According to Ethan Mollick on X, sharing Matt Beane’s new paper, proactive AI assistance can increase user cognitive load and degrade task performance, with models failing to recover once they derail while humans do recover, as reported by the paper on arXiv. According to Matt Beane on X, the study offers quantitative measures showing that AI-initiated suggestions impose measurable cognitive overhead that worsens work outcomes, with evidence gathered over a three-year research effort and published on arXiv. According to the arXiv preprint, the findings imply that product teams should throttle unsolicited AI prompts, stage guidance contextually, and enable quick user reorientation to reduce derailment and restore performance in operational workflows.