Explaining AI Decisions with Contrastive Reasoning

This direction is split into other directions: contrastive explanation. Its last version stays readable and citable here.

Version 1 of 155 papers70 conceptsfrozen 2026-10-04

When a self-driving car suddenly swerves or an automated scheduler denies a vacation request, our first instinct isn't to ask for a complete technical log of the system's internal state. Instead, we ask a specific, counterfactual question: "Why did you do that instead of what I expected?"

This "Why A instead of B?" framing is the cornerstone of contrastive explanation. Research across AI planning, reinforcement learning, and language modeling suggests that for AI to be truly transparent, it must move beyond justifying its own choices and start addressing the alternatives—or "foils"—that humans have in mind.

The Intuition of the Foil

In traditional explainable AI (XAI), a model might provide a list of features that led to a decision. However, as Sarath Sreedharan and Subbarao Kambhampati argue in The Emerging Landscape of Explainable AI Planning and Decision Making, humans are "selective" explainers. We don't want a data dump; we want to understand the trade-offs between what happened (the fact) and what we expected to happen (the foil).

By focusing on the delta—the specific difference in cost, risk, or feasibility between two paths—contrastive explanations reduce cognitive load. If a robot chooses a longer path, telling the user "the shorter path was blocked by a closed door" is more useful than providing a 50-page proof of the longer path's optimality.

Mechanisms of Contrastive Planning

In the field of AI Planning, researchers like Daniele Magazzeni and Michael Cashmore have developed frameworks like Towards Providing Explanations for AI Planner Decisions that treat the user's "Why not B?" as a constraint.

The system takes the user's suggested alternative, compiles it into a new planning problem, and attempts to solve it. By performing a "delta analysis" between the original optimal plan and this new, restricted plan, the system can point to specific reasons why the original was better—such as increased mission time or resource consumption. Benjamin Krarup further refined this in Contrastive Explanations of Plans through Model Restrictions, showing that explicitly highlighting these differences in action sequences and costs significantly improves user satisfaction.

Simulating the "What If" in Reinforcement Learning

For Reinforcement Learning (RL) agents, which often operate in "black-box" neural environments, contrastive reasoning requires a causal understanding of the world. Prashan Madumal proposes using Structural Causal Models (SCMs) to allow agents to answer counterfactual questions.

In Explainable Reinforcement Learning through a Causal Lens, the agent learns an "action influence model." When asked "Why not action B?", the agent simulates the outcome of B within its causal model and compares the resulting rewards and state transitions to its chosen action A. This allows the agent to say, "I chose A because B would have led to a 40% higher chance of failure," a rationale that Francisco Cruz and Peter Vamplew found to be highly intuitive for human observers.

Building Trust Through Rejection

Interestingly, explaining why something wasn't done can be more persuasive than explaining why it was. In the context of Large Language Models, Zichen Chen and Ambuj Singh introduced XplainLLM, which uses Knowledge Graphs to provide "why-not-choose" explanations. They found that clarifying why alternative answers were dismissed helps users identify the model's decision boundaries and reduces the perception of "hallucinations."

This aligns with human-centric studies by Na Du and X. Jessie Yang, whose work on option-centric rationales showed that displaying rejected alternatives alongside a recommendation helps human operators calibrate their trust more quickly. By seeing the "near-misses" or the flawed alternatives the AI considered, humans can better judge the system's competence.

The Social Exchange of Logic

Ultimately, contrastive explanation treats AI interpretability as a social dialogue rather than a technical dump. Whether it is Sarit Kraus using simulated negotiation trees to explain task allocations in CMAOE, or Julio César Bahamón using contrastive choice to reveal character personality in narrative generation, the goal is the same: to ground the machine's "logic" in the user's "expectations." By answering "Why not B?", the system proves it has considered the user's perspective, turning a black-box decision into a transparent, justifiable choice.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_explaining_ai_decisions_with_contrastive_fc6d1796,
  title = {Explaining AI Decisions with Contrastive Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explaining-ai-decisions-with-contrastive-reasoning/rdr_fc6d179642a44018a82dcf5d0f5505bb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_explaining_ai_decisions_with_contrastive_fc6d1796_v1,
  title = {Explaining AI Decisions with Contrastive Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explaining-ai-decisions-with-contrastive-reasoning/rdr_fc6d179642a44018a82dcf5d0f5505bb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}