Semantics and explanation: why counterfactual explanations produce adversarial examples in deep neural networks

Counterfactual explanations and adversarial examples are mathematically the same thing, despite appearing to serve opposite purposes—one aims to explain AI decisions while the other exploits vulnerabilities. This formal equivalence reveals a fundamental tension in how we use these techniques.

The difference between helpful explanations and harmful attacks lies in semantics, not mathematics. The paper argues that what makes a counterfactual explanation meaningful to humans depends on whether we can understand what the AI's internal representations actually mean, not on the technical procedure itself.

Current explainability methods may have a hidden weakness. If counterfactual explanations are mathematically identical to adversarial examples, this suggests that explanations generated by these methods might not be as robust or trustworthy as they appear, raising questions about their reliability for high-stakes decisions.

Understanding AI decision-making requires interpreting "hidden layers," not just inputs and outputs. The paper emphasizes that truly responsible AI explanation demands we develop better methods to decode what's happening inside neural networks, rather than relying on surface-level explanation techniques.

The field needs to rethink its approach to explainability. Rather than treating counterfactual explanations as a solved problem, researchers should focus on the deeper challenge of making AI's internal reasoning semantically interpretable to humans.