Researchers have found that relying solely on prediction-based certification for trustworthy AI is not enough. They've demonstrated a separation theorem showing that two models can have identical performance metrics but differ significantly in their decision-making mechanisms and deployment behavior. This means that current safeguards, such as accuracy and calibration checks, are insufficient to ensure model trustworthiness. To address this issue, the authors propose an operational framework called the competence envelope, which combines prediction and explanation certification into a single criterion. This new approach reveals failure modes that were previously invisible in prediction behavior. --- Why it matters: This matters because it highlights the limitations of current AI certification methods and emphasizes the need for more comprehensive approaches to ensure trustworthy AI systems. Engineers working on AI development will need to consider the decision-making mechanisms behind their models, not just their performance metrics. Source: https://arxiv.org/abs/2608.20825