💻 computer science

A Communication-First Account of Explanation

This paper proposes a formal, communication-first account of causal explanation grounded in conversational pragmatics and interventionist theory, demonstrating how this framework naturally generates recognized explanatory virtues and empirical patterns regarding norms while effectively integrating insights from cognitive science into philosophical analysis.

Original authors: Jacqueline Harding, Tobias Gerstenberg, Thomas Icard

Published 2026-06-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jacqueline Harding, Tobias Gerstenberg, Thomas Icard

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out why your car won't start. You ask a mechanic, "Why won't it start?"

If the mechanic says, "Because the universe is vast and complex," that's technically true, but it's a terrible explanation. If they say, "Because the battery is dead," that's helpful. But what if you already knew the battery was dead? Would they say, "Because the battery is dead, and also the alternator is broken"?

This paper by Jacqueline Harding, Tobias Gerstenberg, and Thomas Icard argues that explanation isn't just about finding the "true" cause in the world; it's about a conversation between two people.

They propose a "Communication-First" approach. Instead of treating explanation as a static fact waiting to be discovered, they treat it like a game of Rational Speech Acts. Think of it as a high-stakes game of "Guess What I'm Thinking," but with a twist: the goal is to help the other person make a better decision.

Here is the breakdown of their ideas using everyday analogies:

1. The Setup: The Speaker, The Listener, and The Goal

Imagine a Speaker (who knows the truth) and a Listener (who is confused and has a specific goal).

The Listener's Goal: The listener isn't just asking for trivia. They usually have a "decision problem." Maybe they need to fix their roof, apologize to a friend, or buy the right paint for a pigeon.
The Speaker's Job: The speaker's job isn't just to dump data. Their job is to pick the one piece of information that will help the listener solve their specific problem most effectively.

The Analogy: Imagine you are playing a video game where you have to navigate a maze. You are the Listener. Your friend is the Speaker. You don't know the map.

If your friend says, "The maze is huge," that's true, but useless.
If your friend says, "Turn left at the red door," that's helpful if you need to get to the exit quickly.
If your friend says, "Turn left at the red door, but also, the ceiling is made of glass," that might be too much info if you just need to get to the exit.

The paper argues that a "good explanation" is simply the message that helps the listener win their game (make the best decision) the most.

2. Why "Context" Matters More Than "Truth"

The authors use a model called a Structural Causal Model. Think of this as a flowchart of how the world works.

The "Roof" Example: Imagine a house catches fire. Was it because the roof was thatched? Or because there was a drought?
- If your goal is to prevent future fires, knowing the roof was thatched is crucial (you should replace your roof). Knowing it was a drought is less useful (you can't control the weather).
- If your goal is to figure out who is to blame, maybe the drought is the key factor.
The Insight: The "best" explanation changes depending on what the listener wants to do. The paper shows that if you build a computer model that tries to maximize the listener's success, it naturally picks the explanation that fits the listener's goal, even if the speaker knows both causes.

3. The "Surprise" Factor (Why we don't explain what you already know)

You might think a good explanation must always be something the listener doesn't know. The paper says: Not necessarily.

The "Late Meeting" Example: Bob is late. He knows he is late. He also knows Charlie is angry. Bob asks, "Why is Charlie angry?"
- Bob already knows "I was late" is a cause.
- But Bob doesn't know if being late was the only cause, or if he also forgot Charlie's birthday.
- If Alice says, "Because you were late," Bob might think, "Oh, so it's only because I was late. I don't need to apologize for the birthday."
- If Alice says, "Because you forgot the birthday," Bob knows he messed up twice.
The Insight: Even if the listener knows a fact (being late), hearing it as an explanation can still be useful because it tells the listener which causal story is true. The explanation isn't about the fact; it's about narrowing down the possibilities.

4. The "Normal" vs. "Abnormal" Rule

The paper explains a weird quirk in human psychology:

If two things must happen together to cause a result (like a thatched roof AND a drought to start a fire), we tend to blame the weird thing (the drought, because droughts are rare).
If either thing alone can cause a result (like a short circuit OR a gas leak causing a fire), we tend to blame the normal thing (the short circuit, because they happen often).

The authors' model predicts this perfectly. It calculates that citing the "weird" thing in the first scenario helps the listener understand the specific conditions needed to stop the fire. Citing the "normal" thing in the second scenario helps the listener understand the most common risk. The model doesn't need to be told this rule; it figures it out by trying to be helpful.

5. Simplicity and "Cost"

Finally, the paper addresses why we prefer short, simple explanations.

The Cost of Speaking: Saying "The short circuit caused the fire" is easier than saying "The most noteworthy event of the year caused the fire," even if they mean the same thing.
The Trade-off: The model includes a "cost" for long or confusing messages. If a long message doesn't give the listener significantly more help than a short one, the "good speaker" will choose the short one to save energy and avoid confusion.

The Big Picture

The authors are saying that for centuries, philosophers have tried to find the "perfect" explanation by looking at the world's hard facts. They argue this is backwards.

Explanation is a tool for communication.

It emerges from the dynamic between a speaker trying to be helpful and a listener trying to solve a problem.
When you get the communication right (knowing the listener's goals and what they already know), the "virtues" of a good explanation (being simple, being surprising, being relevant) just happen naturally. You don't need to force them; they are the result of a good conversation.

In short: A good explanation is the one that helps the listener win their game.

Problem
The philosophical problem of explanation is traditionally framed as locating objective explanatory relations in the world, treating human interests and communicative contexts as peripheral or merely "pragmatic" noise to be filtered out to reveal the core explanatory facts. While philosophers acknowledge that pragmatic factors (such as background knowledge and downstream interests) influence explanatory judgments, they often relegate these to the margins of the explanatory project. Conversely, while psychological research has been influenced by philosophical accounts, philosophers have been reluctant to incorporate insights from cognitive science and formal pragmatics into their theories. Existing formal accounts of explanation, including interventionist approaches, often fail to engage deeply with how people actually produce and interpret explanations, or they treat pragmatic factors as constraints rather than constitutive elements. The central question motivating this paper is: What happens if we "put communication first" in developing an account of explanation?

Methodology
The authors develop a formal account of causal explanation grounded in a theory of conversational pragmatics, specifically utilizing the Rational Speech Acts (RSA) framework. This framework models communication as a hierarchy of agents:

The Literal Listener ( $L_0$ ): Updates their epistemic state based on the semantic content of an utterance.
The Speaker ( $S$ ): Chooses an utterance to maximize the listener's expected utility in a decision problem, taking into account the literal listener's interpretation and the costs of production/processing.
The Pragmatic Listener ( $L$ ): Infers the speaker's intent by reasoning about why the speaker chose that specific utterance, updating their beliefs based on the speaker's optimization strategy.

The authors integrate this with Structural Causal Models (SCMs) to formalize the causal content of explanations. They assume that "why" questions request causal information and that "because" statements communicate relations of actual causal dependence. The listener's decision problem is formalized as a manipulation game, where the goal is to identify variables that, when intervened upon, can change the value of the explanandum across various background conditions.

The model evaluates the "goodness" of an explanation not by intrinsic properties of the causal relation, but by the extent to which the explanation updates the listener's beliefs to improve their performance in downstream decision problems (e.g., manipulation, control, or attribution of responsibility).

Key Contributions and Results
The paper demonstrates that several widely recognized "explanatory virtues" and empirical patterns of causal selection emerge naturally from the dynamics of this communication-first model, without needing to be stipulated as axioms:

Downstream Interests: The model shows that explanations are better when they are useful for the listener's specific decision problems. In cases where a listener faces a forward-looking decision (e.g., replacing a roof), citing a cause that is sensitive to that decision (e.g., the roof material) is preferred over a less relevant cause (e.g., drought), even if both are true causes.
Background Knowledge: The model accounts for the intuition that explanations should cite unknown information. However, it also predicts that citing known causes can be a good explanation if the speaker's choice of that known cause (over a known alternative) allows the listener to infer hidden causal structures (e.g., inferring that a cause is sufficient rather than conjunctive).
Invariance and Causal Selection: Using the manipulation game, the model reproduces empirical findings that speakers prefer to cite "abnormal" causes in conjunctive structures and "normal" causes in disjunctive structures. It further predicts that listeners can infer the normality of causes or the structure of the causal model based on which cause the speaker chooses to cite.
Minimality and Simplicity: The model incorporates production and processing costs. It explains why minimal explanations are often preferred (due to lower costs) but also predicts that longer, more informative explanations are preferred when the cost difference is outweighed by the utility gain (e.g., identifying multiple culprits in a theft scenario). It also accounts for "proportionality" (e.g., preferring "red" over "scarlet" in certain contexts) by modeling the trade-off between the specificity of the information and the cognitive cost of processing it.
Level of Abstraction: The model explains why explanations at the "right" level of abstraction are better. It shows that citing a more abstract cause (e.g., "red") can be superior to a specific one (e.g., "scarlet") if the specific detail is not invariant across the listener's relevant decision problems, or if the cost of processing the specific detail is too high.

Significance
The paper argues that a "communication-first" approach is a necessary prerequisite for philosophical work on explanation. By formalizing explanation as a cooperative communication act aimed at helping a listener achieve downstream goals, the authors show that explanatory virtues are not arbitrary or peripheral but are natural consequences of the interaction between causal structure, listener epistemic states, and decision problems.

The significance of this work lies in its ability to bridge the gap between philosophical accounts of explanation and empirical findings in cognitive science. It provides a simple but powerful framework that incorporates insights from the cognitive sciences (specifically causal cognition and pragmatics) into philosophy. The authors claim that this approach offers a unified account where explanatory virtues emerge from the dynamics of conversation, rather than being imposed as external rules. This suggests that getting clear on the communicative dimension of explanation is essential for understanding the nature of explanation itself. The paper concludes by noting that while the current model is a starting point, it opens avenues for future work on scientific explanation, individual (non-communicative) explanation seeking, variable introduction, and non-causal domains like mathematical explanation.