Misleading headlines abound, claiming that machine learning can "accurately" predict criminality, psychosis, sexual orientation, and bestselling books. But, when practitioners claim their model achieves "high accuracy," it's often bogus. Can AI "tell" if you're going to have a heart attack? Contrary to bold, public claims, no it cannot. This episode unpacks the undeniable yet common "accuracy fallacy," which misleads the public into believing that machine learning can distinguish between positive and negative cases and usually be right about it.
See my Scientific American blog article to dig in further and access many links: https://blogs.scientificamerican.com/observations/the-medias-coverage-of-ai-is-bogus/
Watch my two-part video coverage of the accuracy fallacy: