Abstract
In human-computer interaction, a technically correct answer can still be misleading if it allows a user to maintain a false belief. This paper formalizes a "Revised Maxim of Quality," requiring expert systems to provide additional information to block potential false inferences. By modeling user goals and belief states, the authors provide a logical framework for generating truly cooperative responses.
Introduction: The Literalist Trap
A fundamental challenge in Artificial Intelligence is moving beyond technical accuracy toward true cooperation. Authors Joshi, Webber, and Weischedel argue that an AI acting as an expert must actively prevent users from drawing false conclusions. In human conversation, we follow Gricean Maxims to be cooperative. For instance, if a user asks if a specific bus is running and the system says "Yes" while knowing the bus is severely delayed, it has told a literal truth but failed as a cooperative agent.
The paper addresses the gap between a user's stated goal (the literal question, or S-goal) and their intended goal (what they actually want to achieve, or I-goal). The goal is to computationally recognize when a technically correct answer to an S-goal might lead the user to fail at their I-goal.
Formal Modeling and Belief Representation
To move from theory to computation, the authors employ formal logic and linguistic theory to model expert behavior. They extend Paul Grice’s Maxims of Quality and Quantity with a revised rule: if a speaker believes their statement might imply something false to the hearer, they are obligated to provide information to block that implication.
The system utilizes a logic-based representation to model actions, preconditions, and consequences. Crucially, it tracks nested beliefs—what the Respondent (R) believes the Questioner (Q) believes ($RBQB$). This allows the system to reason about whether an action is "admissible" (possible) or "nonproductive" (fails to lead to the I-goal) within the context of the user's current state ($Sc$).
Categorizing Misleading Responses
The core contribution of the paper is a categorization of why a response might be misleading, which serves as a requirements list for expert system designers:
- Failure of Enabling Conditions: The user requests an action that is currently impossible (e.g., trying to drop a course after the deadline).
- Nonproductive Acts: The requested action is possible but will not achieve the user's underlying I-goal.
- The "Better Way": The user's request is valid, but a more efficient or effective path exists that they are unaware of.
- External Events: Events outside the user's control may impact the success of their plan regardless of their specific actions.
By performing "plan inference"—guessing what the user is actually trying to do—the system can volunteer information that the questioner might not even know they need. Silence on a crucial point is often interpreted by users as "everything is fine," which can be a failure of cooperation.
Conclusion and Future Directions
While the paper provides a formal roadmap for cooperative interfaces, it acknowledges significant hurdles. Inferring a user's I-goal from a single query remains a massive research challenge. Furthermore, systems must balance helpfulness with the "Maxim of Brevity" to avoid becoming overly wordy. Despite these challenges, the work demonstrates that "expert" behavior can be modeled as a form of plan-based reasoning rather than simple database lookup, laying the groundwork for more sophisticated AI assistants.
Summary updated