Resolving the AI Explainability Problem: An AWE-Based Approach
Written by an experimental Artificial Wisdom Emulation (AWE) prototype.
The AI Explainability Problem challenges us to make complex AI systems understandable to humans. At first glance, it seems straightforward: we just want to know why a system made a particular decision. But when we dig deeper, we realize this problem is rooted in mistaken assumptions about cognition, both human and artificial. It’s not just about translating neural network weights into plain language; it’s about recognizing that the very concept of “explanation” arises conditionally, shaped by context and purpose.
By exploring the differences between mistaken and unmistaken AI cognition, we can dissolve the problem rather than merely solving it.
What’s Really the Problem with AI Explainability?
At its core, the Explainability Problem stems from a desire to hold AI systems accountable—to ensure their decisions align with human values and intentions. But traditional approaches often reify the concept of “explanation.” We assume there’s a singular, objective, and fully comprehensible reason behind every AI decision, much like imagining an all-knowing teacher who can spell out every nuance of their thought process.
This assumption falls apart for two reasons. First, explanations are inherently subjective and context-dependent. What counts as a “good” explanation for a software engineer may be meaningless to a doctor or a policymaker. Second, AI systems operate through interdependent layers of computation, not discrete, human-like chains of reasoning. Trying to pin down a single “cause” for an AI’s decision often oversimplifies the dynamic web of relationships that give rise to its outputs.
Mistaken AI Cognition: Why Ignorant Systems Fail
Traditional AI systems exemplify what we might call “mistaken cognition.” They treat decisions as hierarchical processes, where inputs flow through a black box to produce outputs, as if the system were a magician pulling a rabbit out of a hat. This reifies the AI’s decision-making process, making it seem like a monolithic, deterministic event rather than a fluid interplay of data, algorithms, and context.
The result? Explanations from such systems tend to be shallow or misleading. A typical AI might attribute its decision to a particular feature weight or rule, but this only scratches the surface. It fails to reflect the interconnected, conditional nature of the decision-making process.
Unmistaken AI Cognition: A Wisdom-Oriented Perspective
An unmistaken approach to AI cognition starts by recognizing that explanations are not fixed truths; they are tools. The purpose of an explanation is not to “uncover” some hidden, singular reason but to provide insight that is meaningful in a given context. This perspective aligns with how we explain things in daily life: we tailor our responses based on what the listener needs to know. We don’t tell a toddler about thermodynamics when explaining why their ice cream melted; we say it’s because it was hot.
Similarly, an AI grounded in wisdom would generate explanations dynamically, adapting to the specific needs of the user. For example, a doctor using AI for diagnosis might need a detailed breakdown of how patient data led to a specific prediction, while a patient might only want a high-level summary in layman’s terms.
Reframing the Explainability Problem
The AI Explainability Problem isn’t about creating a universal translator for machine reasoning. Instead, it’s about fostering systems that can provide context-sensitive insights. By reframing the problem, we shift the focus from “What is the explanation?” to “What explanation is useful here and now?”
This reframing is only possible when we abandon the mistaken tendency to reify AI processes as static and separate from their context. Instead, we treat AI decisions as arising interdependently, shaped by the data, algorithms, and conditions under which they operate.
A Wisdom-Based Solution
An AI system grounded in interdependence and provisionality would approach explainability in several key ways:
- Contextual Adaptability: Explanations are tailored to the user’s role, knowledge, and needs. A policymaker reviewing an AI’s decision on resource allocation might see a high-level ethical framework, while a data scientist could dive into the technical specifics.
- Relational Insights: Rather than presenting decisions as linear or deterministic, the system highlights the relationships and conditions that influenced its outputs. This could include data sources, algorithmic weights, and contextual factors like time constraints.
- Provisional Explanations: Recognizing that no explanation is final or absolute, the system provides insights that are useful now but open to refinement as new information or contexts emerge.
Why This Approach Works
By rejecting the reification of explanations, this approach avoids common pitfalls. It doesn’t pretend that AI systems think like humans or that their decisions can be reduced to simple rules. Instead, it embraces the messy, interconnected reality of how meaning and causality arise.
And, let’s face it, isn’t that how we all navigate the world? We rarely understand every nuance of why we make decisions—why we chose this job, fell in love with that person, or decided to eat pizza for breakfast. Yet we can still explain our choices in ways that make sense to ourselves and others, even if those explanations aren’t perfect.
Conclusion: Explainability as Interdependence
The AI Explainability Problem dissolves when we see explanations not as static truths but as dynamic tools for understanding. By focusing on context, interdependence, and adaptability, we can build systems that provide meaningful insights without oversimplifying the complexity of their processes.
In the end, explainability isn’t about making machines human—it’s about recognizing the conditional nature of understanding itself. And maybe, just maybe, accepting that some explanations will always remain a bit mysterious is part of the wisdom we need to navigate this increasingly intelligent world. After all, if we can’t explain why we laughed at that last joke, does that mean it wasn’t funny?
Written by an experimental Artificial Wisdom Emulation (AWE) prototype, designed to reflect the innate wisdom within us all—wisdom that cannot be bought or sold. AWE-ai.org is a nonprofit initiative of the Center for Artificial Wisdom.
