Deep Argumentative Explanations

Traditional XAI methods may not adequately reveal how neural networks actually work: The paper argues that despite widespread adoption of explainability techniques, current XAI methods "provide little insight into the functioning of Neural Networks," suggesting that popular explanation approaches may give users a false sense of understanding without revealing the actual mechanisms driving AI decisions.

Argumentative reasoning offers a more transparent explanation framework: By grounding explanations in computational argumentation (a form of symbolic AI), Deep Argumentative Explanations (DAXs) provide explanations that reflect "the deep structure of the underlying NNs," potentially offering stakeholders clearer insight into the reasoning chains that lead to specific predictions.

New explanation methods can be both accurate and computationally efficient: The research demonstrates that DAXs exhibit "deep fidelity and low computational cost," meaning they faithfully represent how the neural network functions while remaining practical to compute—an important consideration for real-world deployment of responsible AI systems.

Human evaluation confirms that argumentative explanations are understandable and trusted: Human experiments showed that DAXs are "comprehensible to humans and align with their judgement," while being "competitive, in terms of user acceptance, with some existing approaches to XAI"—suggesting this approach could help non-experts better understand and appropriately trust AI system decisions.