The limits of AI: When AI falls short and what to do instead

A top-tier AI once gave unsafe cancer treatment advice—inside a major hospital, with doctors watching. In one part of your life, you’re trusting that same kind of technology alone. The twist is: you probably don’t know which part… or how often it quietly gets things wrong.

Stanford’s HELM benchmark recently found that leading language models missed nearly a third of commonsense questions. Not obscure trivia—basic things most 10‑year‑olds breeze through. At the same time, McKinsey estimates 60–70% of our work hours involve tasks that *could* be automated. Those two facts pull in opposite directions: we’re racing to hand more over to systems that still stumble on everyday reasoning.

In your own life, this tension shows up in subtle ways: the budgeting app that confidently mislabels a crucial payment, the “smart” calendar that triple-books you, the writing assistant that invents a source you never checked. None of these failures look dramatic in the moment. They feel like small glitches, easy to dismiss. But stacked together, they point to a bigger question: where, exactly, should you *not* outsource your thinking—and what should you do instead when the algorithm sounds sure but your instincts hesitate?