Deep Dive Deconstructing Causality: The Detective’s Whiteboard and AI’s Quest for Understanding

Imagine a detective standing in front of a whiteboard, piecing together clues to solve a complex crime. The whiteboard is filled with photos, notes, and strings connecting suspects, motives, and timelines. The detective’s task is not just to collect evidence but to construct a coherent narrative that explains why and how the crime occurred. This process of constructing causality—connecting the dots to form a story—is at the heart of human reasoning. But what happens when we ask artificial intelligence to perform the same task?

In this article, we’ll explore the challenges of understanding causality in AI systems through the lens of the detective’s whiteboard. We’ll see how causality is not an objective property of the universe but a constructed framework imposed by our cognitive systems. We’ll examine how current AI, particularly Large Language Models (LLMs), excels at correlation but struggles with genuine causal reasoning. Finally, we’ll propose a solution: augmenting LLMs with knowledge graphs—a “digital whiteboard”—to bridge the gap between correlation and causation.

Part 1: The Fog of Knowing – The Detective’s Limited Perspective

The Whiteboard and the Fishbowl: Limited Perception in a Complex World

Every detective knows that their investigation is constrained by the evidence they can gather. No matter how thorough, the whiteboard will always be incomplete—a partial representation of a much more complex reality. This limitation mirrors the concept of Umwelt, the subjective world experienced by each organism based on its sensory and cognitive capabilities. Just as a fish in a fishbowl perceives only a fraction of the world outside, a detective sees only a slice of the crime’s full context.

Our understanding of causality is similarly constrained. We construct causal narratives based on the evidence available to us, but these narratives are always provisional, always partial. The whiteboard is not the crime scene; it’s a map, a model, a tool for navigating the complexities of reality.

The Tyranny of Language: Clues and Categories

Language plays a crucial role in shaping our understanding of causality. Just as a detective categorizes clues into suspects, motives, and opportunities, language imposes discrete categories on the continuous flow of reality. This categorization is both a strength and a limitation. It allows us to communicate and reason about the world, but it also distorts our understanding by creating artificial boundaries.

For example, consider how a detective might label a witness as “reliable” or “unreliable.” This binary categorization can bias the investigation, leading the detective to overlook nuanced information that doesn’t fit neatly into these labels. Similarly, language divides the visible light spectrum into discrete colors like red, orange, and yellow, even though the spectrum itself is continuous. These linguistic categories shape not just how we describe the world but how we experience it—and ultimately, how we understand causality.

Part 2: The Elusive Cause – Constructing Causal Narratives on the Whiteboard

The Detective’s Dilemma: Hume’s Shadow and the Problem of Induction

David Hume famously argued that we never directly observe causal connections—only regular successions of events. The detective’s whiteboard is a perfect illustration of this dilemma. The detective sees that the butler was in the library at the time of the murder, but they cannot directly observe whether the butler caused the murder. The causal connection is inferred, not observed.

This gives rise to the problem of induction: how can we justify inferring general causal rules from specific observed instances? The detective relies on past experience to connect the dots, but there’s no logical certainty that these patterns will hold. The butler might have been in the library every night, but that doesn’t mean he’s the murderer. This inherent uncertainty in causal reasoning has profound implications for both human knowledge and artificial intelligence.

The Butler Did It: Correlation vs. Causation

One of the most common pitfalls in causal reasoning is mistaking correlation for causation. The classic example is the “butler did it” trope. The detective notices that the butler was present at every crime scene and concludes that he must be the culprit. But this is a spurious correlation—the butler’s presence might be coincidental, or he might be a red herring.

Current AI systems, particularly LLMs, are highly susceptible to this kind of error. They excel at detecting patterns in data, but they lack the ability to distinguish between genuine causal relationships and spurious correlations. Like a detective who jumps to conclusions, LLMs can arrange clues plausibly but often fail to understand the underlying causal mechanisms.

Part 3: The Clever Horse and the Chatbot – AI’s Struggle with Causal Reasoning

The AI Detective: Mimicking Investigation Without Understanding

Imagine a detective who can perfectly mimic the process of investigation but doesn’t actually understand the crime. This is the situation with current LLMs. They can generate plausible-sounding narratives based on patterns in their training data, but they lack genuine causal reasoning. They are like Clever Hans, the horse who appeared to solve arithmetic problems but was actually responding to subtle cues from his handler.

LLMs are trained to predict the next word in a sequence, not to perform logical deduction or causal reasoning. When faced with a question like “Can high blood pressure cause headaches?” an LLM might generate a plausible answer based on patterns in its training data, but it doesn’t truly understand the causal relationship between the two.

Architectural Limitations: The Single-Pass Detective

The fundamental limitation of LLMs lies in their architecture. They operate as a chaotic collection of data, processing information in a single pass without the ability to revisit, refine, or consolidate evidence. This single-pass, autoregressive nature makes it impossible for LLMs to perform the kind of iterative reasoning that a detective uses to solve a complex case. They can’t discern relevance, eliminate spurious correlations, or build a coherent causal narrative over multiple steps.

Fine-tuning and test-time compute can offer partial improvements, but they don’t fundamentally address these architectural limitations. To achieve genuine causal reasoning, we need a different approach.

Part 4: A Symbolic Crutch – The Digital Whiteboard

The Knowledge Graph: A Digital, Enhanced Whiteboard

The solution lies in augmenting LLMs with knowledge graphs—a “digital whiteboard” that provides a separate, persistent symbolic state for reasoning. Just as a detective uses a whiteboard to organize and refine their investigation, a knowledge graph allows an AI system to iteratively build and revise its understanding of causal relationships.

Knowledge graphs represent entities as nodes and relationships as edges, creating a network of interconnected information that can be traversed and queried. For example, a medical knowledge graph might represent the fact that “high blood pressure can cause headaches” as a directed edge between two nodes. This explicit representation allows for formal reasoning over the graph, enabling the system to perform multi-step causal inference.

Iterative Reasoning: Revisiting and Refining the Narrative

One of the key benefits of a knowledge graph is its ability to support iterative reasoning. Unlike the single-pass nature of LLMs, a knowledge graph allows the system to revisit and refine its conclusions over time. This is crucial for causal reasoning, where initial hypotheses often need to be revised in light of new evidence.

For example, if an AI system initially concludes that “the butler did it” based on a spurious correlation, it can later revise this conclusion when new evidence—such as an alibi or a fingerprint analysis—points to a different suspect. This ability to iteratively refine the narrative is a significant advantage over the static, single-pass reasoning of LLMs.

Beyond Iterative Reasoning: Explainability and Transparency

But the benefits of a knowledge graph extend beyond simply enabling iterative reasoning. It also offers a level of explainability currently absent in LLMs. Because the system must explicitly instantiate evidence, clues, and potential links, we remove one of the most insidious aspects of large language models: inherent uncertainty. Unlike the “black box” nature of LLMs, where reasoning processes are opaque, a knowledge graph provides a transparent and auditable trail of inference. For a specific subset of information, we can even achieve 100% stability or deterministic behavior, knowing precisely why the system reached a particular conclusion.

This transparency isn’t just about understanding what the system is doing, but also about enabling human oversight and collaboration. The graph is readable and can be adjusted for human interaction and interpretation. By agreeing to certain elements and links as part of specific cases, we are allowing standardization and improving consistency in the behaviors of the symbolic state. This opens the door to review, delegation, and cataloging of reasoning processes, fostering trust and accountability.

Part 5: Beyond the Crutch – The Future of AI and the Enduring Mysteries

The Limits of the Whiteboard: Representing the Unrepresentable

Even the most sophisticated whiteboard is still a limited representation of reality. Just as a detective’s investigation is constrained by the evidence they can gather, a knowledge graph is limited by the information it can represent. Much of human knowledge—particularly common sense, tacit knowledge, and embodied experience—resists formalization.

This limitation relates directly to Gödel’s Incompleteness Theorems, which demonstrate that any formal system powerful enough to express basic arithmetic must contain truths it cannot prove. Just as Gödel showed that any formal system will have inherent limitations, a knowledge graph, no matter how comprehensive, will always be incomplete. No matter how sophisticated our digital whiteboards become, they will always have blind spots—truths they cannot capture.

The Detective and the Hotpot: Shared Reality, Divergent Perspectives

As we conclude our exploration of causality, cognition, and AI, let’s return to the detective’s whiteboard. Despite its limitations, the whiteboard—and its digital counterpart, the knowledge graph—offers a path towards more transparent, reliable, and accountable AI systems. By combining the strengths of statistical AI with the explicit reasoning capabilities of knowledge graphs, we can move beyond correlation to achieve true causal understanding, and build systems that allow us to understand and validate that understanding. The quest for causal understanding in AI is far from over, but the detective’s whiteboard, now enhanced with the power of symbolic reasoning, will continue to guide us as we navigate the mysteries of causality in a world that is far more complex than our current models can capture. This isn’t just about building smarter AI; it’s about building AI we can rely on and understand.