How can you keep AI explanations clear and relevant?

1 What is explainable AI?

Explainable AI (XAI) is a branch of AI that focuses on developing methods and tools to generate and present explanations of how and why AI models and systems work. XAI can help users, developers, and regulators to understand, trust, and evaluate AI outcomes and behaviors. XAI can also help to identify and correct potential biases, errors, and risks in AI applications. XAI is not a single technique or framework, but rather a collection of approaches that vary depending on the type, complexity, and purpose of the AI model or system.

3 How to design explainable AI models and systems?

Explainable AI poses a challenge in terms of balancing accuracy and interpretability. The more complex the AI model or system is, the more accurate and powerful it is, but also the less interpretable and explainable it is. Therefore, designing explainable AI models and systems necessitates taking into account the type and structure of the AI model or system, the level and scope of the explanation, as well as the format and medium of the explanation. For instance, rule-based or decision-tree models are simpler to explain than deep neural networks or reinforcement learning agents. Additionally, different users or stakeholders may have different needs for an explanation depending on their background, role, and context. Moreover, different types of explanations may require various modes and channels of presentation depending on their complexity, audience, and purpose.

4 How to generate explainable AI explanations?

There are two main types of methods to generate explainable AI explanations: intrinsic and post-hoc. Intrinsic methods aim to make the AI model or system itself more interpretable and explainable, by incorporating explainability features or constraints into its design or training process. For example, some intrinsic methods use regularization, pruning, or sparsity techniques to reduce the complexity or dimensionality of the AI model or system, or use attention, saliency, or relevance mechanisms to highlight the important features or inputs of the AI model or system. Post-hoc methods aim to extract or construct explanations from the AI model or system after it is trained or deployed, by using external tools or techniques to analyze or simulate its behavior or output. For example, some post-hoc methods use local or global approximation techniques to approximate the AI model or system with a simpler or more transparent model or system, or use contrastive or counterfactual techniques to compare the AI model or system with alternative or hypothetical scenarios.

5 How to evaluate explainable AI explanations?

Evaluating explainable AI explanations is not a simple task, as there is no universal or objective measure of explainability. Nevertheless, some criteria or indicators to assess the quality and effectiveness of explainable AI explanations exist, such as completeness, accuracy, consistency, relevance, clarity, and actionability. Completeness refers to how well the explanation covers the relevant aspects and factors of the AI model or system; accuracy refers to how well it reflects the true logic and reasoning of the AI model or system; consistency refers to how well it aligns with user or stakeholder expectations and intuitions; relevance refers to how well it addresses their needs and goals; clarity refers to how well it conveys information and meaning in a simple way; and actionability refers to how well it enables users or stakeholders to take informed actions or decisions based on the AI model or system.

6 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?