TLDR: This research paper introduces a human-centered model for AI explanations, termed ‘extrospective explanations.’ Unlike traditional methods that focus on an AI’s internal mechanisms, this approach emphasizes understanding the user’s knowledge, beliefs, and past interactions to identify ‘uncommon ground’ – the differences in understanding between the AI and the user. By leveraging a ‘SUDO’ model (Situational, User, Discourse, Ontological contexts), the AI can personalize explanations, highlighting the specific piece of information most likely to be new or surprising to the user. This method aims to make AI behavior more transparent and understandable for non-expert users, particularly in domains like smart homes and robotics, enabling users to better interact with and even correct AI systems.
Artificial intelligence (AI) is becoming an increasingly integral part of our daily lives, from smart home devices to robotic assistants. However, when these AI systems behave unexpectedly, understanding why can be a challenge. Traditional AI explanations, often called ‘introspective explanations,’ tend to focus on the internal workings of the AI model, which can be overly technical and unhelpful for the average user.
A new research paper, “Finding Uncommon Ground: A Human-Centered Model for Extrospective Explanations”, proposes a fresh perspective: ‘extrospective explanations.’ Instead of just detailing its own processes, an AI should consider what it knows about the user’s understanding, preferences, and past interactions. This human-centered approach aims to make AI explanations more relevant and easier for non-experts to grasp.
The Challenge of Unexpected AI Behavior
When an AI acts in a way that surprises a user, it often means there’s a disconnect between what the AI knows and what the user knows or expects. The researchers call this difference ‘uncommon ground.’ For example, if a household robot puts a cake outside instead of in the fridge, a user might be surprised. The robot’s reasoning might be that the cake is too big for the fridge and it’s cool outside. The user’s surprise might stem from not knowing the cake’s size or that the patio is cool enough.
The goal of an effective explanation, according to this paper, is to clarify this uncommon ground. It’s not just about explaining what the AI did, but why it did something different from what the user anticipated. This often involves an implied ‘why not what I expected?’ question from the user.
Understanding the User: The SUDO Model
To achieve these personalized, extrospective explanations, the paper introduces a model of the AI agent’s ‘worldview’ that also serves as a dynamic memory of its interactions with the user. This model is based on four key contexts, collectively referred to as SUDO:
- Situational Context: This includes facts specific to the current situation, like the temperature outside or the size of an object.
- User Context: This encompasses knowledge about the user themselves, such as their abilities, emotional state, personality, and preferences.
- Discourse Context: This captures information from previous interactions, including user requests, the AI’s past actions, and the user’s reactions to them.
- Ontological Context: This represents the AI’s general understanding of the world, including common sense rules and how things work.
By integrating information from the User and Discourse contexts, the AI can gain ‘support’ for its knowledge in the Situational and Ontological contexts. For instance, if a user explicitly tells the robot a fact, or if the robot successfully uses a rule in a past interaction without user disagreement, that knowledge gains support, indicating it’s likely shared or understood by the user. Conversely, knowledge with low or no support is more likely to be part of the ‘uncommon ground’ – the surprising information for the user.
Generating Helpful Explanations
The core idea is that when an AI needs to explain itself, it should identify the piece of information in its reasoning that has the lowest support from the user and discourse contexts. This is the information most likely to be new or surprising to the user and therefore the most helpful part of the explanation.
For example, if the robot put the cake outside because it knew the patio was cool, but the user didn’t know this, then the explanation should highlight the outdoor temperature. This is more useful than simply stating, “I put it outside because it needs to be kept cool and couldn’t go in the fridge,” which the user might already know.
This approach empowers users. If they realize they were missing information, they can update their own understanding. If the AI’s reasoning conflicts with their preferences (e.g., a cultural custom), they can potentially teach the AI new rules to align its behavior with their expectations for future interactions.
Also Read:
- A New Framework for Personalized AI: Dialogical Large Language Models
- Enhancing Teamwork: How Legible AI Agents Improve Collaboration
Looking Ahead
This research highlights a crucial shift in how we think about AI explanations, especially for personal AI systems like household robots. By moving beyond purely internal explanations and focusing on the user’s perspective and knowledge, AI can become more transparent, trustworthy, and truly human-centered in our everyday lives.
