Alastair Campbell's Virtua11y 2025 presentation examines AI through an accessibility lens, separating useful applications from overhyped or harmful ones. He warns that large language models trained on typical web content tend to reproduce average, often inaccessible patterns, and are poor substitutes for real usability testing with people with disabilities. The talk also highlights promising areas such as speech-to-text, text-to-speech, and image recognition while urging practitioners to choose the right tool for each accessibility job.