Learn how to explain an AI system in order to improve the user experience.
Design Strategy Overview
Based on industry and academic research, we propose the following design strategy for explainability in AI. This strategy is the beginning of a conversation; we hope you will experiment, play, use, and break what you see here and send us your feedback so that we can continue to iterate!
Who are your users and why do they need an explanation?
Identify the distinct groups of people who are interested in explanations from your AI system and understand the nuances within these groups. There will likely be varying degrees of factors such as domain expertise, self-confidence, attitudes towards AI, and knowledge of how AI works, all of which can influence trust and how people understand the system. For each user group, identify what triggers a need for an explanation as well as the underlying motivations and expectations.
By identifying and understanding your users, you can ensure the explanation matches their needs and capabilities.There are typically four distinct groups to consider:
The people making decisions with AI
Decision makers are people who use the recommendations of an AI system to make a decision, such as a physician or loan officer.
Decision makers seek explanations that can build their trust and confidence in the system’s recommendations and possibly provide them with additional insight to improve their future decisions and understanding of the system. These users will have a high need for domain sophistication, but will also have less tolerance for complex explanations.
The people affected by the decisions
Affected users are people who are impacted by the recommendations made by an AI system, such as patients or loan applicants. In some scenarios, a person may be both a decision maker and affected user.
Affected users seek explanations that can help them understand if they were treated fairly and what factor(s) could be changed to get a different result. These users need reasons for their outcomes communicated in a simple and direct way and often have a lower threshold for both complexity and domain information.
The people checking the system
Regulatory bodies include government agencies who define and enforce relevant policies, such as the European Union’s General Data Protection Regulation (GDPR). In 2016, the “right to explanation” was approved, requiring that data subjects receive meaningful information about the logic involved in automated decision-making systems.
Regulatory bodies seek explanations that enable them to ensure decisions are made in a safe and fair manner. Their needs may be satisfied by showing the overall process, including training data, is free of negative societal impact and may not be able to consume a high level of complexity.
The people behind the system
Internal stakeholders are those who build and deploy an AI system, especially technical individuals such as data scientists and developers.
Internal stakeholders seek explanations that help them know if the system is working as expected, how to diagnose and improve it, and possibly gain insight from its decisions. They are likely to need and understand a more complex explanation of the system’s inner workings to take action accordingly.
Explanations are typically responses to questions, and as such, user needs and triggers for explainability can be written in the form of questions. The following questions are based on common types of information that AI systems can present to users.
What did the system do?
What did the system do?
Why did the system do ____?
Why did the system do ____?
Why did the system not do ___?
Why did the system not decide this plant is safe?
What would the system do if ___ happens?
What would the system predict if the plant was smooth instead of thorny?
How (under what condition) does it do ___?
How does the system decide a plant is poisonous?
What are the changes required for this instance to get a different prediction?
What would need to change for this plant to be predicted safe?
Still Be This
What is the scope of change permitted to still get the same prediction?
How much would need to change for this plant to still be predicted poisonous?
How certain is the system in a prediction or outcome?
How certain is the system that this plant is poisonous?
What data does the system learn from?
What information does the system use to determine whether a plant is safe or poisonous?
What are the possible outputs that the system can produce?
What can the system detect about this plant?
How It Works
What is the overall model of how the system works?
How does the system make its predictions?
When do users need an explanation?
A mental model is a person’s understanding of a system and how it works. Mental models help people set expectations of AI system capabilities, constraints, and value. Expectations impact user satisfaction, behavior, and acceptance of an AI system. When a person’s mental model does not match how the system actually works, it often leads to frustration, misuse, and even product abandonment.
As a result, mental models play a key role in calibrating trust in human-AI interaction. Mental models can change as a person interacts with an AI system, therefore the need for an explanation should be contextualized in the phase of a user’s experience. Within each phase, consider existing mental models and how to calibrate them accordingly.
Onboarding should not only introduce the system but also set the users’ expectations of how it will work, what it can do, and how accurate it is.
During Regular Interaction
Providing explanations during regular interaction enables users to understand the system, identify issues, and intervene as needed.
Explaining system errors can help users understand when the system is likely to err (its “error boundary”), adjust their expectations, and repair trust.
While system updates can improve the AI’s performance, they may also lead to changes that are at odds with the user’s current mental model.
What kind of explanation should be used?
When selecting an explainability method, it is important to consider the type of AI system you are explaining (supervised vs. unsupervised) as well as the relationship between the AI system’s prediction and its explanation. The complexity of an AI system is directly related to its ability to be explained; the more complex the model, the more difficult it is to interpret and explain. There are two main relationships between an AI system and an explainability approach:
Ante-hoc approaches use the same model for predictions and explanations. These approaches are thought to provide full transparency and are typically model-specific because they are designed for and only applicable to a specific model.
Post-hoc approaches use a different model to reverse engineer the inner works of the original model and provide explanations. These approaches are thought to lighten the black box of complex models and are typically model-agnostic because they are designed to work with any type of model.
A common way to categorize explanation methods is by scope: global or local. Global or general system explanations describe how the system behaves while local or specific output explanations discuss the rationale behind a specific output. There is a promising line of work that is focusing on combining the strengths and benefits of both local and global explanations, suggesting a hybridized approach may be a possible human-in-the-loop workflow.
Popular explanation methods for supervised machine learning are shown below.
Global explanations help users understand and evaluate the system.
Describe the weights of features used by the model (including visualization that shows the weights of features).
Decision Tree Approximation
Approximate the model to an interpretable decision tree.
Approximate the model to a set of rules, e.g. if-then rules.
Show what information the system has access to in order to make decisions.
Show what the system is able to do.
Local explanations help users examine individual cases, which can help with identifying fairness discrepancies and calibrate trust on a case-by-case basis.
Feature Importance and Saliency
Show how features of the instance contribute to the model’s prediction (including causes in parts of an image or text).
Rules or Trees
Describe the rules or a decision-tree path that the instance fits to guarantee the prediction.
Contrastive or Counterfactual Features
Describe the feature(s) that will change the prediction if perturbed, absent or present.
Prototypical or Representative Examples
Provide example(s) similar to the instance and with the same record as the prediction.
Provide example(s) with small differences from the instance but with a different record from the prediction.
Feature Influence or Relevance
Show how the prediction changes corresponding to changes of a feature (often in a visualization format).
Explain how certain the AI is in its prediction, e.g. through category (high, low), numbers, N-best alternatives, or visualizations.
Explanation fidelity includes soundness and completeness. Soundness refers to how truthful an explanation is with respect to the underlying predictive model. Completeness measures how well an explanation generalizes; in other words, what extent it covers the underlying predictive model.
Explanations can be static or interactive. Interactive explanations accommodates a wider array of user needs and expectations. Some examples of interactivity include explanations that are reversible, collect and respond to user feedback, and allow adjustment of granularity.
Explanations can be delivered as a summarization (typically with statistics), visualization, text, formal argumentation, or a mixture of the above.
Questions to Explanations
Here are recommended connections between questions and common explanation methods.
How can explanations be assessed and validated?
Explainability approaches can be assessed with product teams (UX, product management, and engineering) using the following dimensions:
Explainability approaches should also be validated by users based on what is being tested. For example, an application level validation approach is best suited for testing concrete applications of an explanation, while a human level approach is best suited for testing more general notions of the quality of an explanation.
- Adadi, A., & Berrada, M. (2018). Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access, 6, 52138-52160.
- Amershi, S., Inkpen, K., Teevan, J., Kikin-Gil, R., Horvitz, E., Weld, D., … Bennett, P. N. (2019). Guidelines for Human-AI Interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI 19.
- Doshi-Velez, F., & Kim, B. (2017). Towards A Rigorous Science of Interpretable Machine Learning. arXiv: Machine Learning.
- Google. (n.d.). People AI Guidebook. Retrieved from https://pair.withgoogle.com/chapter/explainability-trust/
- Hind, M., Wei, D., Campbell, M., Codella, N.C., Dhurandhar, A., Mojsilovic, A., Ramamurthy, K.N., & Varshney, K.R. (2018). TED: Teaching AI to Explain its Decisions. AIES '19.
- Kocielnik, R., Amershi, S., & Bennett, P.N. (2019). Will You Accept an Imperfect AI?: Exploring Designs for Adjusting End-user Expectations of AI Systems. CHI '19.
- Liao, Q. V., Gruen, D., & Miller, S. (2020). Questioning the AI: Informing Design Practices for Explainable AI User Experiences. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems - CHI 20.
- Lim, B.Y., & Dey, A.K. (2009). Assessing demand for intelligibility in context-aware applications. Proceedings of the 11th international conference on Ubiquitous computing.
- Lim, B.Y., Dey, A.K., & Avrahami, D. (2009). Why and why not explanations improve the intelligibility of context-aware intelligent systems. CHI.
- Norman, D. (1988). The Design of Everyday Things.
- Sokol, K., & Flach, P.A. (2020). Explainability fact sheets: a framework for systematic assessment of explainable approaches. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency.
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License