AI still can't reliably read an analog clock

The world's smartest AI models can write software, analyze financial markets, and pass grueling bar exams. But hand them a simple photo of an analog clock on the wall, and they basically fall apart.

A new benchmark called ClockBench, developed by researcher Alek Safar, pit 11 frontier models from six top labs against five human testers. The task was straightforward: read analog clocks. The outcome was brutal. Humans nailed the test with 89.1 percent accuracy. The best AI model on the planet managed just 13.3 percent.

Off by an entire hour

To make sure models couldn't cheat using images scraped from the web, Safar built 180 custom analog clocks from scratch. He used 36 distinct designs featuring Roman numerals, mirrored dials, colorful backdrops, and hands pointing at tricky angles, spanning 720 total questions.

Google took the top spots, but barely. Gemini 2.5 Pro won the benchmark with a meager 13.3 percent accuracy, followed by Gemini 2.5 Flash at 10.5 percent. GPT-5 landed in third place with just 8.4 percent, and giving the model more reasoning budget barely moved the needle. Anthropic’s Claude 4.1 Opus and Claude 4 Sonnet lagged further behind at 5.6 percent and 4.2 percent.

Grok 4 took last place with an astonishing 0.7 percent. The model went into an ultra-paranoid defensive mode, flagging 63.3 percent of the clocks as invalid or broken, even though only 37 out of 180 clocks actually depicted impossible times.

The error margins tell an even uglier story. When humans misread a dial, their median error was just 3 minutes. The best AI model missed by a median of one hour. The lowest-ranking models were off by roughly three hours, which is virtually indistinguishable from guessing blindly on a 12-hour face. Across the whole benchmark, 61.7 percent of the clocks were missed by every single model.

Great at math, terrible at seeing

Interestingly, the models didn't struggle with temporal logic. When an AI accidentally read the clock face correctly, it answered follow-up questions about time zone shifts, hand rotations, and elapsed time with near-perfect accuracy.

The bottleneck isn't reasoning about time. It is pure visual geometry. The models simply cannot resolve the spatial orientation of overlapping hands against background markings and translate those angles into symbolic numbers.

Roman numerals crushed accuracy down to 3.2 percent, and circular numbering dropped it to 4.5 percent. A year ago, a Chinese study found GPT-4o hit 54.8 percent on reading basic gauges and clocks. But under ClockBench's novel, unmemorized dial patterns, current vision capabilities crumbled.

Why it matters

ClockBench exposes the dirty secret of modern multimodal AI: these systems are still language brains bolted onto crude visual interpreters. We think this benchmark proves that high benchmark scores on text-heavy exams don't translate to basic spatial common sense. A human toddler can grasp relative hand lengths and angles, but multi-billion-dollar models still struggle with it.

For developers and enterprises, this is a clear warning sign. If your workflow depends on extracting precise readings from the physical world—like reading legacy utility meters, analog industrial dials, or subtle architectural schematics—do not rely solely on general-purpose vision LLMs. Dedicated computer vision or classic edge detection tools are still far more reliable.

ClockBench is clever because it doesn't test trivia; it tests raw spatial reasoning that cannot be brute-forced with search. It shows that simply scaling compute and throwing more web scrapes at vision encoders might not fix core spatial blind spots. Frontier labs will likely need fundamentally better vision tokenization before AI can truly see the physical world as it is.