TLDRIt is found that vague or incomplete image descriptions, as well as incorrect AI outputs more broadly, can cause both BLV and sighted scientists to abandon AI workflows, and implications for AI-powered scientific QA systems are discussed. Expand