The Pragmatic Turn in Explainable Artificial Intelligence (XAI)

Understanding should be the real goal of XAI, not just explanation: The paper argues that explainable AI efforts should focus on whether stakeholders actually understand a model or decision, rather than simply providing explanations. Without clarity on what "understanding" means, XAI strategies lack a well-defined objective.

We need to move beyond the impossible standard of perfect accuracy: Current XAI approaches often require explanations to be completely factually correct ("factivity condition"), which is unrealistic for many machine learning models. The paper proposes focusing instead on pragmatic conditions—finding the best fit between a model and the methods used to explain it to different audiences.

Approximation models are essential for real interpretability: The research concludes that interpretative or approximation models (simplified versions that capture key aspects of how a model works) are not just helpful but necessary for achieving genuine post-hoc interpretability—the ability to understand a model after it's already been built and deployed.

One-size-fits-all explanations won't work: By emphasizing pragmatic understanding, the paper implies that explanations must be tailored to different stakeholders' needs and backgrounds, rather than assuming a single explanation method will work for regulators, users, and developers alike.

Current XAI approaches may be missing the mark: The paper critiques the "purely functionalist approach" dominating recent XAI literature, suggesting that many current methods for making AI interpretable may not actually achieve meaningful understanding among stakeholders.