What is eXplainable AI?

  Explainability for Multimodal AI Emerging Frontiers Series Introduction: A New Kind of Black Box Imagine asking a state-of-the-art AI system to describe a picture of a cat stalking through tall grass. The AI captions it: “Stealth hunter.” If you press it to explain why it chose those words, what should the answer look like? Was it the elongated posture of the animal? The narrow pupils? The association of tall grass with predation? Or did the model simply learn from thousands of caption–image pairs online that “cat + tall grass” often co-occurs with “hunting”? Welcome to the world of multimodal AI —models that can process and integrate more than one kind of input, such as text, images, and audio. While this ability brings astonishing capabilities—like describing videos, tutoring students with diagrams, or analyzing medical scans—it also creates new challenges for XAI (Explainable Artificial Intelligence) . The central question is not just “Why did the AI produce this output...

  Article 4: The Psychology of AI Explanations: How Much Detail is Too Much? One of the biggest paradoxes of explainability is this: users often say they want more transparency, but in practice, too much detail erodes trust . Cognitive load Psychology gives us a tool here: cognitive load , or the mental effort required to process information. Too little explanation feels dismissive: “Because the model said so.” Too much explanation causes fatigue: long technical justifications confuse or alienate. Just right explanations —the “Goldilocks zone”—help users without overwhelming them. Progressive disclosure UX (User Experience) designers use the principle of progressive disclosure : give users a simple answer first, then let them drill deeper if they wish. Explanations could work the same way: Layer 1 : Simple rationale (“The system recommended this route because it’s faster.”) Layer 2 : Supporting evidence (traffic data, accident reports). Layer 3 : Technic...

  Article 3: Why Doctors, Judges, and Teachers Need Different Kinds of Explanations Not all explanations are created equal. A doctor, a judge, and a teacher may all use AI , but the explanations they need are radically different. Doctors: Causal reasoning Doctors need to know: What symptoms or features led to this recommendation? If an AI suggests pneumonia , the doctor must see the path from data to diagnosis . Otherwise, they cannot justify treatment decisions—or maintain patient trust. Judges: Procedural fairness Judges are not only making decisions; they are upholding legitimacy. They need to know: Did the AI apply rules fairly and consistently? For them, explanations must emphasize due process , not just outcomes. Teachers: Pedagogical insight Teachers want to understand student thinking . If an AI marks an essay low, the teacher needs more than “grammar errors detected.” They need insight into the learning process —so they can guide growth, not just assign grades. C...

  Article 2: Trust, Transparency, and Human-in-the-Loop Systems Trust is not binary. We don’t simply trust or distrust AI. Instead, trust is graded and contextual . You may happily trust Google Maps to guide you through traffic but hesitate to trust an AI that diagnoses cancer . Three concepts— trust, transparency, and human-in-the-loop (HITL) —form a triangle of human-centered XAI . Trust Trust is the willingness to rely on a system even when outcomes are uncertain. In AI, overtrust can be dangerous ( automation bias ), but undertrust can make systems useless. Transparency Transparency means how much the system reveals about itself. Too little transparency makes AI feel manipulative. Too much can overwhelm. Human-in-the-loop (HITL) HITL refers to systems where a person remains involved in critical decision points— reviewing, overriding, or adjusting AI outputs. HITL is often described as the “safety net” for explainability. Case study: Aviation autopilot Modern plane...

Article 1: Explainability for Designers: Making AI Understandable to End-Users When most people hear the phrase XAI (Explainable Artificial Intelligence) , they imagine computer scientists writing white papers or regulators debating accountability. But there’s another group at the center of AI adoption that rarely gets enough credit: designers . Designers don’t create the mathematical guts of AI models. Instead, they shape the way humans encounter AI . And when AI feels like a black box , it’s designers who decide where to cut a window into that box—so that end-users can peek inside without being overwhelmed. Why designers matter in XAI A designer’s role is not just visual polish. It’s cognitive scaffolding: helping users navigate AI-driven decision-making without losing confidence, autonomy, or clarity. Good design : Anticipates user questions— Why did the AI suggest this? Can I trust it? What are my options? Poor design : Drowns users in numbers (confidence scores, p-valu...

  Philosophy of Explainability (Part 4) What AI Can Teach Us About Human Cognition Introduction: The Mirror Effect We often approach AI as though it must catch up to human intelligence . But in the debate over explainability , the opposite happens: AI forces us to confront the limits of our own explanations . When we ask, “Why did the AI make that decision?” we realize that humans themselves rarely provide perfect, transparent explanations for their own actions. In this sense, AI explainability is not just about making machines more human-like. It is about understanding the ways in which humans have always been machine-like: opaque, approximate, and narrative-driven in our reasoning. Human Explanations: Post Hoc Stories Cognitive science research reveals that much of human explanation is post hoc rationalization —stories we tell ourselves after the fact. Split-brain studies show that when one hemisphere takes an action, the other hemisphere invents a plausible reason—e...

  Philosophy of Explainability (Part 3) Explainability as a Moral Imperative Introduction: Why Ethics Enters the Room We could treat explainability as a technical challenge —something for engineers and data scientists to solve. But the stakes are higher. When AI systems deny a loan, recommend prison sentences, or decide who receives scarce medical resources, explanation is not optional. It is a moral imperative . Ethics enters the room because explainability shapes autonomy, fairness, and accountability . To refuse explanation is to deny people the ability to contest, understand, or influence the decisions that affect their lives. Philosophical Grounding: Duties vs. Consequences Philosophers approach moral responsibility in two dominant traditions: Deontological duty ( Kantian ethics ) : People must be treated as rational agents, capable of understanding reasons. If an AI affects you, you deserve an explanation because it is a matter of respect. Consequentialist ethics...