Explainable Artificial Intelligence: First World Conference, xAI 2023, edited by Luca Longo, represents a monumental compilation of cutting-edge research that goes into the burgeoning field of explainable artificial intelligence (XAI). This three-volume set emerges from the proceedings of the inaugural xAI 2023 conference held in Lisbon, Portugal, and encompasses a comprehensive spectrum of 94 rigorously peer-reviewed papers selected from 220 submissions. The collection shows the interdisciplinary efforts to address the opacity inherent in complex AI systems, a challenge that has profound implications across societal, developmental, and business contexts.
In an era where artificial intelligence permeates various facets of human endeavor—from healthcare and finance to manufacturing and autonomous systems—the imperative for transparency and interpretability in AI decision-making processes has never been more critical. The book explores this imperative by organizing the contributions into thematic sections that cover interdisciplinary perspectives, model-agnostic explanations, causality, applications in specific domains like finance and healthcare, visual representations, human-centered explanations, and more. Each section creates a narrative that collectively displays the multifaceted challenges and solutions associated with making AI systems explainable to human stakeholders.
Explainable Artificial Intelligence: First World Conference, xAI 2023 is best approached as a panoramic map of a new landscape rather than a manual filled with formulas. It gathers the most representative work presented at the first worldwide meeting devoted to making artificial intelligence understandable to the people who must live with it, rely on it, and be accountable for its outcomes. The three volumes reflect a simple yet demanding idea: if machines are going to help decide who gets a loan, which medical image needs urgent attention, how to schedule a factory line, or what a vehicle should do next, then their suggestions cannot remain impenetrable. They must be able to give reasons that people can weigh, question, and, when necessary, challenge. The collection shows how that demand for clarity changes the way systems are built, tested, and introduced into real organizations.
The editors have arranged hundreds of voices into a coherent conversation. What emerges is not a single definition of explanation, but a careful framing of the many different things we ask of it. Sometimes we want to know which pieces of information most influenced a decision. Sometimes we want a counter-story — what would have had to change for the outcome to be different. Sometimes we want a global picture of how a system behaves across many cases, rather than an account of one decision in isolation. Sometimes we need a practical assurance about accountability — who is responsible when a recommendation goes wrong, and how we will trace that responsibility. Because these needs differ by domain and by role, no single method can satisfy them all. The volumes embrace that plurality and insist that claims about clarity be matched to the situations in which clarity is required.
One of the most memorable threads in the book follows a story from a car factory. A team built a decision system that learned, over time, how to release painted car bodies from a buffer toward final assembly so as to reduce stoppages and delays. On paper, and in early trials, the system improved the rhythm of the line. Yet supervisors did not adopt its recommendations whenever they could not see why a suggested sequence would be better than their own. The system was designed to advise, not to take control, and responsibility for the line remained human. In that setting, a recommendation without a reason was not simply unsatisfying; it was unusable. The lesson the authors draw is clear and widely applicable: if people remain accountable, then clarity is not an optional extra. It is a design requirement. Explanations must be shaped to the actual questions supervisors ask during their shift — why this order now, what will happen if we delay this unit, which rule or constraint is at stake — and they must arrive in time to be useful. The book treats this not as a cautionary tale about stubborn workers, but as evidence that organizational life has its own logic and that understandable systems must respect it.
A different set of contributions explores how to step back from single decisions and see the larger patterns that guide a system’s behavior. Instead of asking why one image, record, or packet was labeled as it was, the authors ask: across many cases, what are the consistent signals the system pays attention to, and where does its attention vary? They introduce visual summaries that let an analyst scan a complex model the way a physician scans a dashboard: colors and sizes hint at which features matter and how stable their influence is across the dataset. With a few interactions — filtering, zooming, focusing on a subset of interest — one can move from an overall picture to the precise details that prompted it. The important point is not the paint on the chart; it is the structure for inquiry. The visuals do not pretend to reveal the machine’s “true mind.” They help humans decide where to look next, which expectations are confirmed, and which surprises deserve a closer investigation. They turn a thicket of numbers into a navigable landscape.
The volumes also contain a quiet but revealing experiment: take a powerful language system, the kind that has transformed translation and question answering, and ask it to solve tasks that are not obviously linguistic, such as bracketed arithmetic puzzles or predicting the winner in a simple board game. Because these puzzles have a known structure and a known solution path, researchers can watch how the system changes its internal representations while it works and ask whether parts of the system appear to specialize for particular steps. Two conclusions stand out. First, on tasks with a clear sequence of subproblems, the system seems to develop a layered way of working that resembles the step-by-step strategies a human might use, and this observation can be used to adjust and fine-tune it more efficiently. Second, on tasks whose structure clashes with how the data are presented to the system, the usual “look at what it focused on” visualizations can be misleading. The broader message is a modest one: if we want insight into how these systems behave, we should choose tests where we already understand, at least in outline, what a good line of reasoning would look like.
Throughout the three volumes, the same disciplined attitude returns. Explanations are not a charm that automatically creates trust. They are instruments for making and evaluating decisions. They should be judged by the work they do for particular people in particular situations. A plan that aims to help a radiologist spot spurious patterns in scans must be tested with radiologists and real scans; a tool that claims to improve public accountability must be tested against the procedures and records that define accountability in that setting; a technique that claims to reflect how a system actually works must be confronted with checks designed to expose wishful thinking and mere plausibility. The book places success stories next to negative results and sharp critiques, not to dampen enthusiasm but to anchor the field in evidence rather than spectacle.
Because the question touches so many domains, the volumes range widely. In healthcare, the emphasis falls on whether a clearer system truly helps a clinician do their work better, faster, or more safely — and on how to spot seemingly convincing patterns that are, in fact, artifacts of where and how data were collected. In finance, the focus shifts toward fairness, recourse, and the stability of reasons over time, given that people must be able to understand and contest decisions that affect their lives. In systems that learn from networks — social graphs, molecules, transport links — the very idea of a “feature” looks different, and so do the forms of clarity that make sense; what matters are recurring patterns of connection, and not every imagined change is feasible in the world that the system models. At every turn, the editors have sought contributions that translate broad values — transparency, reliability, human oversight — into concrete practices and tests. What counts, finally, is whether a doctor, a loan officer, a factory supervisor, or a regulator can do their job better because a system can show its work.
The setting of the conference itself is part of the story. A first world meeting, with reviewers drawn from many countries and specialties, special tracks and demonstrations, and a doctoral program that invites early-career researchers into the debate, is not only a venue but a method. It brings approaches that might otherwise talk past one another into direct comparison. It encourages authors to say plainly what their contribution claims, what kind of evidence supports it, and where its limits lie. It makes it possible to see common patterns: techniques that travel well from one domain to another, patterns of failure that recur and can be anticipated, evaluation habits that deserve to become standards rather than afterthoughts. The result is an unusually clear snapshot of what “good practice” looked like at a particular moment.
The editorial stance is deliberate. Rather than pretending that there is a single path to clarity, the volumes accept that different audiences need different kinds of reasons and that these reasons carry different costs. An explanation that helps a non-expert make sense of a decision may not satisfy a safety auditor who needs traceable evidence; a mathematically pure account of why a system behaves as it does may be too slow or too complicated to use on the factory floor. The book does not try to erase these tensions. It treats them as design choices that must be made openly, justified, and tested. That insistence on matching claims to contexts is, in a sense, the deepest kind of clarity the field can offer.
Readers who come to these volumes looking for a grand unified recipe will not find one, and that is to the book’s credit. What they will find is a set of patterns and exemplars that can guide practice. When a system’s advice will remain advice and a person will carry the responsibility, make the reasons fit the questions that person actually asks under time pressure. When the stakes are social and legal, build records that show how conclusions were reached and how they can be reconsidered. When the goal is to understand the overall tendencies of a complex model, provide overviews that can be refined and interrogated, rather than point-and-click miracles. When insight into a system’s inner workings is promised, test it on problems where a sensible path is known, and be willing to report where the promise fails.
In this spirit, Explainable Artificial Intelligence: First World Conference, xAI 2023 is more than a record of a meeting. It is a careful argument for how a field should hold itself to account. It turns the urgent, sometimes vague demand that machines “be transparent” into a set of concrete tasks, and it shows, across many examples, how those tasks can be met — and where they cannot. It invites engineers, managers, clinicians, policy-makers, and curious readers to treat clarity not as a decorative layer but as a property that changes how systems are conceived and judged from the start. The result is a demanding, humane, and, above all, useful portrait of what it will take for artificial intelligence to be not only powerful but also answerable to the societies that increasingly rely on it.
Simon Gros 
Leave a comment