Why AI Struggles to Draw the Humble Clock

Artificial intelligence can do your taxes and write your emails, but give it a clock and things go sideways, sometimes literally. A site called AI World Clocks puts major AI models to a basic test: "Create HTML/CSS of an analog clock showing ${time}." The results, displayed side by side, are largely a mess. Numbers drift off the face, hands float in space or point nowhere in particular, and even the least-wrong attempts are subtly off. That gap between a task a child can do and a machine can't is exactly the appeal, says artist Brian Moore, who created the site and calls it "fun and funny" to watch AI fail at something so simple.

But the failures prompted Justin Pot to delve into the question of why at Popular Science. One culprit may be the training data: There may not be enough varied clock images, and clocks are surprisingly hard to describe precisely in text, which language models rely on. Pattern bias could be another issue. Modern large-language models don't really "do math" in the traditional sense; they excel at spotting patterns. With clocks, that backfires.

Search for watch photos and you'll see a familiar time: 10:10. Marketers like that setting because it frames the brand logo and looks vaguely like a smile. So a huge share of clock images online—and thus in AI training sets—show 10:10. Ask an image model to draw a clock at, say, 3:45 and it may still "guess" 10:10, because that's the dominant pattern it has seen. Read the full article at Popular Science or see the wacky clocks yourself here.