behave, but it cannot invent a node that the human creators forgot to include. This is what happened when Palantir’s Maven suggested the Minab School in Iran as a valid enemy target. The database the AI examined was flawed, as it was missing the critical piece of information that the location had been converted to a school from a previous IRGC camp.

What happened in Minab exposes the precise boundary where predictive optimisation ends and the severe danger of automated systems begins, revealing a profound lack of what cognitive scientists call metacognition: the ability to think about one’s own thinking. Where a human analyst might pause to doubt the completeness of their information, a Bayesian network operates under a permanent delusion of total clarity. So, let’s look at what metacognition means in human and AI terms.

The metacognitive deficit: Why algorithms can’t question the map

Metacognition is the uniquely human ability to stand outside our own cognitive architecture and critique it. It is the voice inside your head that asks, “Am I framing this problem entirely wrong?” or “What critical variables am I blind to right now?” An AI agent processing a Bayesian network possesses close to zero metacognition. It operates under a delusion of total information. It calculates conditional probabilities across a database with absolute, cold precision, but it treats the boundaries of that database as the boundaries of reality itself. In the context of this article and the COIN slide, it does not question the validity of the assumption that Afghanistan could be “pacified” under a Western set of ideals (but then, neither did a “human network” at the time challenge this belief, but that is another story—or perhaps not).

In cognitive science and artificial intelligence philosophy, this is a classic manifestation of the Frame Problem.

In Machine Logic, if an intelligence log doesn’t explicitly contain a node for “Local Resentment” or “Cultural Trauma,” the algorithm’s loss function assumes its value is zero. The real-world consequence of that omission is the Minab School paradigm: the AI lacked the metacognitive awareness to doubt the completeness of its own universe. It optimised a lethal strike based on what was in the frame, completely oblivious to the historical conversion that occurred outside it.

Figure 7: The Minab strike illustrated as an example of lack of metacognition.

When humans lack metacognition, we get the 2009 Spaghetti Slide: a well-intentioned but fundamentally hubristic mess of Western institutional assumptions. But when an AI system lacks metacognition, the consequences can be as hyper-efficient as catastrophic.

By automating targeting or strategic forecasting through systems that cannot question their own ontology, we are industrialising our blind spots. We are building systems that can run a million sensitivity simulations a second, yet remain profoundly incapable of asking the single most important question in strategy: “What if our map is entirely wrong?”

How to enable metacognition in AI

If a single algorithm is left alone with a database, it will always suffer from “What You See Is All There Is” (WYSIATI) bias. However, to give an AI system metacognition, we cannot simply feed it more data; we must change its fundamental relationship with certainty. We must transition from unified optimisation pipelines to adversarial, multi-agent frameworks designed to generate deliberate epistemic friction. In a military context, this manifests as a Civilian Contradiction Engine.

Instead of allowing a targeting or predictive algorithm to calculate a single, unmitigated probability score, a Contradiction Engine splits the cognitive load between independent agents with opposed loss functions. The Primary Agent optimises for tactical alignment (e.g., matching a visual signature to a known threat profile). The Contradiction Agent operates on an independent ontology, ingested primarily from civilian, open-source infrastructure registries, local humanitarian reports, and regional historical media. Its sole metric of success is its ability to find data points that invalidate the primary agent’s conclusions.

If the primary model identifies an asset as an active adversary compound, the Contradiction Engine aggressively cross-examines the timeline. If it surfaces a localised text, a municipal registry edit, or a cultural anomaly indicating the site’s function has shifted, it triggers an automated structural halt—forcing the human back into the loop not as a rubber stamp, but as an arbiter of a profound logical paradox.

Introducing a Contradiction Engine is another way to make AI more accountable, as I referred to in my previous article, where I suggested that slop could be avoided with API calls to external validators. This would be expensive, but not as expensive as conducting a 20-year war for 2 trillion dollars with no result.

This is not a novel concept. Andrei Kojukhov and Arkady Bovshover discuss it *here*. The authors suggest decomposing AI systems into autonomous, specialised agents (detection, hypothesis formation, contextualization) overseen by a distinct “meta-cognitive judgment layer” that measures internal agent disagreement and operational uncertainty before granting decision readiness.

Another solution to AI “safety” could be via Deliberative Debate (The Game-Theoretic Approach). This is a framework where two AI agents debate a complex question in front of a human judge. Game theory dictates that it is mathematically easier for an agent to expose a lie or a framing error in its opponent’s argument than it is to successfully maintain a false premise. It is discussed by Christiano and Amodei *here*. Their argument is this:

We could imagine a system trained to both give answers and point out flaws in answers; this gives a third level of difficulty. Flaws themselves may be too hard to judge: flaws could have their own flaws that must be pointed out to a human. And flaws of flaws can have flaws, etc.

This hierarchy of alignment tasks has a natural limit: a debate between competing agents where agents make arguments, other agents poke holes in those arguments, and so on until we have enough information to decide the truth.

This argues that a reliable way for a machine to find the blind spot in its own map could be through adversarial competitive dialogue.

Figure 8: how deliberative debate can fine tune AI chains of thought.

Conclusion

Now we can tie the entire narrative together by returning to General McChrystal and the 2009 PA Consulting slide. The ultimate tragedy of the “Worst PowerPoint Slide Ever Created” was not so much that it was too complicated for human eyes, but that it was a map designed to look like its own conclusion.

True strategic metacognition—whether wielded by a human commander or programmed into a modern machine-learning architecture — lies in recognising that the map is a living, volatile hypothesis. By implementing Contradiction Engines, we accept a core truth of “Scientific Realism”: the goal of AI in conflict (or indeed anywhere) shouldn’t be to accelerate the speed of our actions, but to broaden the perimeter of our lack of knowledge and attempt to fill it with valid data before we take those actions.