How can you avoid common myths about explainable AI?

3 Myth 3: XAI is only for end-users and consumers

Another common myth about XAI is that it is only for end-users and consumers, such as patients, customers, or citizens, that may want to know why an AI system made a certain decision or recommendation that affects them. However, this is not true, as XAI is also for developers and practitioners, such as data scientists, engineers, or analysts, that may want to know how an AI system works, how to improve it, or how to debug it. For example, XAI can help you to understand the logic, behavior, or limitations of your model, to optimize its performance, accuracy, or robustness, or to detect and fix errors, bugs, or anomalies in your model.

4 Myth 4: XAI is only for post-hoc analysis

Another common myth about XAI is that it is only for post-hoc analysis, that is, explaining an AI model or outcome after it has been trained or deployed. However, this is not true, as XAI can also be integrated into the design and development process of an AI system, that is, creating an AI model or outcome that is inherently explainable or interpretable. For example, XAI can help you to choose the most suitable model architecture, feature selection, or regularization technique for your problem, to incorporate domain knowledge, ethical principles, or user feedback into your model, or to evaluate and validate your model against various criteria, such as accuracy, fairness, or robustness.

5 Myth 5: XAI is only for human consumption

Another common myth about XAI is that it is only for human consumption, that is, explaining an AI model or outcome to a human user or stakeholder. However, this is not true, as XAI can also be used for machine consumption, that is, explaining an AI model or outcome to another AI system or agent. For example, XAI can help you to enable collaboration, coordination, or communication between multiple AI systems or agents, to facilitate learning, adaptation, or transfer of knowledge or skills between different AI systems or agents, or to monitor, control, or audit the behavior or performance of an AI system or agent.

6 Myth 6: XAI is a solved problem

The last common myth about XAI is that it is a solved problem, that is, there are already enough tools, methods, or standards for creating and evaluating explainable AI models or outcomes. However, this is not true, as XAI is still an active and evolving research field, that faces many challenges and opportunities for further development and innovation. For example, XAI needs to address the trade-offs, limitations, or risks of different explanation techniques, such as accuracy, complexity, scalability, or privacy. XAI also needs to explore new ways to measure, compare, or improve the quality, usefulness, or impact of explanations, such as validity, reliability, relevance, or satisfaction. XAI also needs to adapt to the changing needs, expectations, or preferences of different users, domains, or scenarios, such as personalization, contextualization, or interaction.

7 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?