STILL CAN'T DO IT
Things that are easy for people and still hard for AI. Each one has a dated source, or comes from our own test, where all five models in the blind test get the same prompt, again whenever a model is updated and at least once a month. When AI gets the hang of one, it moves to "Solved since".
Prediction market: Manifold traders give a 63% chance that, at the end of 2026, there will still be a simple text task most people can solve but top AI can't (37 traders, play money). What traders believe, not a test result. Market · as of Sun 22 Nov
- SEEING AND READING
Read the time on an analog clock
On the ClockBench test, the best model read analog clocks correctly 50.6% of the time; people managed 90.1%.
Stanford AI Index 2026, technical performance - PHYSICAL WORLD
Do everyday household chores
Robots succeed in only 12% of real household tasks, even though they do well in simulations.
Stanford AI Index 2026, technical performance - PHYSICAL WORLD
Fold a basket of laundry on its own
The one home laundry robot on sale ($8,000, Bay Area only) can't cope with bedsheets, large blankets or inside-out clothes, and a remote human steps in when it gets stuck.
9to5Mac, 12 Feb 2026 - RELIABILITY
Finish multi-step computer tasks every time
AI agents still fail roughly one in three attempts on structured benchmarks.
Stanford AI Index 2026, technical performance
SOLVED SINCE…
Nothing has come off the wall yet.