Helpful, Misleading or Confusing: How Humans Perceive Fundamental Building Blocks of Artificial Intelligence Explanations

Different explanation formats serve different audiences: The paper emphasizes that mathematical formulas, graphs, and text summaries are the building blocks of AI explanations, and diverse stakeholders (engineers, researchers, consumers, regulators) likely perceive their comprehensibility differently—suggesting no single explanation format works universally.

Current evaluation methods for AI explanations are inadequate: The authors note that "explainable artificial intelligence techniques are developed at breakneck speed, but suitable evaluation approaches lag behind," highlighting a critical gap where new explanation methods are created faster than we can properly assess whether they actually help people understand AI decisions.

Complexity and scope of explanations significantly impact understanding: The paper identifies that textual explanations vary in "complexity and scope," suggesting that how much detail is included and how technical the language is fundamentally affects whether people find explanations helpful, misleading, or confusing.

User studies with real stakeholders are essential for responsible AI: Rather than assuming explanations work, the authors advocate for "setting up and executing relevant user studies" with actual representatives from different groups who will use AI systems, emphasizing that responsible AI requires empirical evidence of comprehensibility.

Simple models are a necessary starting point: By studying explanations of basic decision-making models first, the research establishes a foundation for understanding how people process AI explanations before tackling more complex systems—a methodologically sound approach to building trustworthy AI.