Dear XAI Community, We Need to Talk! Fundamental Misconceptions in Current XAI Research

Many XAI techniques lack clear purpose: The paper finds that explanation methods are often proposed without clearly defining what they're meant to accomplish, instead relying on visually appealing heatmaps or benchmarks that may not actually measure what matters.

"Building trust" is an insufficient and problematic goal: The authors argue that using explainability primarily to build trust in AI systems is conceptually flawed, suggesting this motivation needs to be reconsidered and replaced with more substantive objectives.

Unfounded assumptions about how AI learns are widespread: Current XAI research often assumes deep learning algorithms learn interpretable "concepts," but these assumptions lack solid evidence and may mislead researchers about what explanations can actually reveal about model behavior.

The field is growing in the wrong direction: Rather than improving, the paper warns that problematic practices in XAI research are actually increasing, with more techniques being proposed on shaky conceptual and methodological grounds.

XAI needs stronger foundational standards: The authors call for the field to establish clearer ethical frameworks, more rigorous methodology, and better-defined purposes before continuing to develop new explanation techniques—essentially asking the XAI community to pause and reconsider its fundamental approach.